├── LICENSE ├── README.md ├── code_for_figures ├── benchmarking │ ├── .DS_Store │ ├── notebooks │ │ ├── .DS_Store │ │ ├── brain │ │ │ ├── get_ground_truth.ipynb │ │ │ ├── graphst_results.ipynb │ │ │ ├── knn_results.ipynb │ │ │ ├── radius_results.ipynb │ │ │ ├── stagate_results.ipynb │ │ │ └── utag_results.ipynb │ │ ├── gut │ │ │ ├── get_ground_truth.ipynb │ │ │ ├── graphst_results.ipynb │ │ │ ├── knn_results.ipynb │ │ │ ├── radius_results.ipynb │ │ │ ├── stagate_results.ipynb │ │ │ └── utag_results.ipynb │ │ ├── plot_performance.ipynb │ │ └── simulation │ │ │ ├── get_ground_truth.ipynb │ │ │ ├── graphst_results.ipynb │ │ │ ├── knn_results.ipynb │ │ │ ├── radius_results.ipynb │ │ │ ├── stagate_results.ipynb │ │ │ └── utag_results.ipynb │ └── scripts │ │ ├── .DS_Store │ │ ├── brain │ │ ├── brain_graphst.sh │ │ ├── brain_knn.sh │ │ ├── brain_radius.sh │ │ ├── brain_stagate.sh │ │ ├── brain_utag.sh │ │ └── performance_brain.sh │ │ ├── graphst_.py │ │ ├── gut │ │ ├── gut_graphst.sh │ │ ├── gut_knn.sh │ │ ├── gut_radius.sh │ │ ├── gut_stagate.sh │ │ ├── gut_utag.sh │ │ └── performance_gut.sh │ │ ├── knn.py │ │ ├── modified_existing_methods │ │ ├── graphst │ │ │ └── GraphST_mod.py │ │ ├── stagate │ │ │ └── utils.py │ │ └── utag │ │ │ └── segmentation.py │ │ ├── radius.py │ │ ├── simulation │ │ ├── performance_simulation.sh │ │ ├── simulation_graphst.sh │ │ ├── simulation_knn.sh │ │ ├── simulation_radius.sh │ │ ├── simulation_stagate.sh │ │ └── simulation_utag.sh │ │ ├── stagate_.py │ │ └── utag_.py └── generalization │ ├── .DS_Store │ ├── brain │ ├── concat.ipynb │ ├── length_scales.ipynb │ ├── n_pcs.ipynb │ ├── nmf.ipynb │ ├── schematic.ipynb │ ├── shuffling.ipynb │ ├── xspecies.ipynb │ └── xtech.ipynb │ └── gut │ └── xcondition.ipynb ├── docs ├── imgs │ ├── github_idea_1_dark.png │ ├── github_idea_1_light.png │ ├── github_idea_2_dark.png │ ├── github_idea_2_light.png │ ├── github_obstacle_1_dark.png │ ├── github_obstacle_1_light.png │ ├── github_obstacle_2_dark.png │ └── github_obstacle_2_light.png └── tutorials │ └── tutorial.ipynb ├── pyproject.toml └── src └── spin ├── __init__.py ├── cli.py └── spin.py /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # SPIN: spatial integration of spatially resolved transcriptomics (SRT) data 2 | [![Biorxiv badge](https://zenodo.org/badge/doi/10.1101/2023.06.30.547258.svg)](https://doi.org/10.1101/2023.06.30.547258) ⬅️ manuscript
3 | [![Zenodo badge](https://zenodo.org/badge/doi/10.5281/zenodo.13883268.svg)](https://doi.org/10.5281/zenodo.13883268) ⬅️ data
4 | 5 | SPIN is a simple, Scanpy-based implementation of the subsampling and smoothing approach described in the manuscript *Mitigating autocorrelation during spatially resolved transcriptomics data analysis*. It enables the alignment and analysis of transcriptionally defined tissue regions across multiple SRT datasets, regardless of morphology or experimental technology, using conventional single-cell tools. Here we include information regarding: 6 | 7 | 1. A conceptual overview of the approach 8 | 2. Package requirements 9 | 3. Installation instructions 10 | 4. Basic usage principles 11 | 12 | For examples of downstream analysis (e.g. differentially expressed gene analysis and trajectory inference), see the [tutorial](docs/tutorials/tutorial.ipynb) notebook. For further details on SPIN parameters, import SPIN into Python as shown below and run `help(spin)`. 13 | 14 | ## 1. Conceptual overview 15 | * Conventional single-cell analysis can identify molecular *cell types* by considering each cell individually. 16 | * However, it does not incorporate spatial information. 17 | 18 | 19 | 20 | 21 | 22 | 23 | * Arguably the simplest way to incorporate spatial information and identify molecular *tissue regions* is to spatially smooth gene expression features across neighboring cells in the tissue. 24 | * This can be done by setting the features of each cell to the average of its spatial neighborhood. 25 | 26 | 27 | 28 | 29 | 30 | 31 | * However, a problem arises when smoothed representations of each cell are compared to one another. 32 | * Physically adjacent cells will have almost identical neighborhoods and thus almost identical smoothed representations. 33 | 34 | 35 | 36 | 37 | 38 | 39 | * Thus, we end up with nearest neighbors in feature space that are just nearest neighbors in physical space. 40 | * Because conventional methods for downstream anlaysis rely on the nearest neighbors graph in feature space, this leads to reconstruction of physical space in latent space rather than representing the true underlying large scale molecular patterns. 41 | * Here, we implement an approach in which each cell's spatial neighborhood is randomly subsampled before averaging, allowing the *exact neighborhood* composition to vary while still maintaining the *general molecular* composition. 42 | 43 | 44 | 45 | 46 | 47 | 48 | Ultimately, this approach enables the application of conventional single-cell tools to spatial molecular features in SRT data, yielding regional analogies for each tool. For more details and examples, please refer to the manuscript and [tutorial](docs/tutorials/tutorial.ipynb). 49 | 50 | ## 2. Requirements: 51 | 52 | ### Software: 53 | * Tested on MacOS (Monterey, Ventura) and Linux (Red Hat Enterprise Linux 7). 54 | * Command Line Tools is required for `pip` installing this package from GitHub. While it comes standard on most machines, those without it may encounter an `xcrun: error` when following the installation instructions below. See [here](https://apple.stackexchange.com/questions/254380/why-am-i-getting-an-invalid-active-developer-path-when-attempting-to-use-git-a) for simple instructions on how to install it. 55 | * Python >= 3.9 56 | * The only dependency is Scanpy. For details, see [`pyproject.toml`](pyproject.toml). 57 | 58 | ### Data: 59 | * One or more SRT datasets in `.h5ad` format 60 | * An expression matrix under `.X` (both sparse and dense representations supported) 61 | * Spatial coordinates under `.obsm` (key can be specified with argument `spatial_key`) 62 | * Batch information 63 | * If multiple batches in single dataset, batch labels provided under column in `.obs` with column name `batch_key`. 64 | * If multiple batches in separate datasets, batch labels for each dataset provided as input. 65 | 66 | ## 3. Installation 67 | 68 | ### From GitHub: 69 | ``` 70 | pip install git+https://github.com/wanglab-broad/spin@main 71 | ``` 72 | Takes ~5 mins. 73 | 74 | ## 4. Usage 75 | ### In Python: 76 | Consider the marmoset and mouse data from the manuscript which we provide as a demo: 77 | ```python 78 | import scanpy as sc 79 | 80 | adata_marmoset = sc.read( 81 | 'data/marmoset.h5ad', 82 | backup_url='https://zenodo.org/record/8092024/files/marmoset.h5ad?download=1' 83 | ) 84 | adata_mouse = sc.read( 85 | 'data/mouse.h5ad', 86 | backup_url='https://zenodo.org/record/8092024/files/mouse.h5ad?download=1' 87 | ) 88 | ``` 89 | 90 | These datasets can be spatially integrated and clustered using `spin`. The `batch_key` argument corresponds to the name of a new column in `adata.obs` that stores the batch labels for each dataset. The `batch_labels` argument is a list of these batch labels in the same order as the input AnnDatas: 91 | ```python 92 | from spin import spin 93 | 94 | adata = spin( 95 | adatas=[adata_marmoset, adata_mouse], 96 | batch_key='species', 97 | batch_labels=['marmoset', 'mouse'], 98 | resolution=0.7 99 | ) 100 | ``` 101 | This performs the following steps: 102 | * `integrate`: 103 | 1. Subsampling and smoothing of each dataset individually (stored under `adata.layers['smooth']`) 104 | 2. Joint PCA across both smoothed datasets 105 | 3. Integration of the resulting PCs using Harmony (stored under `adata.obsm['X_pca_spin']`) 106 | * `cluster`: 107 | 1. Latent nearest neighbor search 108 | 2. Leiden clustering with a resolution of 0.7 (stored under `adata.obs['region']`) 109 | 3. UMAP (stored under `adata.obsm['X_umap_spin']`) 110 | 111 | Note that `spin` can equivalently take as input a single AnnData containing multiple labeled batches. It can also take a single AnnData containing one batch for finding regions in a single dataset. For examples, see the [tutorial](docs/tutorials/tutorial.ipynb). 112 | 113 | 114 | The resulting region clusters can then be visualized using standard Scanpy functions: 115 | ```python 116 | # In physical space 117 | sc.set_figure_params(figsize=(7,5)) 118 | sc.pl.embedding(adata, basis='spatial', color='region') 119 | 120 | # In UMAP space 121 | sc.set_figure_params(figsize=(4,4)) 122 | sc.pl.embedding(adata, basis='X_umap_spin', color='region') 123 | ``` 124 | Downstream analysis (e.g. DEG analysis, trajectory inference) can then be performed using standard Scanpy functions as well. 125 | For examples of downstream analysis, see the [tutorial](docs/tutorials/tutorial.ipynb). 126 | For further details on the parameters of `spin`, import SPIN into Python as shown above and run `help(spin)`. 127 | 128 | ### From the shell: 129 | SPIN can be executed from the shell using the `spin` command as shown below (the path is identified automatically; see [`spin_cli`](src/spin/cli.py) and [`pyproject.toml`](pyproject.toml)) 130 | 131 | Shell submission requires a read path to the relevant dataset(s) as well as a write path for the output dataset. Otherwise, provide the same parameters you would when running in Python as above: 132 | ```python 133 | spin \ 134 | --adata_paths data/marmoset.h5ad data/mouse.h5ad \ 135 | --write_path data/marmoset_mouse_spin.h5ad \ 136 | --batch_key species \ 137 | --batch_labels marmoset mouse \ 138 | --resolution "0.7" 139 | ``` 140 | 141 | Just as when running in Python, a single AnnData containing multiple batches can be passed in instead, as well as just a single dataset containing a single batch. 142 | -------------------------------------------------------------------------------- /code_for_figures/benchmarking/.DS_Store: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/wanglab-broad/spin/53c174bfcc029f2614bf381f3d61557f3e1824d1/code_for_figures/benchmarking/.DS_Store -------------------------------------------------------------------------------- /code_for_figures/benchmarking/notebooks/.DS_Store: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/wanglab-broad/spin/53c174bfcc029f2614bf381f3d61557f3e1824d1/code_for_figures/benchmarking/notebooks/.DS_Store -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/.DS_Store: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/wanglab-broad/spin/53c174bfcc029f2614bf381f3d61557f3e1824d1/code_for_figures/benchmarking/scripts/.DS_Store -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/brain/brain_graphst.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=80G 4 | #$ -l h_rt=01:30:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/brain/graphst/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/graphst/ 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | k_physical=50 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/graphst_.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/brain/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --k_physical $k_physical \ 28 | --k_latent $k_latent \ 29 | --leiden_res "0.2" \ 30 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/brain/graphst/reconstructions_${sample_rate}_${k_physical}physical_${k_latent}latent/ \ 31 | --random_seed $random_seed \ 32 | --n_epochs 50 \ -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/brain/brain_knn.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=16G 4 | #$ -l h_rt=01:30:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/brain/knn/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/xax 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | k_physical=50 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/knn.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/brain/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --k_physical $k_physical \ 28 | --k_latent $k_latent \ 29 | --n_pcs 50 \ 30 | --leiden_res "0.2" \ 31 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/brain/knn/reconstructions_${sample_rate}_${k_physical}physical_${k_latent}latent/ \ 32 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/brain/brain_radius.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=16G 4 | #$ -l h_rt=01:30:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/brain/radius/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/xax 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | radius=550 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/radius.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/brain/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --radius $radius \ 28 | --k_latent $k_latent \ 29 | --n_pcs 50 \ 30 | --leiden_res "0.2" \ 31 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/brain/radius/reconstructions_${sample_rate}_${radius}radius_${k_latent}latent/ \ 32 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/brain/brain_stagate.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=64G 4 | #$ -l h_rt=02:00:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/brain/stagate/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/stagate_pyg/ 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | k_physical=50 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/stagate_.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/brain/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --k_physical $k_physical \ 28 | --k_latent $k_latent \ 29 | --leiden_res "0.2" \ 30 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/brain/stagate/reconstructions_${sample_rate}_${k_physical}physical_${k_latent}latent/ \ 31 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/brain/brain_utag.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=86G 4 | #$ -l h_rt=02:30:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/brain/utag/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/bridget/envs/utag/ 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | radius=550 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/utag_.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/brain/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --radius $radius \ 28 | --k_latent $k_latent \ 29 | --n_pcs 50 \ 30 | --leiden_res "0.2" \ 31 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/brain/utag/reconstructions_${sample_rate}_${radius}radius_${k_latent}latent/ \ 32 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/brain/performance_brain.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | tissue=brain 4 | 5 | methods=(knn radius utag stagate graphst) 6 | jobids=(42435516 42435533 42435534 42435535 42435536) 7 | n_methods="${#methods[@]}" 8 | 9 | # Write job resource metrics to disk 10 | for i in $(seq 0 $((n_methods-1))); do 11 | method=${methods[i]} 12 | jobid=${jobids[i]} 13 | write_path=/stanley/WangLab/kamal/data/projects/spin/reviews/results/${tissue}/${method}/ 14 | performance=${write_path}performance_${tissue}_${method}_${jobid}.txt 15 | echo "Writing $performance" 16 | touch $performance 17 | nonsub_tasks=($(seq 11 11 110)) # only compare using non-subsampling runs 18 | for task in ${nonsub_tasks[@]}; do 19 | qacct -j $jobid -t $task >> $performance; 20 | done; 21 | done 22 | echo "Done" -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/graphst_.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | import argparse 4 | 5 | from GraphST import GraphST_mod 6 | import torch 7 | import scanpy as sc 8 | import numpy as np 9 | from sklearn.neighbors import NearestNeighbors 10 | from sklearn.decomposition import PCA 11 | import scipy as sp 12 | from sklearn.cluster import KMeans 13 | from sklearn.metrics import adjusted_rand_score 14 | 15 | 16 | def smooth( 17 | adata, 18 | k_physical, 19 | sample_rate, 20 | n_epochs, 21 | learning_rate, 22 | alpha, 23 | beta, 24 | gamma, 25 | random_seed, 26 | ): 27 | 28 | # Find latent neighbors 29 | nbrs = NearestNeighbors(n_neighbors=k_physical).fit(adata.obsm['spatial']) 30 | A = nbrs.kneighbors_graph() 31 | adata.obsp['spatial_connectivities'] = A.copy() 32 | 33 | # Subsample 34 | if sample_rate: 35 | adata.obsp['spatial_connectivities_subsampled'] = A.copy() 36 | for i in range(len(adata)): 37 | nbr_idxs = adata.obsp['spatial_connectivities'][i].nonzero()[1] 38 | idxs_to_drop = np.random.choice( 39 | nbr_idxs, 40 | size=int(len(nbr_idxs)*(1-sample_rate)), 41 | replace=False, 42 | ) 43 | for j in idxs_to_drop: 44 | adata.obsp['spatial_connectivities_subsampled'][i,j] = 0 45 | adata.obsp['spatial_connectivities_subsampled'].eliminate_zeros() 46 | else: 47 | adata.obsp['spatial_connectivities_subsampled'] = np.eye(adata.shape[0]) 48 | 49 | # Store in GraphST-readable format 50 | adata.obsm['adj'] = adata.obsp['spatial_connectivities_subsampled'] 51 | adata.obsm['graph_neigh'] = adata.obsp['spatial_connectivities_subsampled'] 52 | 53 | # Avoid preprocessing (method consistency argument; choosing simplicity) 54 | adata.var['highly_variable'] = np.ones(adata.shape[1], dtype=bool) 55 | 56 | # Train model 57 | device = torch.device('cpu') 58 | model = GraphST_mod.GraphST( 59 | adata, 60 | datatype='Stereo', 61 | device=device, 62 | epochs=n_epochs, 63 | learning_rate=learning_rate, 64 | alpha=alpha, 65 | beta=beta, 66 | gamma=gamma, 67 | ) 68 | adata = model.train() 69 | adata.obsm['X_smoothed'] = adata.obsm['emb'].copy() 70 | 71 | return adata 72 | 73 | 74 | def binary_search_leiden( 75 | adata, 76 | leiden_res, 77 | n_regions, 78 | jitter=0.1 79 | ): 80 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 81 | n_clusters = len(adata.obs['leiden'].unique()) 82 | lo, hi = 0, 5 83 | iters = 0 84 | while n_clusters != n_regions: 85 | if n_clusters < n_regions: 86 | lo = leiden_res 87 | leiden_res = (leiden_res+hi)/2 88 | elif n_clusters > n_regions: 89 | hi = leiden_res 90 | leiden_res = (leiden_res+lo)/2 91 | iters += 1 92 | if iters > 10: 93 | leiden_res += np.random.choice([-1,1]) * jitter 94 | hi += jitter 95 | lo -= max(0,jitter) 96 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 97 | n_clusters = len(adata.obs['leiden'].unique()) 98 | print(leiden_res, n_clusters, flush=True) 99 | print() 100 | 101 | 102 | def cluster( 103 | adata, 104 | n_pcs, 105 | k_latent, 106 | n_regions, 107 | leiden_res, 108 | ): 109 | 110 | # Calculate latent representation 111 | pca = PCA(n_components=n_pcs) 112 | adata.obsm['X_smoothed_pca'] = pca.fit_transform(adata.obsm['X_smoothed']) 113 | 114 | # Find latent neighbors 115 | nbrs = NearestNeighbors(n_neighbors=k_latent).fit(adata.obsm['X_smoothed_pca']) 116 | adata.obsp['smoothed_connectivities'] = nbrs.kneighbors_graph(mode='connectivity') 117 | adata.obsp['smoothed_distances'] = nbrs.kneighbors_graph(mode='distance') 118 | 119 | # Add metadata expected by Scanpy's Leiden and UMAP 120 | adata.uns['smoothed'] = {'connectivities_key': 'smoothed_connectivities', 121 | 'distances_key': 'smoothed_distances', 122 | 'params': { 123 | 'n_neighbors': k_latent, 124 | 'method': 'umap', 125 | 'random_state': 0, 126 | 'metric': 'euclidean', 127 | 'use_rep': 'X_smoothed_pca' 128 | } 129 | } 130 | 131 | # Perform k-means clustering 132 | kmeans = KMeans(n_clusters=n_regions, n_init=10) 133 | kmeans.fit(adata.obsm['X_smoothed_pca']) 134 | adata.obs['kmeans'] = kmeans.labels_.astype(str) 135 | 136 | # Perform Leiden clustering 137 | binary_search_leiden(adata, leiden_res, n_regions) 138 | 139 | 140 | def quantify(adata, n_regions): 141 | 142 | # Quantify spatial reconstruction 143 | Cp = adata.obsp['spatial_connectivities'] 144 | Cl = adata.obsp['smoothed_connectivities'] 145 | recon_score = Cp.multiply(Cl).count_nonzero() / (Cp+Cl).count_nonzero() 146 | 147 | # Quantify k-means accuracy 148 | kmeans_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['kmeans']) 149 | 150 | # Quantify Leiden accuracy 151 | leiden_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['leiden']) 152 | 153 | return recon_score, kmeans_ari, leiden_ari 154 | 155 | 156 | def main( 157 | read_path, 158 | sample_rate, 159 | k_physical, 160 | k_latent, 161 | n_pcs, 162 | leiden_res, 163 | write_path, 164 | random_seed, 165 | decimals, 166 | n_epochs, 167 | learning_rate, 168 | alpha, 169 | beta, 170 | gamma, 171 | ): 172 | 173 | # Initialize logger 174 | logger = logging.getLogger('GraphST') 175 | logger.setLevel(logging.INFO) 176 | formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') 177 | ch = logging.StreamHandler() 178 | ch.setLevel(logging.INFO) 179 | ch.setFormatter(formatter) 180 | logger.addHandler(ch) 181 | 182 | # Set random seed 183 | np.random.seed(random_seed) 184 | 185 | # Load data 186 | logger.info(f'Loading data from {read_path}') 187 | adata = sc.read_h5ad(read_path) 188 | n_regions = len(adata.obs['region_true'].unique()) 189 | 190 | # Smooth 191 | logger.info('Smoothing') 192 | logger.info(f'\tsample_rate={sample_rate}') 193 | logger.info(f'\trandom_seed={random_seed}') 194 | adata = smooth( 195 | adata, 196 | k_physical, 197 | sample_rate, 198 | n_epochs, 199 | learning_rate, 200 | alpha, 201 | beta, 202 | gamma, 203 | random_seed, 204 | ) 205 | 206 | # Cluster 207 | logger.info('Clustering') 208 | cluster(adata, n_pcs, k_latent, n_regions, leiden_res) 209 | 210 | # Quantify results 211 | logger.info('Quantifying results') 212 | recon_score, kmeans_ari, leiden_ari = quantify(adata, n_regions) 213 | 214 | # Write results to disk 215 | os.makedirs(write_path, exist_ok=True) 216 | results = f'{np.round(recon_score, decimals=decimals)}_' + \ 217 | f'{np.round(kmeans_ari, decimals=decimals)}_' + \ 218 | f'{np.round(leiden_ari, decimals=decimals)}' 219 | filename_txt = f'{random_seed}_{results}.txt' 220 | write_path_txt = os.path.join(write_path, filename_txt) 221 | logger.info(f'Writing results to {write_path_txt}') 222 | f = open(write_path_txt, 'a') 223 | f.close() 224 | 225 | # Write example AnnData to disk 226 | if random_seed == 0: 227 | filename_adata = f'adata_{sample_rate}.h5ad' 228 | write_path_adata = os.path.join(write_path, filename_adata) 229 | logger.info(f'Saving AnnData to {write_path_adata}') 230 | adata.write(write_path_adata) 231 | 232 | logger.info('Done') 233 | 234 | 235 | if __name__=='__main__': 236 | 237 | parser = argparse.ArgumentParser() 238 | parser.add_argument('--read_path', type=str) 239 | parser.add_argument('--sample_rate', type=float, default=None) 240 | parser.add_argument('--k_physical', type=int, default=50) 241 | parser.add_argument('--k_latent', type=int, default=15) 242 | parser.add_argument('--n_pcs', type=int) 243 | parser.add_argument('--leiden_res', type=float) 244 | parser.add_argument('--write_path', type=str) 245 | parser.add_argument('--random_seed', type=int) 246 | parser.add_argument('--decimals', type=int, default=5) 247 | parser.add_argument('--n_epochs', type=int, default=100) 248 | parser.add_argument('--learning_rate', type=float, default=0.001) 249 | parser.add_argument('--alpha', type=float, default=10) 250 | parser.add_argument('--beta', type=float, default=1) 251 | parser.add_argument('--gamma', type=float, default=1) 252 | args = parser.parse_args() 253 | 254 | main( 255 | args.read_path, 256 | args.sample_rate, 257 | args.k_physical, 258 | args.k_latent, 259 | args.n_pcs, 260 | args.leiden_res, 261 | args.write_path, 262 | args.random_seed, 263 | args.decimals, 264 | args.n_epochs, 265 | args.learning_rate, 266 | args.alpha, 267 | args.beta, 268 | args.gamma, 269 | ) -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/gut/gut_graphst.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=400G 4 | #$ -l h_rt=04:00:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/gut/graphst/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/graphst/ 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | k_physical=50 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/graphst_.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/gut/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --k_physical $k_physical \ 28 | --k_latent $k_latent \ 29 | --leiden_res "0.2" \ 30 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/gut/graphst/reconstructions_${sample_rate}_${k_physical}physical_${k_latent}latent/ \ 31 | --random_seed $random_seed \ 32 | --n_epochs 50 \ -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/gut/gut_knn.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=48G 4 | #$ -l h_rt=02:00:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/gut/knn/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/xax 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | k_physical=50 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/knn.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/gut/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --k_physical $k_physical \ 28 | --k_latent $k_latent \ 29 | --n_pcs 50 \ 30 | --leiden_res "0.2" \ 31 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/gut/knn/reconstructions_${sample_rate}_${k_physical}physical_${k_latent}latent/ \ 32 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/gut/gut_radius.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=48G 4 | #$ -l h_rt=02:00:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/gut/radius/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/xax 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | radius=260 # ~ 50 nbrs 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/radius.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/gut/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --radius $radius \ 28 | --k_latent $k_latent \ 29 | --n_pcs 50 \ 30 | --leiden_res "0.2" \ 31 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/gut/radius/reconstructions_${sample_rate}_${radius}radius_${k_latent}latent/ \ 32 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/gut/gut_stagate.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=120G 4 | #$ -l h_rt=05:00:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/gut/stagate/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/stagate_pyg/ 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | k_physical=50 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/stagate_.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/gut/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --k_physical $k_physical \ 28 | --k_latent $k_latent \ 29 | --leiden_res "0.2" \ 30 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/gut/stagate/reconstructions_${sample_rate}_${k_physical}physical_${k_latent}latent/ \ 31 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/gut/gut_utag.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=256G 4 | #$ -l h_rt=01:30:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/gut/utag/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/bridget/envs/utag/ 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | radius=260 # ~ 50 nbrs 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/utag_.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/gut/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --radius $radius \ 28 | --k_latent $k_latent \ 29 | --n_pcs 50 \ 30 | --leiden_res "0.2" \ 31 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/gut/utag/reconstructions_${sample_rate}_${radius}radius_${k_latent}latent/ \ 32 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/gut/performance_gut.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | tissue=gut 4 | 5 | methods=(knn radius utag stagate graphst) 6 | jobids=(42437430 42437787 42437788 42437789 42437790) 7 | n_methods="${#methods[@]}" 8 | 9 | # Write job resource metrics to disk 10 | for i in $(seq 0 $((n_methods-1))); do 11 | method=${methods[i]} 12 | jobid=${jobids[i]} 13 | write_path=/stanley/WangLab/kamal/data/projects/spin/reviews/results/${tissue}/${method}/ 14 | performance=${write_path}performance_${tissue}_${method}_${jobid}.txt 15 | echo "Writing $performance" 16 | touch $performance 17 | nonsub_tasks=($(seq 11 11 110)) # only compare using non-subsampling runs 18 | for task in ${nonsub_tasks[@]}; do 19 | qacct -j $jobid -t $task >> $performance; 20 | done; 21 | done 22 | echo "Done" -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/knn.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | import argparse 4 | 5 | import scanpy as sc 6 | import numpy as np 7 | from sklearn.neighbors import NearestNeighbors 8 | from sklearn.decomposition import PCA 9 | import scipy as sp 10 | from sklearn.cluster import KMeans 11 | from sklearn.metrics import adjusted_rand_score 12 | 13 | 14 | def smooth( 15 | adata, 16 | k_physical, 17 | subsample_rate, 18 | ): 19 | 20 | # Get neighbor info 21 | nbrs = NearestNeighbors(n_neighbors=k_physical).fit(adata.obsm['spatial']) 22 | A = nbrs.kneighbors_graph() 23 | adata.obsp['spatial_connectivities'] = A.copy() 24 | 25 | # Subsample 26 | if subsample_rate: 27 | adata.obsp['spatial_connectivities_subsampled'] = A.copy() 28 | for i in range(len(adata)): 29 | nbr_idxs = adata.obsp['spatial_connectivities'][i].nonzero()[1] 30 | idxs_to_drop = np.random.choice( 31 | nbr_idxs, 32 | size=int(len(nbr_idxs)*(1-subsample_rate)), 33 | replace=False, 34 | ) 35 | for j in idxs_to_drop: 36 | adata.obsp['spatial_connectivities_subsampled'][i,j] = 0 37 | adata.obsp['spatial_connectivities_subsampled'].eliminate_zeros() 38 | else: 39 | adata.obsp['spatial_connectivities_subsampled'] = np.eye(adata.shape[0]) 40 | 41 | # Smooth 42 | adata.obsm['X_smoothed'] = adata.obsp['spatial_connectivities_subsampled'] @ adata.X 43 | 44 | 45 | def binary_search_leiden( 46 | adata, 47 | leiden_res, 48 | n_regions, 49 | jitter=0.1 50 | ): 51 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 52 | n_clusters = len(adata.obs['leiden'].unique()) 53 | lo, hi = 0, 3 54 | iters = 0 55 | while n_clusters != n_regions: 56 | if n_clusters < n_regions: 57 | lo = leiden_res 58 | leiden_res = (leiden_res+hi)/2 59 | elif n_clusters > n_regions: 60 | hi = leiden_res 61 | leiden_res = (leiden_res+lo)/2 62 | iters += 1 63 | if iters > 20: 64 | leiden_res += np.random.choice([-1,1]) * jitter 65 | # hi += jitter 66 | # lo -= max(0,jitter) 67 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 68 | n_clusters = len(adata.obs['leiden'].unique()) 69 | print(leiden_res, n_clusters, flush=True) 70 | print() 71 | 72 | 73 | def cluster( 74 | adata, 75 | n_pcs, 76 | k_latent, 77 | n_regions, 78 | leiden_res, 79 | ): 80 | 81 | # Calculate latent representation 82 | pca = PCA(n_components=n_pcs) 83 | adata.obsm['X_smoothed_pca'] = pca.fit_transform(adata.obsm['X_smoothed']) 84 | 85 | # Find latent neighbors 86 | nbrs = NearestNeighbors(n_neighbors=k_latent).fit(adata.obsm['X_smoothed_pca']) 87 | adata.obsp['smoothed_connectivities'] = nbrs.kneighbors_graph(mode='connectivity') 88 | adata.obsp['smoothed_distances'] = nbrs.kneighbors_graph(mode='distance') 89 | 90 | # Add metadata expected by Scanpy's Leiden and UMAP 91 | adata.uns['smoothed'] = {'connectivities_key': 'smoothed_connectivities', 92 | 'distances_key': 'smoothed_distances', 93 | 'params': { 94 | 'n_neighbors': k_latent, 95 | 'method': 'umap', 96 | 'random_state': 0, 97 | 'metric': 'euclidean', 98 | 'use_rep': 'X_smoothed_pca' 99 | } 100 | } 101 | 102 | # Perform k-means clustering 103 | kmeans = KMeans(n_clusters=n_regions, n_init=10) 104 | kmeans.fit(adata.obsm['X_smoothed_pca']) 105 | adata.obs['kmeans'] = kmeans.labels_.astype(str) 106 | 107 | # Perform Leiden clustering 108 | binary_search_leiden(adata, leiden_res, n_regions) 109 | 110 | 111 | def quantify(adata, n_regions): 112 | 113 | # Quantify spatial reconstruction 114 | Cp = adata.obsp['spatial_connectivities'] 115 | Cl = adata.obsp['smoothed_connectivities'] 116 | recon_score = Cp.multiply(Cl).count_nonzero() / (Cp+Cl).count_nonzero() 117 | 118 | # Quantify k-means accuracy 119 | kmeans_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['kmeans']) 120 | 121 | # Quantify Leiden accuracy 122 | leiden_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['leiden']) 123 | 124 | return recon_score, kmeans_ari, leiden_ari 125 | 126 | 127 | def main( 128 | read_path, 129 | sample_rate, 130 | k_physical, 131 | k_latent, 132 | n_pcs, 133 | leiden_res, 134 | write_path, 135 | random_seed, 136 | decimals, 137 | ): 138 | 139 | # Initialize logger 140 | logger = logging.getLogger('kNN') 141 | logger.setLevel(logging.INFO) 142 | formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') 143 | ch = logging.StreamHandler() 144 | ch.setLevel(logging.INFO) 145 | ch.setFormatter(formatter) 146 | logger.addHandler(ch) 147 | 148 | # Set random seed 149 | np.random.seed(random_seed) 150 | 151 | # Load data 152 | logger.info(f'Loading data from {read_path}') 153 | adata = sc.read_h5ad(read_path) 154 | n_regions = len(adata.obs['region_true'].unique()) 155 | 156 | # Smooth 157 | logger.info('Smoothing') 158 | logger.info(f'\tsample_rate={sample_rate}') 159 | logger.info(f'\trandom_seed={random_seed}') 160 | smooth(adata, k_physical, sample_rate) 161 | 162 | # Cluster 163 | logger.info('Clustering') 164 | cluster(adata, n_pcs, k_latent, n_regions, leiden_res) 165 | 166 | # Quantify results 167 | logger.info('Quantifying results') 168 | recon_score, kmeans_ari, leiden_ari = quantify(adata, n_regions) 169 | 170 | # Write results to disk 171 | os.makedirs(write_path, exist_ok=True) 172 | results = f'{np.round(recon_score, decimals=decimals)}_' + \ 173 | f'{np.round(kmeans_ari, decimals=decimals)}_' + \ 174 | f'{np.round(leiden_ari, decimals=decimals)}' 175 | filename_txt = f'{random_seed}_{results}.txt' 176 | write_path_txt = os.path.join(write_path, filename_txt) 177 | logger.info(f'Writing results to {write_path_txt}') 178 | f = open(write_path_txt, 'a') 179 | f.close() 180 | 181 | # Write example AnnData to disk 182 | filename_adata = f'adata_{sample_rate}.h5ad' 183 | write_path_adata = os.path.join(write_path, filename_adata) 184 | if not os.path.exists(write_path_adata): 185 | logger.info(f'Saving AnnData to {write_path_adata}') 186 | adata.write(write_path_adata) 187 | 188 | logger.info('Done') 189 | 190 | 191 | if __name__=='__main__': 192 | 193 | parser = argparse.ArgumentParser() 194 | parser.add_argument('--read_path', type=str) 195 | parser.add_argument('--sample_rate', type=float, default=None) 196 | parser.add_argument('--k_physical', type=int, default=50) 197 | parser.add_argument('--k_latent', type=int, default=15) 198 | parser.add_argument('--n_pcs', type=int) 199 | parser.add_argument('--leiden_res', type=float) 200 | parser.add_argument('--write_path', type=str) 201 | parser.add_argument('--random_seed', type=int) 202 | parser.add_argument('--decimals', type=int, default=5) 203 | args = parser.parse_args() 204 | 205 | main( 206 | args.read_path, 207 | args.sample_rate, 208 | args.k_physical, 209 | args.k_latent, 210 | args.n_pcs, 211 | args.leiden_res, 212 | args.write_path, 213 | args.random_seed, 214 | args.decimals, 215 | ) -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/modified_existing_methods/graphst/GraphST_mod.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from .preprocess import preprocess_adj, preprocess_adj_sparse, preprocess, construct_interaction, construct_interaction_KNN, add_contrastive_label, get_feature, permutation, fix_seed 3 | import time 4 | import random 5 | import numpy as np 6 | from .model import Encoder, Encoder_sparse, Encoder_map, Encoder_sc 7 | from tqdm import tqdm 8 | from torch import nn 9 | import torch.nn.functional as F 10 | from scipy.sparse.csc import csc_matrix 11 | from scipy.sparse.csr import csr_matrix 12 | import pandas as pd 13 | 14 | class GraphST(): 15 | def __init__(self, 16 | adata, 17 | adata_sc = None, 18 | device= torch.device('cpu'), 19 | learning_rate=0.001, 20 | learning_rate_sc = 0.01, 21 | weight_decay=0.00, 22 | epochs=600, 23 | dim_input=3000, 24 | dim_output=64, 25 | random_seed = 41, 26 | alpha = 10, 27 | beta = 1, 28 | gamma = 1, 29 | theta = 0.1, 30 | lamda1 = 10, 31 | lamda2 = 1, 32 | deconvolution = False, 33 | datatype = '10X' 34 | ): 35 | '''\ 36 | 37 | Parameters 38 | ---------- 39 | adata : anndata 40 | AnnData object of spatial data. 41 | adata_sc : anndata, optional 42 | AnnData object of scRNA-seq data. adata_sc is needed for deconvolution. The default is None. 43 | device : string, optional 44 | Using GPU or CPU? The default is 'cpu'. 45 | learning_rate : float, optional 46 | Learning rate for ST representation learning. The default is 0.001. 47 | learning_rate_sc : float, optional 48 | Learning rate for scRNA representation learning. The default is 0.01. 49 | weight_decay : float, optional 50 | Weight factor to control the influence of weight parameters. The default is 0.00. 51 | epochs : int, optional 52 | Epoch for model training. The default is 600. 53 | dim_input : int, optional 54 | Dimension of input feature. The default is 3000. 55 | dim_output : int, optional 56 | Dimension of output representation. The default is 64. 57 | random_seed : int, optional 58 | Random seed to fix model initialization. The default is 41. 59 | alpha : float, optional 60 | Weight factor to control the influence of reconstruction loss in representation learning. 61 | The default is 10. 62 | beta : float, optional 63 | Weight factor to control the influence of contrastive loss in representation learning. 64 | The default is 1. 65 | lamda1 : float, optional 66 | Weight factor to control the influence of reconstruction loss in mapping matrix learning. 67 | The default is 10. 68 | lamda2 : float, optional 69 | Weight factor to control the influence of contrastive loss in mapping matrix learning. 70 | The default is 1. 71 | deconvolution : bool, optional 72 | Deconvolution task? The default is False. 73 | datatype : string, optional 74 | Data type of input. Our model supports 10X Visium ('10X'), Stereo-seq ('Stereo'), and Slide-seq/Slide-seqV2 ('Slide') data. 75 | Returns 76 | ------- 77 | The learned representation 'self.emb_rec'. 78 | 79 | ''' 80 | self.adata = adata.copy() 81 | self.device = device 82 | self.learning_rate=learning_rate 83 | self.learning_rate_sc = learning_rate_sc 84 | self.weight_decay=weight_decay 85 | self.epochs=epochs 86 | self.random_seed = random_seed 87 | self.alpha = alpha 88 | self.beta = beta 89 | self.gamma = gamma 90 | self.theta = theta 91 | self.lamda1 = lamda1 92 | self.lamda2 = lamda2 93 | self.deconvolution = deconvolution 94 | self.datatype = datatype 95 | 96 | fix_seed(self.random_seed) 97 | 98 | if 'highly_variable' not in adata.var.keys(): 99 | preprocess(self.adata) 100 | 101 | if 'adj' not in adata.obsm.keys(): 102 | if self.datatype in ['Stereo', 'Slide']: 103 | construct_interaction_KNN(self.adata) 104 | else: 105 | construct_interaction(self.adata) 106 | 107 | if 'label_CSL' not in adata.obsm.keys(): 108 | add_contrastive_label(self.adata) 109 | 110 | if 'feat' not in adata.obsm.keys(): 111 | get_feature(self.adata) 112 | 113 | self.features = torch.FloatTensor(self.adata.obsm['feat'].copy()).to(self.device) 114 | self.features_a = torch.FloatTensor(self.adata.obsm['feat_a'].copy()).to(self.device) 115 | self.label_CSL = torch.FloatTensor(self.adata.obsm['label_CSL']).to(self.device) 116 | self.adj = self.adata.obsm['adj'] 117 | self.graph_neigh = torch.FloatTensor(self.adata.obsm['graph_neigh'].copy() + np.eye(self.adj.shape[0])).to(self.device) 118 | 119 | self.dim_input = self.features.shape[1] 120 | self.dim_output = dim_output 121 | 122 | if self.datatype in ['Stereo', 'Slide']: 123 | #using sparse 124 | print('Building sparse matrix ...') 125 | self.adj = preprocess_adj_sparse(self.adj).to(self.device) 126 | else: 127 | # standard version 128 | self.adj = preprocess_adj(self.adj) 129 | self.adj = torch.FloatTensor(self.adj).to(self.device) 130 | 131 | if self.deconvolution: 132 | self.adata_sc = adata_sc.copy() 133 | 134 | if isinstance(self.adata.X, csc_matrix) or isinstance(self.adata.X, csr_matrix): 135 | self.feat_sp = adata.X.toarray()[:, ] 136 | else: 137 | self.feat_sp = adata.X[:, ] 138 | if isinstance(self.adata_sc.X, csc_matrix) or isinstance(self.adata_sc.X, csr_matrix): 139 | self.feat_sc = self.adata_sc.X.toarray()[:, ] 140 | else: 141 | self.feat_sc = self.adata_sc.X[:, ] 142 | 143 | # fill nan as 0 144 | self.feat_sc = pd.DataFrame(self.feat_sc).fillna(0).values 145 | self.feat_sp = pd.DataFrame(self.feat_sp).fillna(0).values 146 | 147 | self.feat_sc = torch.FloatTensor(self.feat_sc).to(self.device) 148 | self.feat_sp = torch.FloatTensor(self.feat_sp).to(self.device) 149 | 150 | if self.adata_sc is not None: 151 | self.dim_input = self.feat_sc.shape[1] 152 | 153 | self.n_cell = adata_sc.n_obs 154 | self.n_spot = adata.n_obs 155 | 156 | def train(self): 157 | if self.datatype in ['Stereo', 'Slide']: 158 | self.model = Encoder_sparse(self.dim_input, self.dim_output, self.graph_neigh).to(self.device) 159 | else: 160 | self.model = Encoder(self.dim_input, self.dim_output, self.graph_neigh).to(self.device) 161 | self.loss_CSL = nn.BCEWithLogitsLoss() 162 | 163 | self.optimizer = torch.optim.Adam(self.model.parameters(), self.learning_rate, 164 | weight_decay=self.weight_decay) 165 | 166 | print('Begin to train ST data...') 167 | self.model.train() 168 | 169 | for epoch in tqdm(range(self.epochs)): 170 | self.model.train() 171 | 172 | self.features_a = permutation(self.features) 173 | self.hiden_feat, self.emb, ret, ret_a = self.model(self.features, self.features_a, self.adj) 174 | 175 | self.loss_sl_1 = self.loss_CSL(ret, self.label_CSL) 176 | self.loss_sl_2 = self.loss_CSL(ret_a, self.label_CSL) 177 | self.loss_feat = F.mse_loss(self.features, self.emb) 178 | 179 | loss = self.alpha*self.loss_feat + self.beta*self.loss_sl_1 + self.gamma*self.loss_sl_2 180 | 181 | self.optimizer.zero_grad() 182 | loss.backward() 183 | self.optimizer.step() 184 | 185 | print("Optimization finished for ST data!") 186 | 187 | with torch.no_grad(): 188 | self.model.eval() 189 | if self.deconvolution: 190 | self.emb_rec = self.model(self.features, self.features_a, self.adj)[1] 191 | 192 | return self.emb_rec 193 | else: 194 | if self.datatype in ['Stereo', 'Slide']: 195 | self.emb_rec = self.model(self.features, self.features_a, self.adj)[1] 196 | self.emb_rec = F.normalize(self.emb_rec, p=2, dim=1).detach().cpu().numpy() 197 | else: 198 | self.emb_rec = self.model(self.features, self.features_a, self.adj)[1].detach().cpu().numpy() 199 | self.adata.obsm['emb'] = self.emb_rec 200 | 201 | return self.adata 202 | 203 | def train_sc(self): 204 | self.model_sc = Encoder_sc(self.dim_input, self.dim_output).to(self.device) 205 | self.optimizer_sc = torch.optim.Adam(self.model_sc.parameters(), lr=self.learning_rate_sc) 206 | 207 | print('Begin to train scRNA data...') 208 | for epoch in tqdm(range(self.epochs)): 209 | self.model_sc.train() 210 | 211 | emb = self.model_sc(self.feat_sc) 212 | loss = F.mse_loss(emb, self.feat_sc) 213 | 214 | self.optimizer_sc.zero_grad() 215 | loss.backward() 216 | self.optimizer_sc.step() 217 | 218 | print("Optimization finished for cell representation learning!") 219 | 220 | with torch.no_grad(): 221 | self.model_sc.eval() 222 | emb_sc = self.model_sc(self.feat_sc) 223 | 224 | return emb_sc 225 | 226 | def train_map(self): 227 | emb_sp = self.train() 228 | emb_sc = self.train_sc() 229 | 230 | self.adata.obsm['emb_sp'] = emb_sp.detach().cpu().numpy() 231 | self.adata_sc.obsm['emb_sc'] = emb_sc.detach().cpu().numpy() 232 | 233 | # Normalize features for consistence between ST and scRNA-seq 234 | emb_sp = F.normalize(emb_sp, p=2, eps=1e-12, dim=1) 235 | emb_sc = F.normalize(emb_sc, p=2, eps=1e-12, dim=1) 236 | 237 | self.model_map = Encoder_map(self.n_cell, self.n_spot).to(self.device) 238 | 239 | self.optimizer_map = torch.optim.Adam(self.model_map.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay) 240 | 241 | print('Begin to learn mapping matrix...') 242 | for epoch in tqdm(range(self.epochs)): 243 | self.model_map.train() 244 | self.map_matrix = self.model_map() 245 | 246 | loss_recon, loss_NCE = self.loss(emb_sp, emb_sc) 247 | 248 | loss = self.lamda1*loss_recon + self.lamda2*loss_NCE 249 | 250 | self.optimizer_map.zero_grad() 251 | loss.backward() 252 | self.optimizer_map.step() 253 | 254 | print("Mapping matrix learning finished!") 255 | 256 | # take final softmax w/o computing gradients 257 | with torch.no_grad(): 258 | self.model_map.eval() 259 | emb_sp = emb_sp.cpu().numpy() 260 | emb_sc = emb_sc.cpu().numpy() 261 | map_matrix = F.softmax(self.map_matrix, dim=1).cpu().numpy() # dim=1: normalization by cell 262 | 263 | self.adata.obsm['emb_sp'] = emb_sp 264 | self.adata_sc.obsm['emb_sc'] = emb_sc 265 | self.adata.obsm['map_matrix'] = map_matrix.T # spot x cell 266 | 267 | return self.adata, self.adata_sc 268 | 269 | def loss(self, emb_sp, emb_sc): 270 | '''\ 271 | Calculate loss 272 | 273 | Parameters 274 | ---------- 275 | emb_sp : torch tensor 276 | Spatial spot representation matrix. 277 | emb_sc : torch tensor 278 | scRNA cell representation matrix. 279 | 280 | Returns 281 | ------- 282 | Loss values. 283 | 284 | ''' 285 | # cell-to-spot 286 | map_probs = F.softmax(self.map_matrix, dim=1) # dim=0: normalization by cell 287 | self.pred_sp = torch.matmul(map_probs.t(), emb_sc) 288 | 289 | loss_recon = F.mse_loss(self.pred_sp, emb_sp, reduction='mean') 290 | loss_NCE = self.Noise_Cross_Entropy(self.pred_sp, emb_sp) 291 | 292 | return loss_recon, loss_NCE 293 | 294 | def Noise_Cross_Entropy(self, pred_sp, emb_sp): 295 | '''\ 296 | Calculate noise cross entropy. Considering spatial neighbors as positive pairs for each spot 297 | 298 | Parameters 299 | ---------- 300 | pred_sp : torch tensor 301 | Predicted spatial gene expression matrix. 302 | emb_sp : torch tensor 303 | Reconstructed spatial gene expression matrix. 304 | 305 | Returns 306 | ------- 307 | loss : float 308 | Loss value. 309 | 310 | ''' 311 | 312 | mat = self.cosine_similarity(pred_sp, emb_sp) 313 | k = torch.exp(mat).sum(axis=1) - torch.exp(torch.diag(mat, 0)) 314 | 315 | # positive pairs 316 | p = torch.exp(mat) 317 | p = torch.mul(p, self.graph_neigh).sum(axis=1) 318 | 319 | ave = torch.div(p, k) 320 | loss = - torch.log(ave).mean() 321 | 322 | return loss 323 | 324 | def cosine_similarity(self, pred_sp, emb_sp): #pres_sp: spot x gene; emb_sp: spot x gene 325 | '''\ 326 | Calculate cosine similarity based on predicted and reconstructed gene expression matrix. 327 | ''' 328 | 329 | M = torch.matmul(pred_sp, emb_sp.T) 330 | Norm_c = torch.norm(pred_sp, p=2, dim=1) 331 | Norm_s = torch.norm(emb_sp, p=2, dim=1) 332 | Norm = torch.matmul(Norm_c.reshape((pred_sp.shape[0], 1)), Norm_s.reshape((emb_sp.shape[0], 1)).T) + -5e-12 333 | M = torch.div(M, Norm) 334 | 335 | if torch.any(torch.isnan(M)): 336 | M = torch.where(torch.isnan(M), torch.full_like(M, 0.4868), M) 337 | 338 | return M 339 | -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/modified_existing_methods/stagate/utils.py: -------------------------------------------------------------------------------- 1 | import pandas as pd 2 | import numpy as np 3 | import sklearn.neighbors 4 | import scipy.sparse as sp 5 | import seaborn as sns 6 | import matplotlib.pyplot as plt 7 | 8 | import torch 9 | from torch_geometric.data import Data 10 | 11 | def Transfer_pytorch_Data(adata): 12 | G_df = adata.uns['Spatial_Net'].copy() 13 | cells = np.array(adata.obs_names) 14 | cells_id_tran = dict(zip(cells, range(cells.shape[0]))) 15 | G_df['Cell1'] = G_df['Cell1'].map(cells_id_tran) 16 | G_df['Cell2'] = G_df['Cell2'].map(cells_id_tran) 17 | 18 | G = sp.coo_matrix((np.ones(G_df.shape[0]), (G_df['Cell1'], G_df['Cell2'])), shape=(adata.n_obs, adata.n_obs)) 19 | G = G + sp.eye(G.shape[0]) 20 | 21 | edgeList = np.nonzero(G) 22 | if type(adata.X) == np.ndarray: 23 | data = Data(edge_index=torch.LongTensor(np.array( 24 | [edgeList[0], edgeList[1]])), x=torch.FloatTensor(adata.X)) # .todense() 25 | else: 26 | data = Data(edge_index=torch.LongTensor(np.array( 27 | [edgeList[0], edgeList[1]])), x=torch.FloatTensor(adata.X.todense())) # .todense() 28 | return data 29 | 30 | def Batch_Data(adata, num_batch_x, num_batch_y, spatial_key=['X', 'Y'], plot_Stats=False): 31 | Sp_df = adata.obs.loc[:, spatial_key].copy() 32 | Sp_df = np.array(Sp_df) 33 | batch_x_coor = [np.percentile(Sp_df[:, 0], (1/num_batch_x)*x*100) for x in range(num_batch_x+1)] 34 | batch_y_coor = [np.percentile(Sp_df[:, 1], (1/num_batch_y)*x*100) for x in range(num_batch_y+1)] 35 | 36 | Batch_list = [] 37 | for it_x in range(num_batch_x): 38 | for it_y in range(num_batch_y): 39 | min_x = batch_x_coor[it_x] 40 | max_x = batch_x_coor[it_x+1] 41 | min_y = batch_y_coor[it_y] 42 | max_y = batch_y_coor[it_y+1] 43 | temp_adata = adata.copy() 44 | temp_adata = temp_adata[temp_adata.obs[spatial_key[0]].map(lambda x: min_x <= x <= max_x)] 45 | temp_adata = temp_adata[temp_adata.obs[spatial_key[1]].map(lambda y: min_y <= y <= max_y)] 46 | Batch_list.append(temp_adata) 47 | if plot_Stats: 48 | f, ax = plt.subplots(figsize=(1, 3)) 49 | plot_df = pd.DataFrame([x.shape[0] for x in Batch_list], columns=['#spot/batch']) 50 | sns.boxplot(y='#spot/batch', data=plot_df, ax=ax) 51 | sns.stripplot(y='#spot/batch', data=plot_df, ax=ax, color='red', size=5) 52 | return Batch_list 53 | 54 | def Cal_Spatial_Net(adata, rad_cutoff=None, k_cutoff=None, model='Radius', verbose=True, sample_rate=None): 55 | """\ 56 | Construct the spatial neighbor networks. 57 | 58 | Parameters 59 | ---------- 60 | adata 61 | AnnData object of scanpy package. 62 | rad_cutoff 63 | radius cutoff when model='Radius' 64 | k_cutoff 65 | The number of nearest neighbors when model='KNN' 66 | model 67 | The network construction model. When model=='Radius', the spot is connected to spots whose distance is less than rad_cutoff. When model=='KNN', the spot is connected to its first k_cutoff nearest neighbors. 68 | 69 | Returns 70 | ------- 71 | The spatial networks are saved in adata.uns['Spatial_Net'] 72 | """ 73 | 74 | assert(model in ['Radius', 'KNN']) 75 | if verbose: 76 | print('------Calculating spatial graph...') 77 | coor = pd.DataFrame(adata.obsm['spatial']) 78 | coor.index = adata.obs.index 79 | coor.columns = ['imagerow', 'imagecol'] 80 | 81 | if model == 'Radius': 82 | nbrs = sklearn.neighbors.NearestNeighbors(radius=rad_cutoff).fit(coor) 83 | distances, indices = nbrs.radius_neighbors(coor, return_distance=True) 84 | KNN_list = [] 85 | for it in range(indices.shape[0]): 86 | KNN_list.append(pd.DataFrame(zip([it]*indices[it].shape[0], indices[it], distances[it]))) 87 | 88 | if model == 'KNN': 89 | nbrs = sklearn.neighbors.NearestNeighbors(n_neighbors=k_cutoff+1).fit(coor) 90 | distances, indices = nbrs.kneighbors(coor) 91 | if sample_rate: 92 | #if subsample: 93 | np.random.seed(0) 94 | n_samples = int(k_cutoff * sample_rate) 95 | indices = np.array([np.random.choice(idxs, size=n_samples, replace=False) 96 | for idxs in indices]) 97 | distances = np.array([np.random.choice(dists, size=n_samples, replace=False) 98 | for dists in distances]) 99 | KNN_list = [] 100 | for it in range(indices.shape[0]): 101 | KNN_list.append(pd.DataFrame(zip([it]*indices.shape[1],indices[it,:], distances[it,:]))) 102 | 103 | KNN_df = pd.concat(KNN_list) 104 | KNN_df.columns = ['Cell1', 'Cell2', 'Distance'] 105 | 106 | Spatial_Net = KNN_df.copy() 107 | Spatial_Net = Spatial_Net.loc[Spatial_Net['Distance']>0,] 108 | id_cell_trans = dict(zip(range(coor.shape[0]), np.array(coor.index), )) 109 | Spatial_Net['Cell1'] = Spatial_Net['Cell1'].map(id_cell_trans) 110 | Spatial_Net['Cell2'] = Spatial_Net['Cell2'].map(id_cell_trans) 111 | if verbose: 112 | print('The graph contains %d edges, %d cells.' %(Spatial_Net.shape[0], adata.n_obs)) 113 | print('%.4f neighbors per cell on average.' %(Spatial_Net.shape[0]/adata.n_obs)) 114 | 115 | adata.uns['Spatial_Net'] = Spatial_Net 116 | 117 | 118 | def Cal_Spatial_Net_3D(adata, rad_cutoff_2D, rad_cutoff_Zaxis, 119 | key_section='Section_id', section_order=None, verbose=True): 120 | """\ 121 | Construct the spatial neighbor networks. 122 | 123 | Parameters 124 | ---------- 125 | adata 126 | AnnData object of scanpy package. 127 | rad_cutoff_2D 128 | radius cutoff for 2D SNN construction. 129 | rad_cutoff_Zaxis 130 | radius cutoff for 2D SNN construction for consturcting SNNs between adjacent sections. 131 | key_section 132 | The columns names of section_ID in adata.obs. 133 | section_order 134 | The order of sections. The SNNs between adjacent sections are constructed according to this order. 135 | 136 | Returns 137 | ------- 138 | The 3D spatial networks are saved in adata.uns['Spatial_Net']. 139 | """ 140 | adata.uns['Spatial_Net_2D'] = pd.DataFrame() 141 | adata.uns['Spatial_Net_Zaxis'] = pd.DataFrame() 142 | num_section = np.unique(adata.obs[key_section]).shape[0] 143 | if verbose: 144 | print('Radius used for 2D SNN:', rad_cutoff_2D) 145 | print('Radius used for SNN between sections:', rad_cutoff_Zaxis) 146 | for temp_section in np.unique(adata.obs[key_section]): 147 | if verbose: 148 | print('------Calculating 2D SNN of section ', temp_section) 149 | temp_adata = adata[adata.obs[key_section] == temp_section, ] 150 | Cal_Spatial_Net( 151 | temp_adata, rad_cutoff=rad_cutoff_2D, verbose=False) 152 | temp_adata.uns['Spatial_Net']['SNN'] = temp_section 153 | if verbose: 154 | print('This graph contains %d edges, %d cells.' % 155 | (temp_adata.uns['Spatial_Net'].shape[0], temp_adata.n_obs)) 156 | print('%.4f neighbors per cell on average.' % 157 | (temp_adata.uns['Spatial_Net'].shape[0]/temp_adata.n_obs)) 158 | adata.uns['Spatial_Net_2D'] = pd.concat( 159 | [adata.uns['Spatial_Net_2D'], temp_adata.uns['Spatial_Net']]) 160 | for it in range(num_section-1): 161 | section_1 = section_order[it] 162 | section_2 = section_order[it+1] 163 | if verbose: 164 | print('------Calculating SNN between adjacent section %s and %s.' % 165 | (section_1, section_2)) 166 | Z_Net_ID = section_1+'-'+section_2 167 | temp_adata = adata[adata.obs[key_section].isin( 168 | [section_1, section_2]), ] 169 | Cal_Spatial_Net( 170 | temp_adata, rad_cutoff=rad_cutoff_Zaxis, verbose=False) 171 | spot_section_trans = dict( 172 | zip(temp_adata.obs.index, temp_adata.obs[key_section])) 173 | temp_adata.uns['Spatial_Net']['Section_id_1'] = temp_adata.uns['Spatial_Net']['Cell1'].map( 174 | spot_section_trans) 175 | temp_adata.uns['Spatial_Net']['Section_id_2'] = temp_adata.uns['Spatial_Net']['Cell2'].map( 176 | spot_section_trans) 177 | used_edge = temp_adata.uns['Spatial_Net'].apply( 178 | lambda x: x['Section_id_1'] != x['Section_id_2'], axis=1) 179 | temp_adata.uns['Spatial_Net'] = temp_adata.uns['Spatial_Net'].loc[used_edge, ] 180 | temp_adata.uns['Spatial_Net'] = temp_adata.uns['Spatial_Net'].loc[:, [ 181 | 'Cell1', 'Cell2', 'Distance']] 182 | temp_adata.uns['Spatial_Net']['SNN'] = Z_Net_ID 183 | if verbose: 184 | print('This graph contains %d edges, %d cells.' % 185 | (temp_adata.uns['Spatial_Net'].shape[0], temp_adata.n_obs)) 186 | print('%.4f neighbors per cell on average.' % 187 | (temp_adata.uns['Spatial_Net'].shape[0]/temp_adata.n_obs)) 188 | adata.uns['Spatial_Net_Zaxis'] = pd.concat( 189 | [adata.uns['Spatial_Net_Zaxis'], temp_adata.uns['Spatial_Net']]) 190 | adata.uns['Spatial_Net'] = pd.concat( 191 | [adata.uns['Spatial_Net_2D'], adata.uns['Spatial_Net_Zaxis']]) 192 | if verbose: 193 | print('3D SNN contains %d edges, %d cells.' % 194 | (adata.uns['Spatial_Net'].shape[0], adata.n_obs)) 195 | print('%.4f neighbors per cell on average.' % 196 | (adata.uns['Spatial_Net'].shape[0]/adata.n_obs)) 197 | 198 | def Stats_Spatial_Net(adata): 199 | import matplotlib.pyplot as plt 200 | Num_edge = adata.uns['Spatial_Net']['Cell1'].shape[0] 201 | Mean_edge = Num_edge/adata.shape[0] 202 | plot_df = pd.value_counts(pd.value_counts(adata.uns['Spatial_Net']['Cell1'])) 203 | plot_df = plot_df/adata.shape[0] 204 | fig, ax = plt.subplots(figsize=[3,2]) 205 | plt.ylabel('Percentage') 206 | plt.xlabel('') 207 | plt.title('Number of Neighbors (Mean=%.2f)'%Mean_edge) 208 | ax.bar(plot_df.index, plot_df) 209 | 210 | def mclust_R(adata, num_cluster, modelNames='EEE', used_obsm='STAGATE', random_seed=2020): 211 | """\ 212 | Clustering using the mclust algorithm. 213 | The parameters are the same as those in the R package mclust. 214 | """ 215 | 216 | np.random.seed(random_seed) 217 | import rpy2.robjects as robjects 218 | robjects.r.library("mclust") 219 | 220 | import rpy2.robjects.numpy2ri 221 | rpy2.robjects.numpy2ri.activate() 222 | r_random_seed = robjects.r['set.seed'] 223 | r_random_seed(random_seed) 224 | rmclust = robjects.r['Mclust'] 225 | 226 | res = rmclust(rpy2.robjects.numpy2ri.numpy2rpy(adata.obsm[used_obsm]), num_cluster, modelNames) 227 | mclust_res = np.array(res[-2]) 228 | 229 | adata.obs['mclust'] = mclust_res 230 | adata.obs['mclust'] = adata.obs['mclust'].astype('int') 231 | adata.obs['mclust'] = adata.obs['mclust'].astype('category') 232 | return adata 233 | -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/modified_existing_methods/utag/segmentation.py: -------------------------------------------------------------------------------- 1 | import typing as tp 2 | import warnings 3 | import os 4 | 5 | import scanpy as sc 6 | import squidpy as sq 7 | import numpy as np 8 | import pandas as pd 9 | import matplotlib.pyplot as plt 10 | from tqdm import tqdm 11 | import anndata 12 | import parmap 13 | 14 | from utag.types import Path, Array, AnnData 15 | from utag.utils import sparse_matrix_dstack 16 | 17 | 18 | def utag_mod( 19 | adata: AnnData, 20 | random_state: int, 21 | channels_to_use: tp.Sequence[str] = None, 22 | slide_key: tp.Optional[str] = "Slide", 23 | save_key: str = "UTAG Label", 24 | filter_by_variance: bool = False, 25 | max_dist: float = 0.1, 26 | normalization_mode: str = "l1_norm", 27 | keep_spatial_connectivity: bool = False, 28 | pca_kwargs: tp.Dict[str, tp.Any] = dict(n_comps=10), 29 | apply_umap: bool = False, 30 | umap_kwargs: tp.Dict[str, tp.Any] = dict(), 31 | apply_clustering: bool = True, 32 | clustering_method: tp.Sequence[str] = ["leiden", "kmeans"], 33 | resolution: float = 0.5, 34 | n_regions: int = 4, 35 | leiden_kwargs: tp.Dict[str, tp.Any] = None, 36 | parallel: bool = True, 37 | processes: int = None, 38 | sample_rate: float = 1.0 39 | ) -> AnnData: 40 | """ 41 | Discover tissue architechture in single-cell imaging data 42 | by combining phenotypes and positional information of cells. 43 | 44 | Parameters 45 | ---------- 46 | adata: AnnData 47 | AnnData object with spatial positioning of cells in obsm 'spatial' slot. 48 | channels_to_use: Optional[Sequence[str]] 49 | An optional sequence of strings used to subset variables to use. 50 | Default (None) is to use all variables. 51 | max_dist: float 52 | Maximum distance to cut edges within a graph. 53 | Should be adjusted depending on resolution of images. 54 | For imaging mass cytometry, where resolution is 1um, 20 often gives good results. 55 | Default is 20. 56 | slide_key: {str, None} 57 | Key of adata.obs containing information on the batch structure of the data. 58 | In general, for image data this will often be a variable indicating the image 59 | so image-specific effects are removed from data. 60 | Default is "Slide". 61 | save_key: str 62 | Key to be added to adata object holding the UTAG clusters. 63 | Depending on the values of `clustering_method` and `resolutions`, 64 | the final keys will be of the form: {save_key}_{method}_{resolution}". 65 | Default is "UTAG Label". 66 | filter_by_variance: bool 67 | Whether to filter vairiables by variance. 68 | Default is False, which keeps all variables. 69 | max_dist: float 70 | Recommended values are between 20 to 50 depending on magnification. 71 | Default is 20. 72 | normalization_mode: str 73 | Method to normalize adjacency matrix. 74 | Default is "l1_norm", any other value will not use normalization. 75 | keep_spatial_connectivity: bool 76 | Whether to keep sparse matrices of spatial connectivity and distance in the obsp attribute of the 77 | resulting anndata object. This could be useful in downstream applications. 78 | Default is not to (False). 79 | pca_kwargs: Dict[str, Any] 80 | Keyword arguments to be passed to scanpy.pp.pca for dimensionality reduction after message passing. 81 | Default is to pass n_comps=10, which uses 10 Principal Components. 82 | apply_umap: bool 83 | Whether to build a UMAP representation after message passing. 84 | Default is False. 85 | umap_kwargs: Dict[str, Any] 86 | Keyword arguments to be passed to scanpy.tl.umap for dimensionality reduction after message passing. 87 | Default is 10.0. 88 | apply_clustering: bool 89 | Whether to cluster the message passed matrix. 90 | Default is True. 91 | resolution: float 92 | What resolution should the methods in `clustering_method` be run at. 93 | leiden_kwargs: dict[str, Any] 94 | Keyword arguments to pass to scanpy.tl.leiden. 95 | parallel: bool 96 | Whether to run message passing part of algorithm in parallel. 97 | Will accelerate the process but consume more memory. 98 | Default is True. 99 | processes: int 100 | Number of processes to use in parallel. 101 | Default is to use all available (-1). 102 | 103 | Returns 104 | ------- 105 | adata: AnnData 106 | AnnData object with UTAG domain predictions for each cell in adata.obs, column `save_key`. 107 | """ 108 | ad = adata.copy() 109 | np.random.seed(random_state) 110 | 111 | if channels_to_use: 112 | ad = ad[:, channels_to_use] 113 | 114 | if filter_by_variance: 115 | ad = low_variance_filter(ad) 116 | 117 | if isinstance(clustering_method, list): 118 | clustering_method = [m.upper() for m in clustering_method] 119 | elif isinstance(clustering_method, str): 120 | clustering_method = [clustering_method.upper()] 121 | else: 122 | print( 123 | "Invalid Clustering Method. Clustering Method Should Either be a string or a list" 124 | ) 125 | return 126 | assert all(m in ["LEIDEN", "PARC", "KMEANS"] for m in clustering_method) 127 | 128 | if "PARC" in clustering_method: 129 | from parc import PARC # early fail if not available 130 | if "KMEANS" in clustering_method: 131 | from sklearn.cluster import KMeans 132 | 133 | print("Applying UTAG Algorithm...", flush=True) 134 | if slide_key: 135 | ads = [ 136 | ad[ad.obs[slide_key] == slide].copy() for slide in ad.obs[slide_key].unique() 137 | ] 138 | ad_list = parmap.map( 139 | _parallel_message_pass, 140 | ads, 141 | radius=max_dist, 142 | 143 | coord_type="generic", 144 | set_diag=True, 145 | mode=normalization_mode, 146 | pm_pbar=True, 147 | pm_parallel=parallel, 148 | pm_processes=processes, 149 | ) 150 | ad_result = anndata.concat(ad_list) 151 | if keep_spatial_connectivity: 152 | ad_result.obsp["spatial_connectivities"] = sparse_matrix_dstack( 153 | [x.obsp["spatial_connectivities"] for x in ad_list] 154 | ) 155 | ad_result.obsp["spatial_distances"] = sparse_matrix_dstack( 156 | [x.obsp["spatial_distances"] for x in ad_list] 157 | ) 158 | else: 159 | sq.gr.spatial_neighbors(ad, radius=max_dist, coord_type="generic", set_diag=True) 160 | ad_result = custom_message_passing(ad, mode=normalization_mode, sample_rate=sample_rate) 161 | 162 | if apply_clustering: 163 | if "n_comps" in pca_kwargs: 164 | if pca_kwargs["n_comps"] > ad_result.shape[1]: 165 | pca_kwargs["n_comps"] = ad_result.shape[1] - 1 166 | print( 167 | f"Overwriding provided number of PCA dimensions to match number of features: {pca_kwargs['n_comps']}" 168 | ) 169 | pca_kwargs["n_comps"] = int(pca_kwargs["n_comps"]) 170 | sc.tl.pca(ad_result, **pca_kwargs) 171 | sc.pp.neighbors(ad_result) 172 | 173 | if apply_umap: 174 | print("Running UMAP on Input Dataset...", flush=True) 175 | sc.tl.umap(ad_result, **umap_kwargs) 176 | 177 | res_key1 = "leiden" 178 | res_key3 = "kmeans" 179 | 180 | if "LEIDEN" in clustering_method: 181 | 182 | def binary_search_leiden(adata, leiden_res, n_regions, kwargs): 183 | sc.tl.leiden(adata, resolution=leiden_res) 184 | n_clusters = len(adata.obs['leiden'].unique()) 185 | lo, hi = 0, 5 186 | while n_clusters != n_regions: 187 | if n_clusters < n_regions: 188 | lo = leiden_res 189 | leiden_res = (leiden_res+hi)/2 190 | elif n_clusters > n_regions: 191 | hi = leiden_res 192 | leiden_res = (leiden_res+lo)/2 193 | sc.tl.leiden(adata, resolution=leiden_res) 194 | n_clusters = len(adata.obs['leiden'].unique()) 195 | print(leiden_res, n_clusters, flush=True) 196 | 197 | print(f"Applying Leiden Clustering...", flush=True) 198 | kwargs = dict() 199 | kwargs.update(leiden_kwargs or {}) 200 | 201 | binary_search_leiden(ad_result, resolution, n_regions, kwargs) 202 | 203 | add_probabilities_to_centroid(ad_result, res_key1) 204 | 205 | if "KMEANS" in clustering_method: 206 | print(f"Applying K-means Clustering...", flush=True) 207 | kmeans = KMeans(n_clusters=n_regions, random_state=random_state).fit(ad_result.obsm["X_pca"]) 208 | ad_result.obs[res_key3] = pd.Categorical(kmeans.labels_.astype(str)) 209 | add_probabilities_to_centroid(ad_result, res_key3) 210 | 211 | return ad_result 212 | 213 | 214 | def _parallel_message_pass( 215 | ad: AnnData, 216 | radius: float, 217 | coord_type: str, 218 | set_diag: bool, 219 | mode: str, 220 | sample_rate: float, 221 | ): 222 | sq.gr.spatial_neighbors(ad, radius=radius, coord_type=coord_type, set_diag=set_diag) 223 | ad = custom_message_passing(ad, mode=mode, sample_rate=sample_rate) 224 | return ad 225 | 226 | 227 | def custom_message_passing(adata: AnnData, mode: str = "l1_norm", sample_rate=1) -> AnnData: 228 | from scipy.linalg import sqrtm 229 | from scipy.sparse import csr_matrix 230 | 231 | if sample_rate == 0: 232 | adata.obsp['spatial_connectivities'] = np.eye(adata.shape[0]) 233 | #print(adata.obsp['spatial_connectivities'], flush=0) 234 | elif sample_rate < 1: 235 | A = adata.obsp['spatial_connectivities'] 236 | n_cells = A.shape[0] 237 | A_subsampled = np.zeros((n_cells,n_cells)) 238 | 239 | for i in range(n_cells): 240 | nbr_idxs = A[i].nonzero()[1] 241 | n_samples = int(len(nbr_idxs) * sample_rate) 242 | nbr_idxs_subsampled = np.random.choice(nbr_idxs, size=n_samples, replace=False) 243 | for j in nbr_idxs_subsampled: 244 | A_subsampled[i,j] = 1 245 | print(len(nbr_idxs), flush=True) 246 | print(n_samples, flush=True) 247 | adata.obsp['spatial_connectivities'] = A_subsampled 248 | 249 | adata.obsp['spatial_connectivities'] = csr_matrix(adata.obsp['spatial_connectivities']) 250 | 251 | if mode == "l1_norm": 252 | A = adata.obsp["spatial_connectivities"] 253 | A_mod = np.asarray(A + np.eye(A.shape[0])) 254 | from sklearn.preprocessing import normalize 255 | affinity = normalize(A_mod, axis=1, norm="l1") 256 | else: 257 | # Plain A_mod multiplication 258 | A = adata.obsp["spatial_connectivities"] 259 | affinity = A 260 | 261 | adata.obsm['X_smoothed'] = affinity @ adata.X 262 | return adata 263 | 264 | 265 | def low_variance_filter(adata: AnnData) -> AnnData: 266 | return adata[:, adata.var["std"] > adata.var["std"].median()] 267 | 268 | 269 | def add_probabilities_to_centroid( 270 | adata: AnnData, col: str, name_to_output: str = None 271 | ) -> AnnData: 272 | from utag.utils import z_score 273 | from scipy.special import softmax 274 | 275 | if name_to_output is None: 276 | name_to_output = col + "_probabilities" 277 | 278 | mean = z_score(adata.to_df()).groupby(adata.obs[col]).mean() 279 | probs = softmax(adata.to_df() @ mean.T, axis=1) 280 | adata.obsm[name_to_output] = probs 281 | return adata 282 | 283 | 284 | def evaluate_performance( 285 | adata: AnnData, 286 | batch_key: str = "Slide", 287 | truth_key: str = "DOM_argmax", 288 | pred_key: str = "cluster", 289 | method: str = "rand", 290 | ) -> Array: 291 | assert method in ["rand", "homogeneity"] 292 | from sklearn.metrics import rand_score, homogeneity_score 293 | 294 | score_list = [] 295 | for key in adata.obs[batch_key].unique(): 296 | batch = adata[adata.obs[batch_key] == key] 297 | if method == "rand": 298 | score = rand_score(batch.obs[truth_key], batch.obs[pred_key]) 299 | elif method == "homogeneity": 300 | score = homogeneity_score(batch.obs[truth_key], batch.obs[pred_key]) 301 | score_list.append(score) 302 | return score_list 303 | -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/radius.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | import argparse 4 | 5 | import scanpy as sc 6 | import numpy as np 7 | from sklearn.neighbors import NearestNeighbors 8 | from sklearn.decomposition import PCA 9 | import scipy as sp 10 | from sklearn.cluster import KMeans 11 | from sklearn.metrics import adjusted_rand_score 12 | 13 | 14 | def smooth( 15 | adata, 16 | radius, 17 | subsample_rate, 18 | ): 19 | 20 | # Get neighbor info 21 | nbrs = NearestNeighbors(radius=radius).fit(adata.obsm['spatial']) 22 | A = nbrs.radius_neighbors_graph() 23 | adata.obsp['spatial_connectivities'] = A.copy() 24 | 25 | # Subsample 26 | if subsample_rate: 27 | adata.obsp['spatial_connectivities_subsampled'] = A.copy() 28 | for i in range(len(adata)): 29 | nbr_idxs = adata.obsp['spatial_connectivities'][i].nonzero()[1] 30 | idxs_to_drop = np.random.choice( 31 | nbr_idxs, 32 | size=int(len(nbr_idxs)*(1-subsample_rate)), 33 | replace=False, 34 | ) 35 | for j in idxs_to_drop: 36 | adata.obsp['spatial_connectivities_subsampled'][i,j] = 0 37 | adata.obsp['spatial_connectivities_subsampled'].eliminate_zeros() 38 | else: 39 | adata.obsp['spatial_connectivities_subsampled'] = np.eye(adata.shape[0]) 40 | 41 | # Smooth 42 | adata.obsm['X_smoothed'] = adata.obsp['spatial_connectivities_subsampled'] @ adata.X 43 | 44 | 45 | def binary_search_leiden( 46 | adata, 47 | leiden_res, 48 | n_regions, 49 | jitter=0.1 50 | ): 51 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 52 | n_clusters = len(adata.obs['leiden'].unique()) 53 | lo, hi = 0, 5 54 | iters = 0 55 | while n_clusters != n_regions: 56 | if n_clusters < n_regions: 57 | lo = leiden_res 58 | leiden_res = (leiden_res+hi)/2 59 | elif n_clusters > n_regions: 60 | hi = leiden_res 61 | leiden_res = (leiden_res+lo)/2 62 | iters += 1 63 | if iters > 10: 64 | leiden_res += np.random.choice([-1,1]) * jitter 65 | hi += jitter 66 | lo -= max(0,jitter) 67 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 68 | n_clusters = len(adata.obs['leiden'].unique()) 69 | print(leiden_res, n_clusters, flush=True) 70 | print() 71 | 72 | 73 | def cluster( 74 | adata, 75 | n_pcs, 76 | k_latent, 77 | n_regions, 78 | leiden_res, 79 | ): 80 | 81 | # Calculate latent representation 82 | pca = PCA(n_components=n_pcs) 83 | adata.obsm['X_smoothed_pca'] = pca.fit_transform(adata.obsm['X_smoothed']) 84 | 85 | # Find latent neighbors 86 | nbrs = NearestNeighbors(n_neighbors=k_latent).fit(adata.obsm['X_smoothed_pca']) 87 | adata.obsp['smoothed_connectivities'] = nbrs.kneighbors_graph(mode='connectivity') 88 | adata.obsp['smoothed_distances'] = nbrs.kneighbors_graph(mode='distance') 89 | 90 | # Add metadata expected by Scanpy's Leiden and UMAP 91 | adata.uns['smoothed'] = {'connectivities_key': 'smoothed_connectivities', 92 | 'distances_key': 'smoothed_distances', 93 | 'params': { 94 | 'n_neighbors': k_latent, 95 | 'method': 'umap', 96 | 'random_state': 0, 97 | 'metric': 'euclidean', 98 | 'use_rep': 'X_smoothed_pca' 99 | } 100 | } 101 | 102 | # Perform k-means clustering 103 | kmeans = KMeans(n_clusters=n_regions, n_init=10) 104 | kmeans.fit(adata.obsm['X_smoothed_pca']) 105 | adata.obs['kmeans'] = kmeans.labels_.astype(str) 106 | 107 | # Perform Leiden clustering 108 | binary_search_leiden(adata, leiden_res, n_regions) 109 | 110 | 111 | def quantify(adata, n_regions): 112 | 113 | # Quantify spatial reconstruction 114 | Cp = adata.obsp['spatial_connectivities'] 115 | Cl = adata.obsp['smoothed_connectivities'] 116 | recon_score = Cp.multiply(Cl).count_nonzero() / (Cp+Cl).count_nonzero() 117 | 118 | # Quantify k-means accuracy 119 | kmeans_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['kmeans']) 120 | 121 | # Quantify Leiden accuracy 122 | leiden_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['leiden']) 123 | 124 | return recon_score, kmeans_ari, leiden_ari 125 | 126 | 127 | def main( 128 | read_path, 129 | sample_rate, 130 | radius, 131 | k_latent, 132 | n_pcs, 133 | leiden_res, 134 | write_path, 135 | random_seed, 136 | decimals, 137 | ): 138 | 139 | # Initialize logger 140 | logger = logging.getLogger('Radius') 141 | logger.setLevel(logging.INFO) 142 | formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') 143 | ch = logging.StreamHandler() 144 | ch.setLevel(logging.INFO) 145 | ch.setFormatter(formatter) 146 | logger.addHandler(ch) 147 | 148 | # Set random seed 149 | np.random.seed(random_seed) 150 | 151 | # Load data 152 | logger.info(f'Loading data from {read_path}') 153 | adata = sc.read_h5ad(read_path) 154 | n_regions = len(adata.obs['region_true'].unique()) 155 | 156 | # Smooth 157 | logger.info('Smoothing') 158 | logger.info(f'\tsample_rate={sample_rate}') 159 | logger.info(f'\trandom_seed={random_seed}') 160 | smooth(adata, radius, sample_rate) 161 | 162 | # Cluster 163 | logger.info('Clustering') 164 | cluster(adata, n_pcs, k_latent, n_regions, leiden_res) 165 | 166 | # Quantify results 167 | logger.info('Quantifying results') 168 | recon_score, kmeans_ari, leiden_ari = quantify(adata, n_regions) 169 | 170 | # Write results to disk 171 | os.makedirs(write_path, exist_ok=True) 172 | results = f'{np.round(recon_score, decimals=decimals)}_' + \ 173 | f'{np.round(kmeans_ari, decimals=decimals)}_' + \ 174 | f'{np.round(leiden_ari, decimals=decimals)}' 175 | filename_txt = f'{random_seed}_{results}.txt' 176 | write_path_txt = os.path.join(write_path, filename_txt) 177 | logger.info(f'Writing results to {write_path_txt}') 178 | f = open(write_path_txt, 'a') 179 | f.close() 180 | 181 | # Write example AnnData to disk 182 | if random_seed == 0: 183 | filename_adata = f'adata_{sample_rate}.h5ad' 184 | write_path_adata = os.path.join(write_path, filename_adata) 185 | logger.info(f'Saving AnnData to {write_path_adata}') 186 | adata.write(write_path_adata) 187 | 188 | logger.info('Done') 189 | 190 | 191 | if __name__=='__main__': 192 | 193 | parser = argparse.ArgumentParser() 194 | parser.add_argument('--read_path', type=str) 195 | parser.add_argument('--sample_rate', type=float, default=None) 196 | parser.add_argument('--radius', type=float, default=None) 197 | parser.add_argument('--k_latent', type=int, default=50) 198 | parser.add_argument('--n_pcs', type=int) 199 | parser.add_argument('--leiden_res', type=float) 200 | parser.add_argument('--write_path', type=str) 201 | parser.add_argument('--random_seed', type=int) 202 | parser.add_argument('--decimals', type=int, default=5) 203 | args = parser.parse_args() 204 | 205 | main( 206 | args.read_path, 207 | args.sample_rate, 208 | args.radius, 209 | args.k_latent, 210 | args.n_pcs, 211 | args.leiden_res, 212 | args.write_path, 213 | args.random_seed, 214 | args.decimals, 215 | ) -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/simulation/performance_simulation.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | tissue=simulation 4 | 5 | simulation_methods=(knn radius utag stagate graphst) 6 | simulation_jobids=(42429477 42429759 42429768 42430818 42430894) 7 | n_methods="${#simulation_methods[@]}" 8 | 9 | # Write job resource metrics to disk 10 | for i in $(seq 0 $((n_methods-1))); do 11 | method=${simulation_methods[i]} 12 | jobid=${simulation_jobids[i]} 13 | write_path=/stanley/WangLab/kamal/data/projects/spin/reviews/results/${tissue}/${method}/ 14 | performance=${write_path}performance_${tissue}_${method}_${jobid}.txt 15 | echo "Writing $performance" 16 | touch $performance 17 | nonsub_tasks=($(seq 11 11 110)) # only compare using non-subsampling runs 18 | for task in ${nonsub_tasks[@]}; do 19 | qacct -j $jobid -t $task >> $performance; 20 | done; 21 | done 22 | echo "Done" -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/simulation/simulation_graphst.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=16G 4 | #$ -l h_rt=00:10:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/simulation/graphst/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/graphst/ 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | k_physical=50 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/graphst_.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/simulation/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --k_physical $k_physical \ 28 | --k_latent $k_latent \ 29 | --leiden_res "0.2" \ 30 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/simulation/graphst/reconstructions_${sample_rate}_${k_physical}physical_${k_latent}latent/ \ 31 | --random_seed $random_seed \ 32 | --n_epochs 50 \ -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/simulation/simulation_knn.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=8G 4 | #$ -l h_rt=00:30:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/simulation/knn/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/xax 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | k_physical=50 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/knn.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/simulation/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --k_physical $k_physical \ 28 | --k_latent $k_latent \ 29 | --n_pcs 50 \ 30 | --leiden_res "0.2" \ 31 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/simulation/knn/reconstructions_${sample_rate}_${k_physical}physical_${k_latent}latent/ \ 32 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/simulation/simulation_radius.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=8G 4 | #$ -l h_rt=00:30:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/simulation/radius/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/xax 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | radius=0.1 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/radius.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/simulation/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --radius $radius \ 28 | --k_latent $k_latent \ 29 | --n_pcs 50 \ 30 | --leiden_res "0.2" \ 31 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/simulation/radius/reconstructions_${sample_rate}_${radius}radius_${k_latent}latent/ \ 32 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/simulation/simulation_stagate.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=16G 4 | #$ -l h_rt=00:10:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/simulation/stagate/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/kamal/envs/stagate_pyg/ 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | k_physical=50 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/stagate_.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/simulation/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --k_physical $k_physical \ 28 | --k_latent $k_latent \ 29 | --leiden_res "0.2" \ 30 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/simulation/stagate/reconstructions_${sample_rate}_${k_physical}physical_${k_latent}latent/ \ 31 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/simulation/simulation_utag.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #$ -l h_vmem=16G 4 | #$ -l h_rt=00:10:00 5 | #$ -o /stanley/WangLab/kamal/outputs/spin/reviews/simulation/utag/ 6 | #$ -j y 7 | #$ -t 1-110 8 | 9 | source /broad/software/scripts/useuse 10 | reuse Anaconda3 11 | source activate /stanley/WangLab/bridget/envs/utag/ 12 | 13 | sample_rates=($(seq 0 0.1 1)) 14 | random_seeds=($(seq 0 1 10)) 15 | # random_seeds=($(seq 0 0)) # test using a single random seed 16 | 17 | n_sample_rates=${#sample_rates[@]} 18 | sample_rate=${sample_rates[($((SGE_TASK_ID-1))%$n_sample_rates)]} 19 | random_seed=${random_seeds[($((SGE_TASK_ID-1))/$n_sample_rates)]} 20 | 21 | radius=0.1 22 | k_latent=15 23 | 24 | python /stanley/WangLab/kamal/code/projects/spin/reviews/scripts/utag_.py \ 25 | --read_path /stanley/WangLab/kamal/data/projects/spin/reviews/simulation/adata.h5ad \ 26 | --sample_rate $sample_rate \ 27 | --radius $radius \ 28 | --k_latent $k_latent \ 29 | --n_pcs 50 \ 30 | --leiden_res "0.2" \ 31 | --write_path /stanley/WangLab/kamal/data/projects/spin/reviews/results/simulation/utag/reconstructions_${sample_rate}_${radius}radius_${k_latent}latent/ \ 32 | --random_seed $random_seed -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/stagate_.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | import argparse 4 | 5 | import STAGATE_pyG_sub 6 | import scanpy as sc 7 | import numpy as np 8 | from sklearn.neighbors import NearestNeighbors 9 | from sklearn.decomposition import PCA 10 | import scipy as sp 11 | from sklearn.cluster import KMeans 12 | from sklearn.metrics import adjusted_rand_score 13 | 14 | 15 | def smooth( 16 | adata, 17 | k_physical, 18 | sample_rate, 19 | n_epochs, 20 | alpha, 21 | random_seed, 22 | ): 23 | 24 | # Get neighbors 25 | STAGATE_pyG_sub.Cal_Spatial_Net( 26 | adata, 27 | model='KNN', 28 | k_cutoff=k_physical, 29 | sample_rate=sample_rate, 30 | ) 31 | 32 | # Train model 33 | adata = STAGATE_pyG_sub.train_STAGATE( 34 | adata, 35 | n_epochs=n_epochs, 36 | key_added='stagate', 37 | alpha=alpha, 38 | random_seed=random_seed, 39 | ) 40 | 41 | return adata 42 | 43 | 44 | def binary_search_leiden( 45 | adata, 46 | leiden_res, 47 | n_regions, 48 | jitter=0.1 49 | ): 50 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 51 | n_clusters = len(adata.obs['leiden'].unique()) 52 | lo, hi = 0, 5 53 | iters = 0 54 | while n_clusters != n_regions: 55 | if n_clusters < n_regions: 56 | lo = leiden_res 57 | leiden_res = (leiden_res+hi)/2 58 | elif n_clusters > n_regions: 59 | hi = leiden_res 60 | leiden_res = (leiden_res+lo)/2 61 | iters += 1 62 | if iters > 10: 63 | leiden_res += np.random.choice([-1,1]) * jitter 64 | hi += jitter 65 | lo -= max(0,jitter) 66 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 67 | n_clusters = len(adata.obs['leiden'].unique()) 68 | print(leiden_res, n_clusters, flush=True) 69 | print() 70 | 71 | 72 | def cluster( 73 | adata, 74 | n_pcs, 75 | k_latent, 76 | n_regions, 77 | leiden_res, 78 | ): 79 | 80 | # Find latent neighbors 81 | nbrs = NearestNeighbors(n_neighbors=k_latent).fit(adata.obsm['stagate']) 82 | adata.obsp['smoothed_connectivities'] = nbrs.kneighbors_graph(mode='connectivity') 83 | adata.obsp['smoothed_distances'] = nbrs.kneighbors_graph(mode='distance') 84 | 85 | # Add metadata expected by Scanpy's Leiden and UMAP 86 | adata.uns['smoothed'] = {'connectivities_key': 'smoothed_connectivities', 87 | 'distances_key': 'smoothed_distances', 88 | 'params': { 89 | 'n_neighbors': k_latent, 90 | 'method': 'umap', 91 | 'random_state': 0, 92 | 'metric': 'euclidean', 93 | 'use_rep': 'stagate' 94 | } 95 | } 96 | 97 | # Perform k-means clustering 98 | kmeans = KMeans(n_clusters=n_regions, n_init=10) 99 | kmeans.fit(adata.obsm['stagate']) 100 | adata.obs['kmeans'] = kmeans.labels_.astype(str) 101 | 102 | # Perform Leiden clustering 103 | binary_search_leiden(adata, leiden_res, n_regions) 104 | 105 | 106 | def quantify(adata, n_regions): 107 | 108 | # Quantify spatial reconstruction 109 | Cp = adata.obsp['spatial_connectivities'] 110 | Cl = adata.obsp['smoothed_connectivities'] 111 | recon_score = Cp.multiply(Cl).count_nonzero() / (Cp+Cl).count_nonzero() 112 | 113 | # Quantify k-means accuracy 114 | kmeans_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['kmeans']) 115 | 116 | # Quantify Leiden accuracy 117 | leiden_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['leiden']) 118 | 119 | return recon_score, kmeans_ari, leiden_ari 120 | 121 | 122 | def main( 123 | read_path, 124 | sample_rate, 125 | k_physical, 126 | k_latent, 127 | n_pcs, 128 | leiden_res, 129 | write_path, 130 | random_seed, 131 | decimals, 132 | n_epochs, 133 | alpha, 134 | ): 135 | 136 | # Initialize logger 137 | logger = logging.getLogger('STAGATE') 138 | logger.setLevel(logging.INFO) 139 | formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') 140 | ch = logging.StreamHandler() 141 | ch.setLevel(logging.INFO) 142 | ch.setFormatter(formatter) 143 | logger.addHandler(ch) 144 | 145 | # Set random seed 146 | np.random.seed(random_seed) 147 | 148 | # Load data 149 | logger.info(f'Loading data from {read_path}') 150 | adata = sc.read_h5ad(read_path) 151 | n_regions = len(adata.obs['region_true'].unique()) 152 | 153 | # Smooth 154 | logger.info('Smoothing') 155 | logger.info(f'\tsample_rate={sample_rate}') 156 | logger.info(f'\trandom_seed={random_seed}') 157 | adata = smooth(adata, k_physical, sample_rate, n_epochs, alpha, random_seed) 158 | 159 | # Cluster 160 | logger.info('Clustering') 161 | cluster(adata, n_pcs, k_latent, n_regions, leiden_res) 162 | 163 | # Quantify results 164 | logger.info('Quantifying results') 165 | recon_score, kmeans_ari, leiden_ari = quantify(adata, n_regions) 166 | 167 | # Write results to disk 168 | os.makedirs(write_path, exist_ok=True) 169 | results = f'{np.round(recon_score, decimals=decimals)}_' + \ 170 | f'{np.round(kmeans_ari, decimals=decimals)}_' + \ 171 | f'{np.round(leiden_ari, decimals=decimals)}' 172 | filename_txt = f'{random_seed}_{results}.txt' 173 | write_path_txt = os.path.join(write_path, filename_txt) 174 | logger.info(f'Writing results to {write_path_txt}') 175 | f = open(write_path_txt, 'a') 176 | f.close() 177 | 178 | # Write example AnnData to disk 179 | if random_seed == 0: 180 | filename_adata = f'adata_{sample_rate}.h5ad' 181 | write_path_adata = os.path.join(write_path, filename_adata) 182 | logger.info(f'Saving AnnData to {write_path_adata}') 183 | adata.write(write_path_adata) 184 | 185 | logger.info('Done') 186 | 187 | 188 | if __name__=='__main__': 189 | 190 | parser = argparse.ArgumentParser() 191 | parser.add_argument('--read_path', type=str) 192 | parser.add_argument('--sample_rate', type=float, default=None) 193 | parser.add_argument('--k_physical', type=int, default=50) 194 | parser.add_argument('--k_latent', type=int, default=15) 195 | parser.add_argument('--n_pcs', type=int) 196 | parser.add_argument('--leiden_res', type=float) 197 | parser.add_argument('--write_path', type=str) 198 | parser.add_argument('--random_seed', type=int) 199 | parser.add_argument('--decimals', type=int, default=5) 200 | parser.add_argument('--n_epochs', type=int, default=50) 201 | parser.add_argument('--alpha', type=float, default=0) 202 | args = parser.parse_args() 203 | 204 | main( 205 | args.read_path, 206 | args.sample_rate, 207 | args.k_physical, 208 | args.k_latent, 209 | args.n_pcs, 210 | args.leiden_res, 211 | args.write_path, 212 | args.random_seed, 213 | args.decimals, 214 | args.n_epochs, 215 | args.alpha 216 | ) -------------------------------------------------------------------------------- /code_for_figures/benchmarking/scripts/utag_.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | import argparse 4 | 5 | from utag.segmentation import utag_mod 6 | import scanpy as sc 7 | import numpy as np 8 | from sklearn.neighbors import NearestNeighbors 9 | from sklearn.decomposition import PCA 10 | import scipy as sp 11 | from sklearn.cluster import KMeans 12 | from sklearn.metrics import adjusted_rand_score 13 | 14 | 15 | def smooth( 16 | adata, 17 | radius, 18 | sample_rate, 19 | random_seed, 20 | ): 21 | 22 | adata = utag_mod( 23 | adata=adata, 24 | max_dist=radius, 25 | sample_rate=sample_rate, 26 | random_state=random_seed, 27 | slide_key=None, 28 | apply_clustering=False, 29 | apply_umap=False, 30 | ) 31 | 32 | return adata 33 | 34 | 35 | def binary_search_leiden( 36 | adata, 37 | leiden_res, 38 | n_regions, 39 | jitter=0.1 40 | ): 41 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 42 | n_clusters = len(adata.obs['leiden'].unique()) 43 | lo, hi = 0, 5 44 | iters = 0 45 | while n_clusters != n_regions: 46 | if n_clusters < n_regions: 47 | lo = leiden_res 48 | leiden_res = (leiden_res+hi)/2 49 | elif n_clusters > n_regions: 50 | hi = leiden_res 51 | leiden_res = (leiden_res+lo)/2 52 | iters += 1 53 | if iters > 10: 54 | leiden_res += np.random.choice([-1,1]) * jitter 55 | hi += jitter 56 | lo -= max(0,jitter) 57 | sc.tl.leiden(adata, resolution=leiden_res, neighbors_key='smoothed') 58 | n_clusters = len(adata.obs['leiden'].unique()) 59 | print(leiden_res, n_clusters, flush=True) 60 | print() 61 | 62 | 63 | def cluster( 64 | adata, 65 | n_pcs, 66 | k_latent, 67 | n_regions, 68 | leiden_res, 69 | ): 70 | 71 | # Calculate latent representation 72 | pca = PCA(n_components=n_pcs) 73 | adata.obsm['X_smoothed_pca'] = pca.fit_transform(adata.obsm['X_smoothed']) 74 | 75 | # Find latent neighbors 76 | nbrs = NearestNeighbors(n_neighbors=k_latent).fit(adata.obsm['X_smoothed_pca']) 77 | adata.obsp['smoothed_connectivities'] = nbrs.kneighbors_graph(mode='connectivity') 78 | adata.obsp['smoothed_distances'] = nbrs.kneighbors_graph(mode='distance') 79 | 80 | # Add metadata expected by Scanpy's Leiden and UMAP 81 | adata.uns['smoothed'] = {'connectivities_key': 'smoothed_connectivities', 82 | 'distances_key': 'smoothed_distances', 83 | 'params': { 84 | 'n_neighbors': k_latent, 85 | 'method': 'umap', 86 | 'random_state': 0, 87 | 'metric': 'euclidean', 88 | 'use_rep': 'X_smoothed_pca' 89 | } 90 | } 91 | 92 | # Perform k-means clustering 93 | kmeans = KMeans(n_clusters=n_regions, n_init=10) 94 | kmeans.fit(adata.obsm['X_smoothed_pca']) 95 | adata.obs['kmeans'] = kmeans.labels_.astype(str) 96 | 97 | # Perform Leiden clustering 98 | binary_search_leiden(adata, leiden_res, n_regions) 99 | 100 | 101 | def quantify(adata, n_regions): 102 | 103 | # Quantify spatial reconstruction 104 | Cp = adata.obsp['spatial_connectivities'] 105 | Cl = adata.obsp['smoothed_connectivities'] 106 | recon_score = Cp.multiply(Cl).count_nonzero() / (Cp+Cl).count_nonzero() 107 | 108 | # Quantify k-means accuracy 109 | kmeans_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['kmeans']) 110 | 111 | # Quantify Leiden accuracy 112 | leiden_ari = adjusted_rand_score(adata.obs['region_true'], adata.obs['leiden']) 113 | 114 | return recon_score, kmeans_ari, leiden_ari 115 | 116 | 117 | def main( 118 | read_path, 119 | sample_rate, 120 | radius, 121 | k_latent, 122 | n_pcs, 123 | leiden_res, 124 | write_path, 125 | random_seed, 126 | decimals, 127 | ): 128 | 129 | # Initialize logger 130 | logger = logging.getLogger('UTAG') 131 | logger.setLevel(logging.INFO) 132 | formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') 133 | ch = logging.StreamHandler() 134 | ch.setLevel(logging.INFO) 135 | ch.setFormatter(formatter) 136 | logger.addHandler(ch) 137 | 138 | # Set random seed 139 | np.random.seed(random_seed) 140 | 141 | # Load data 142 | logger.info(f'Loading data from {read_path}') 143 | adata = sc.read_h5ad(read_path) 144 | n_regions = len(adata.obs['region_true'].unique()) 145 | 146 | # Smooth 147 | logger.info('Smoothing') 148 | logger.info(f'\tsample_rate={sample_rate}') 149 | logger.info(f'\trandom_seed={random_seed}') 150 | adata = smooth(adata, radius, sample_rate, random_seed) 151 | 152 | # Cluster 153 | logger.info('Clustering') 154 | cluster(adata, n_pcs, k_latent, n_regions, leiden_res) 155 | 156 | # Quantify results 157 | logger.info('Quantifying results') 158 | recon_score, kmeans_ari, leiden_ari = quantify(adata, n_regions) 159 | 160 | # Write results to disk 161 | os.makedirs(write_path, exist_ok=True) 162 | results = f'{np.round(recon_score, decimals=decimals)}_' + \ 163 | f'{np.round(kmeans_ari, decimals=decimals)}_' + \ 164 | f'{np.round(leiden_ari, decimals=decimals)}' 165 | filename_txt = f'{random_seed}_{results}.txt' 166 | write_path_txt = os.path.join(write_path, filename_txt) 167 | logger.info(f'Writing results to {write_path_txt}') 168 | f = open(write_path_txt, 'a') 169 | f.close() 170 | 171 | # Write example AnnData to disk 172 | if random_seed == 0: 173 | filename_adata = f'adata_{sample_rate}.h5ad' 174 | write_path_adata = os.path.join(write_path, filename_adata) 175 | logger.info(f'Saving AnnData to {write_path_adata}') 176 | adata.write(write_path_adata) 177 | 178 | logger.info('Done') 179 | 180 | 181 | if __name__=='__main__': 182 | 183 | parser = argparse.ArgumentParser() 184 | parser.add_argument('--read_path', type=str) 185 | parser.add_argument('--sample_rate', type=float, default=None) 186 | parser.add_argument('--radius', type=float, default=None) 187 | parser.add_argument('--k_latent', type=int, default=50) 188 | parser.add_argument('--n_pcs', type=int) 189 | parser.add_argument('--leiden_res', type=float) 190 | parser.add_argument('--write_path', type=str) 191 | parser.add_argument('--random_seed', type=int) 192 | parser.add_argument('--decimals', type=int, default=5) 193 | args = parser.parse_args() 194 | 195 | main( 196 | args.read_path, 197 | args.sample_rate, 198 | args.radius, 199 | args.k_latent, 200 | args.n_pcs, 201 | args.leiden_res, 202 | args.write_path, 203 | args.random_seed, 204 | args.decimals, 205 | ) -------------------------------------------------------------------------------- /code_for_figures/generalization/.DS_Store: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/wanglab-broad/spin/53c174bfcc029f2614bf381f3d61557f3e1824d1/code_for_figures/generalization/.DS_Store -------------------------------------------------------------------------------- /code_for_figures/generalization/gut/xcondition.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": null, 6 | "metadata": {}, 7 | "outputs": [], 8 | "source": [ 9 | "import scanpy as sc\n", 10 | "import numpy as np\n", 11 | "from matplotlib.patches import Rectangle\n", 12 | "import seaborn as sns \n", 13 | "import matplotlib.pyplot as plt\n", 14 | "import os \n", 15 | "from scipy.stats import mannwhitneyu" 16 | ] 17 | }, 18 | { 19 | "cell_type": "markdown", 20 | "metadata": {}, 21 | "source": [ 22 | "## Load data and set up params" 23 | ] 24 | }, 25 | { 26 | "cell_type": "code", 27 | "execution_count": null, 28 | "metadata": {}, 29 | "outputs": [], 30 | "source": [ 31 | "adata = sc.read_h5ad('/stanley/WangLab/kamal/data/mouse/gut/ileum_spin_filtered.h5ad')\n", 32 | "basepath = '/stanley/WangLab/kamal/code/remote_notebooks/gut_xcondition/figures_kamal/'\n", 33 | "dpi = 300\n", 34 | "ctypes = adata.obs['general_annos'].unique()\n", 35 | "ctype_cmap = {ctypes[i]:sc.pl.palettes.default_102[i] for i in range(len(ctypes))}\n", 36 | "region_cmap = {str(i):sc.pl.palettes.default_102[i] for i in range(100)}\n", 37 | "figsize = (7,4.5)" 38 | ] 39 | }, 40 | { 41 | "cell_type": "code", 42 | "execution_count": null, 43 | "metadata": {}, 44 | "outputs": [], 45 | "source": [ 46 | "adata" 47 | ] 48 | }, 49 | { 50 | "cell_type": "code", 51 | "execution_count": null, 52 | "metadata": {}, 53 | "outputs": [], 54 | "source": [ 55 | "# Get vmin and vmax of gene expression across samples for colorbar normalization\n", 56 | "def getMinMax(adata, gene, layer='smooth'):\n", 57 | " return (adata[:,gene].layers[layer].min(), adata[:,gene].layers[layer].max())" 58 | ] 59 | }, 60 | { 61 | "attachments": {}, 62 | "cell_type": "markdown", 63 | "metadata": {}, 64 | "source": [ 65 | "# Full tissue slices" 66 | ] 67 | }, 68 | { 69 | "cell_type": "markdown", 70 | "metadata": {}, 71 | "source": [ 72 | "## Tissue and latent colored by cell type" 73 | ] 74 | }, 75 | { 76 | "cell_type": "code", 77 | "execution_count": null, 78 | "metadata": {}, 79 | "outputs": [], 80 | "source": [ 81 | "sc.set_figure_params(figsize=figsize)\n", 82 | "sc.pl.embedding(adata, basis='spatial', color='general_annos', palette=ctype_cmap, s=3, title='', legend_loc=None, return_fig=True)\n", 83 | "plt.axis('off')\n", 84 | "figname = f'tissue_colored_by_celltype.png'\n", 85 | "savepath = os.path.join(basepath, figname)\n", 86 | "plt.savefig(savepath, bbox_inches='tight', transparent=True, dpi=dpi)" 87 | ] 88 | }, 89 | { 90 | "cell_type": "code", 91 | "execution_count": null, 92 | "metadata": {}, 93 | "outputs": [], 94 | "source": [ 95 | "sc.set_figure_params(figsize=(5,5))\n", 96 | "sc.pl.embedding(adata, basis='X_umap', color='general_annos', palette=ctype_cmap, s=3, title='', legend_loc='', return_fig=True)\n", 97 | "plt.axis('off')\n", 98 | "figname = f'latent_colored_by_celltype.png'\n", 99 | "savepath = os.path.join(basepath, figname)\n", 100 | "plt.savefig(savepath, bbox_inches='tight', transparent=True, dpi=dpi)" 101 | ] 102 | }, 103 | { 104 | "cell_type": "code", 105 | "execution_count": null, 106 | "metadata": {}, 107 | "outputs": [], 108 | "source": [ 109 | "sc.set_figure_params(figsize=(5,5))\n", 110 | "sc.pl.embedding(adata, basis='X_umap', color='sample', s=3, title='', legend_loc='', return_fig=True)\n", 111 | "plt.axis('off')\n", 112 | "figname = f'latent_colored_by_condition_celltype.png'\n", 113 | "savepath = os.path.join(basepath, figname)\n", 114 | "plt.savefig(savepath, bbox_inches='tight', transparent=True, dpi=dpi)" 115 | ] 116 | }, 117 | { 118 | "cell_type": "code", 119 | "execution_count": null, 120 | "metadata": {}, 121 | "outputs": [], 122 | "source": [ 123 | "sc.set_figure_params(figsize=(5,5))\n", 124 | "random_idxs = np.random.choice(np.arange(len(adata)), size=len(adata), replace=False)\n", 125 | "sc.pl.embedding(adata[random_idxs], basis='X_umap', color='sample', s=3, title='', legend_loc='', return_fig=True)\n", 126 | "plt.axis('off')\n", 127 | "figname = f'latent_colored_by_condition_celltype.png'\n", 128 | "savepath = os.path.join(basepath, figname)\n", 129 | "plt.savefig(savepath, bbox_inches='tight', transparent=True, dpi=dpi)" 130 | ] 131 | }, 132 | { 133 | "cell_type": "code", 134 | "execution_count": null, 135 | "metadata": {}, 136 | "outputs": [], 137 | "source": [ 138 | "sc.set_figure_params(figsize=(10,10))\n", 139 | "sc.pl.embedding(adata, basis='X_umap', color='general_annos', palette=ctype_cmap, s=10, title='', legend_loc='on data')" 140 | ] 141 | }, 142 | { 143 | "cell_type": "code", 144 | "execution_count": null, 145 | "metadata": {}, 146 | "outputs": [], 147 | "source": [ 148 | "sc.set_figure_params(figsize=(10,10))\n", 149 | "sc.pl.embedding(adata, basis='X_umap', color='sample', s=10, title='')" 150 | ] 151 | }, 152 | { 153 | "cell_type": "markdown", 154 | "metadata": {}, 155 | "source": [ 156 | "## Tissue and latent colored by region" 157 | ] 158 | }, 159 | { 160 | "cell_type": "code", 161 | "execution_count": null, 162 | "metadata": {}, 163 | "outputs": [], 164 | "source": [ 165 | "sc.set_figure_params(figsize=figsize)\n", 166 | "sc.pl.embedding(adata, basis='spatial', color='region', palette=region_cmap, s=3, title='', legend_loc=None, return_fig=True)\n", 167 | "plt.axis('off')\n", 168 | "figname = f'tissue_colored_by_region.png'\n", 169 | "savepath = os.path.join(basepath, figname)\n", 170 | "plt.savefig(savepath, bbox_inches='tight', transparent=True, dpi=dpi)" 171 | ] 172 | }, 173 | { 174 | "cell_type": "code", 175 | "execution_count": null, 176 | "metadata": {}, 177 | "outputs": [], 178 | "source": [ 179 | "sc.set_figure_params(figsize=(5,5))\n", 180 | "sc.pl.embedding(adata, basis='X_umap_spin', color='region', palette=region_cmap, s=3, title='', legend_loc=None, return_fig=True)\n", 181 | "plt.axis('off')\n", 182 | "figname = f'latent_colored_by_region.png'\n", 183 | "savepath = os.path.join(basepath, figname)\n", 184 | "plt.savefig(savepath, bbox_inches='tight', transparent=True, dpi=dpi)" 185 | ] 186 | }, 187 | { 188 | "cell_type": "code", 189 | "execution_count": null, 190 | "metadata": {}, 191 | "outputs": [], 192 | "source": [ 193 | "sc.set_figure_params(figsize=(5,5))\n", 194 | "sc.pl.embedding(adata, basis='X_umap_spin', color='sample', s=3, title='', legend_loc=None, return_fig=True)\n", 195 | "plt.axis('off')\n", 196 | "figname = f'latent_colored_by_condition.png'\n", 197 | "savepath = os.path.join(basepath, figname)\n", 198 | "plt.savefig(savepath, bbox_inches='tight', transparent=True, dpi=dpi)" 199 | ] 200 | }, 201 | { 202 | "cell_type": "code", 203 | "execution_count": null, 204 | "metadata": {}, 205 | "outputs": [], 206 | "source": [ 207 | "sc.set_figure_params(figsize=(5,5))\n", 208 | "random_idxs = np.random.choice(np.arange(len(adata)), size=len(adata), replace=False)\n", 209 | "sc.pl.embedding(adata[random_idxs], basis='X_umap_spin', color='sample', s=3, title='', legend_loc=None, return_fig=True)\n", 210 | "plt.axis('off')\n", 211 | "figname = f'latent_colored_by_condition.png'\n", 212 | "savepath = os.path.join(basepath, figname)\n", 213 | "plt.savefig(savepath, bbox_inches='tight', transparent=True, dpi=dpi)" 214 | ] 215 | }, 216 | { 217 | "cell_type": "markdown", 218 | "metadata": {}, 219 | "source": [ 220 | "# Zooms" 221 | ] 222 | }, 223 | { 224 | "attachments": {}, 225 | "cell_type": "markdown", 226 | "metadata": {}, 227 | "source": [ 228 | "## Get zoom regions" 229 | ] 230 | }, 231 | { 232 | "cell_type": "code", 233 | "execution_count": null, 234 | "metadata": {}, 235 | "outputs": [], 236 | "source": [ 237 | "def get_zoom(adata, x, y, width, height, theta):\n", 238 | " x1, x2 = x, x+width\n", 239 | " y1, y2 = y, y+height\n", 240 | " R = np.array([\n", 241 | " [np.cos(theta), -np.sin(theta)],\n", 242 | " [np.sin(theta), np.cos(theta)]\n", 243 | " ])\n", 244 | " rdata = adata.copy()\n", 245 | " rdata.obsm['spatial'] = rdata.obsm['spatial'] @ R.T\n", 246 | " zoomdata = rdata[rdata.obsm['spatial'][:,0]>x1]\n", 247 | " zoomdata = zoomdata[zoomdata.obsm['spatial'][:,0]y1]\n", 249 | " zoomdata = zoomdata[zoomdata.obsm['spatial'][:,1]=3.2,<4"] 3 | build-backend = "flit_core.buildapi" 4 | 5 | [project] 6 | name = "spin" 7 | version = "0.0.1" 8 | description = "SPatial INtegration of spatially resolved transcriptomics data" 9 | authors = [ 10 | {name = "Kamal Maher", email = "kmaher@mit.edu"}, 11 | ] 12 | readme = {file = "README.md", content-type="text/markdown"} 13 | license = {file = "LICENSE"} 14 | requires-python = ">=3.9" 15 | dependencies = [ 16 | "scanpy[leiden,harmony]", 17 | ] 18 | 19 | [tool.setuptools] 20 | package-dir = {"" = "src"} 21 | include-package-data = true 22 | 23 | [project.scripts] 24 | spin = "spin.cli:spin_cli" 25 | -------------------------------------------------------------------------------- /src/spin/__init__.py: -------------------------------------------------------------------------------- 1 | """ 2 | Spatially integrate spatially resolved transcriptomics (SRT) datasets. 3 | """ 4 | 5 | from .spin import spin 6 | -------------------------------------------------------------------------------- /src/spin/cli.py: -------------------------------------------------------------------------------- 1 | """ 2 | Command-line interface for using SPIN from the shell. 3 | """ 4 | 5 | import argparse 6 | 7 | from .spin import spin 8 | 9 | 10 | def spin_cli(): 11 | 12 | # Parse arguments 13 | parser = argparse.ArgumentParser( 14 | description='SPatially INtegrate and cluster one or more spatially resolved \ 15 | transcriptomics datasets' 16 | ) 17 | parser.add_argument('--adata_paths', type=str, nargs='+', default=None) 18 | parser.add_argument('--write_path', type=str, default=None) 19 | parser.add_argument('--batch_key', type=str, default=None) 20 | parser.add_argument('--batch_labels', type=str, nargs='+', default=None) 21 | parser.add_argument('--n_nbrs', type=int, nargs='+', default=30) 22 | parser.add_argument('--n_samples', type=int, nargs='+', default=None) 23 | parser.add_argument('--spatial_key', type=str, default='spatial') 24 | parser.add_argument('--n_pcs', type=int, default=50) 25 | parser.add_argument('--svd_solver', default='randomized') 26 | parser.add_argument('--pca_key', default='X_pca_spin') 27 | parser.add_argument('--region_key', default='region') 28 | parser.add_argument('--umap_key', default='X_umap_spin') 29 | parser.add_argument('--resolution', type=float, default=0.5) 30 | parser.add_argument('--verbose', type=bool, default=True) 31 | parser.add_argument('--random_state', type=int, default=0) 32 | args = parser.parse_args() 33 | 34 | # Run SPIN 35 | spin( 36 | adata_paths=args.adata_paths, 37 | write_path=args.write_path, 38 | batch_key=args.batch_key, 39 | batch_labels=args.batch_labels, 40 | n_nbrs=args.n_nbrs, 41 | n_samples=args.n_samples, 42 | spatial_key=args.spatial_key, 43 | n_pcs=args.n_pcs, 44 | svd_solver=args.svd_solver, 45 | pca_key=args.pca_key, 46 | region_key=args.region_key, 47 | umap_key=args.umap_key, 48 | resolution=args.resolution, 49 | verbose=args.verbose, 50 | random_state=args.random_state, 51 | ) 52 | -------------------------------------------------------------------------------- /src/spin/spin.py: -------------------------------------------------------------------------------- 1 | """ 2 | SPatially INtegrate spatially resolved transcriptomics (SRT) datasets. 3 | """ 4 | 5 | from __future__ import annotations 6 | 7 | import logging 8 | from typing import Optional, Collection 9 | 10 | import scanpy as sc 11 | import numpy as np 12 | from sklearn.neighbors import NearestNeighbors 13 | from sklearn.decomposition import PCA 14 | from anndata import AnnData 15 | 16 | 17 | # Create logger 18 | logger = logging.getLogger('SPIN') 19 | logger.setLevel(logging.INFO) 20 | formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') 21 | ch = logging.StreamHandler() 22 | ch.setLevel(logging.INFO) 23 | ch.setFormatter(formatter) 24 | logger.addHandler(ch) 25 | 26 | 27 | def spin( 28 | adatas: Optional[Collection[AnnData] | AnnData] = None, 29 | adata_paths: Optional[Collection[str]] = None, 30 | write_path: Optional[str] = None, 31 | batch_key: Optional[str] = None, 32 | batch_labels: Optional[Collection[str]] = None, 33 | n_nbrs: Collection[int] | int = 30, 34 | n_samples: Collection[Optional[int]] | Optional[int] = None, 35 | spatial_key: str = 'spatial', 36 | n_pcs: int = 50, 37 | svd_solver: str = 'randomized', 38 | pca_key: str = 'X_pca_spin', 39 | region_key: str = 'region', 40 | umap_key: str = 'X_umap_spin', 41 | resolution: float = 0.5, 42 | verbose: bool = True, 43 | random_state: int = 0, 44 | ): 45 | """\ 46 | Spatially integrate and cluster SRT data using SPIN. 47 | 48 | Parameters 49 | ---------- 50 | adatas 51 | One or more SRT datasets. 52 | Assumed to be have been normalized prior. 53 | adata_paths 54 | Paths to one or more SRT datasets. 55 | write_path 56 | Path to write integrated data to. 57 | batch_key 58 | The key to batch information within `adata.obs`. 59 | batch_labels 60 | Labels corresponding to each batch. Relevant when integrating across multiple 61 | `adatas`. Will be stored under `adata.obs[]`. 62 | n_nbrs 63 | Number of nearest neighbors to find for each cell. 64 | n_samples 65 | Number of random neighbor samples used for averaging. 66 | spatial_key 67 | The key to spatial coordinates within `adata.obsm`. 68 | n_pcs 69 | Number of PCs to calculate for dimension reduction. 70 | svd_solver 71 | SVD solver to use. 72 | pca_key 73 | The key to store PCA output under within `adata.obsm`. 74 | region_key 75 | The key to store region labels under within `adata.obs`. 76 | umap_key 77 | The key to store UMAP output under within `adata.obsm`. 78 | resolution 79 | Resolution for Leiden clustering 80 | verbose 81 | Display updates on function progress. 82 | random_state 83 | Random seed used for smoothing, PCA, and Harmony. 84 | """ 85 | adata_source_types = np.array([type(adatas), type(adata_paths)]) 86 | n_adata_sources = np.sum(adata_source_types==type(None)) 87 | assert n_adata_sources == 1, "Requires either a list of paths or a list of AnnDatas" 88 | 89 | # Read data 90 | if adata_paths: 91 | adatas = [] 92 | for path in adata_paths: 93 | if verbose: 94 | logger.info(f'Reading {path}') 95 | adatas.append(sc.read_h5ad(path)) 96 | 97 | # Integrate spatial features across samples 98 | adata = _integrate( 99 | adatas, 100 | batch_key=batch_key, 101 | batch_labels=batch_labels, 102 | n_nbrs=n_nbrs, 103 | n_samples=n_samples, 104 | spatial_key=spatial_key, 105 | n_pcs=n_pcs, 106 | svd_solver=svd_solver, 107 | pca_key=pca_key, 108 | random_state=random_state, 109 | verbose=verbose, 110 | ) 111 | 112 | # Cluster integrated samples to find regions 113 | adata = _cluster( 114 | adata, 115 | pca_key=pca_key, 116 | region_key=region_key, 117 | umap_key=umap_key, 118 | resolution=resolution, 119 | verbose=verbose, 120 | ) 121 | 122 | # Write data 123 | if write_path: 124 | adata.write(write_path) 125 | if verbose: 126 | logger.info(f'Written to {write_path}') 127 | else: 128 | return adata 129 | 130 | 131 | def _integrate( 132 | adatas: Collection[AnnData] | AnnData, 133 | batch_key: Optional[str] = None, 134 | batch_labels: Optional[Collection[str]] = None, 135 | n_nbrs: Collection[int] | int = 30, 136 | n_samples: Collection[Optional[int]] | Optional[int] = None, 137 | spatial_key: str = 'spatial', 138 | n_pcs: int = 50, 139 | svd_solver: str = 'randomized', 140 | pca_key: str = 'X_pca_spin', 141 | random_state: int = 0, 142 | verbose: bool = True, 143 | ) -> AnnData: 144 | """\ 145 | Smooth and integrate SRT datasets. 146 | 147 | Parameters 148 | ---------- 149 | adatas 150 | One or more SRT datasets. 151 | Assumed to be have been normalized prior. 152 | batch_key 153 | The key to batch information within `adata.obs`. Relevant when integrating across 154 | multiple samples. 155 | If only analyzing a single sample, leave as `None`. 156 | batch_labels 157 | Labels corresponding to each adata. Will be stored under `adata.obs[]`. 158 | Required if passing in multiple adatas. Otherwise, leave as `None`. 159 | n_nbrs 160 | Number of nearest neighbors to find for each cell. 161 | Either provide n_nbrs for each dataset or a single n_nbrs for all datasets. 162 | Needs to be the same length as n_samples. 163 | n_samples 164 | Number of random neighbor samples used for averaging. 165 | Either provide n_samples for each dataset or a single n_samples for all datasets. 166 | Needs to be the same length as n_nbrs. 167 | spatial_key 168 | The key to spatial coordinates within `adata.obsm`. 169 | n_pcs 170 | Number of PCs to calculate for dimension reduction. 171 | svd_solver 172 | SVD solver to use. 173 | pca_key 174 | The key to store PCA output under within `adata.obsm`. 175 | random_state 176 | Random seed used for smoothing, PCA, and Harmony. 177 | verbose 178 | Display updates on function progress. 179 | 180 | Returns 181 | ------- 182 | Copy of adata input containing integrated spatial expression PCs. 183 | """ 184 | # Handle non-Collection input (e.g. single AnnData, n_nbrs, and/or n_samples) 185 | if type(adatas) == AnnData: 186 | adatas = [adatas] 187 | if type(n_nbrs) == int: 188 | n_nbrs = [n_nbrs] 189 | if (type(n_samples) == int) or (n_samples == None): 190 | n_samples = [n_samples] 191 | 192 | # Split single AnnData by batch, if batch_key provided 193 | n_adatas = len(adatas) 194 | if (n_adatas == 1): 195 | if batch_key: 196 | if verbose: 197 | logger.info(f'Splitting adata by `{batch_key}`') 198 | batch_labels = adatas[0].obs[batch_key].unique() 199 | adatas = [adatas[0][adatas[0].obs[batch_key]==batch] for batch in batch_labels] 200 | n_adatas = len(adatas) 201 | 202 | # If single n_nbrs/n_samples for multiple batches, clone 203 | if len(n_nbrs) < n_adatas: 204 | n_nbrs *= n_adatas 205 | if len(n_samples) < n_adatas: 206 | n_samples *= n_adatas 207 | 208 | # Smooth each batch independently 209 | if verbose: 210 | logger.info('Smoothing') 211 | for i in range(n_adatas): 212 | _get_nbrs( 213 | adatas[i], 214 | n_nbrs[i], 215 | spatial_key, 216 | ) 217 | _smooth( 218 | adatas[i], 219 | n_samples[i], 220 | random_state, 221 | ) 222 | 223 | # Concatenate batches and add metadata 224 | if n_adatas > 1: 225 | adata = sc.concat(adatas, keys=batch_labels, label=batch_key, join='inner') 226 | adata.uns['n_nbrs'] = [adatas[i].uns['n_nbrs'] for i in range(n_adatas)] 227 | adata.uns['n_samples'] = [adatas[i].uns['n_samples'] for i in range(n_adatas)] 228 | else: 229 | adata = adatas[0] 230 | 231 | # Run PCA 232 | if verbose: 233 | logger.info('Performing PCA') 234 | pca = PCA(n_components=n_pcs, svd_solver=svd_solver, random_state=random_state) 235 | adata.obsm[pca_key] = pca.fit_transform(adata.layers['smooth']) 236 | adata.varm[pca_key] = pca.components_.T 237 | 238 | # Integrate PCs 239 | if batch_labels: 240 | if verbose: 241 | logger.info('Integrating') 242 | sc.external.pp.harmony_integrate( 243 | adata, 244 | batch_key, 245 | basis=pca_key, 246 | adjusted_basis=pca_key, 247 | random_state=random_state, 248 | verbose=verbose, 249 | ) 250 | if verbose: 251 | logger.info('Integration complete') 252 | 253 | return adata.copy() # copying necessary for multiple runs on single AnnData 254 | 255 | 256 | def _cluster( 257 | adata: AnnData, 258 | pca_key: str = 'X_pca_spin', 259 | region_key: str = 'region', 260 | umap_key: str = 'X_umap_spin', 261 | resolution: float = 0.5, 262 | verbose: bool = True, 263 | ) -> AnnData: 264 | """\ 265 | Create nearest neighbors graph in latent space and perform UMAP and Leiden. 266 | 267 | Parameters 268 | -------- 269 | adata 270 | SRT dataset 271 | pca_key 272 | The key to PCA output within `adata.obsm`. 273 | region_key 274 | The key to store region labels under within `adata.obs`. 275 | umap_key 276 | The key to store UMAP output under within `adata.obsm`. 277 | resolution 278 | Resolution for Leiden clustering. 279 | verbose 280 | Display updates on function progress. 281 | 282 | Returns 283 | -------- 284 | Copy of adata input containing region cluster labels and UMAP coordinates 285 | """ 286 | if verbose: 287 | logger.info('Finding latent neighbors') 288 | sc.pp.neighbors( 289 | adata, 290 | use_rep=pca_key, 291 | key_added=region_key, 292 | ) 293 | if umap_key: 294 | if verbose: 295 | logger.info('Performing UMAP') 296 | umap = sc.tl.umap( 297 | adata, 298 | neighbors_key=region_key, 299 | copy=True, 300 | ).obsm['X_umap'] 301 | adata.obsm[umap_key] = umap 302 | if verbose: 303 | logger.info('Leiden clustering') 304 | sc.tl.leiden( 305 | adata, 306 | resolution=resolution, 307 | key_added=region_key, 308 | neighbors_key=region_key, 309 | ) 310 | if verbose: 311 | logger.info('Clustering complete') 312 | 313 | return adata.copy() 314 | 315 | 316 | def _get_nbrs( 317 | adata: AnnData, 318 | n_nbrs: int, 319 | spatial_key: str, 320 | ) -> np.ndarray: 321 | """\ 322 | Find spatial nearest neighbors of each cell. 323 | 324 | Parameters 325 | ---------- 326 | adata 327 | SRT dataset. 328 | n_nbrs 329 | Number of nearest neighbors to find for each cell. 330 | spatial_key 331 | The key to spatial coordinates within `adata.obsm`. 332 | 333 | Returns 334 | ------- 335 | Updates `adata` with the following fields: 336 | 337 | `.obsm['nbr_idxs']`: 338 | Matrix of neighbor indices of shape `n_obs` × `n_nbrs` 339 | `.uns['n_nbrs']`: 340 | Number of neighbors used 341 | `.uns['spatial_key']`: 342 | The key to `.obsm` for the domain used to find neighbors 343 | """ 344 | # Find spatial nearest neighbors 345 | coordinates = adata.obsm[spatial_key] 346 | nbrs = NearestNeighbors(n_neighbors=n_nbrs) 347 | nbrs.fit(coordinates) 348 | _, nbr_idxs = nbrs.kneighbors(coordinates) 349 | 350 | # Save metadata 351 | adata.obsm['nbr_idxs'] = nbr_idxs 352 | adata.uns['n_nbrs'] = n_nbrs 353 | adata.uns['spatial_key'] = spatial_key 354 | 355 | 356 | def _smooth( 357 | adata: AnnData, 358 | n_samples: Optional[int], 359 | random_state: int, 360 | ) -> np.ndarray: 361 | """\ 362 | Set each cell's representation to the average of its randomly-subsampled neighborhood. 363 | 364 | Parameters 365 | ---------- 366 | adata 367 | SRT dataset. 368 | nbr_idxs 369 | Matrix of neighbor indices of shape `n_obs` × `n_nbrs`. Rows correspond to cells 370 | and columns to the adata indices of their nearest neighbors. 371 | n_samples 372 | Number of random neighbor samples used for averaging. 373 | If None, sample 1/3 of n_nbrs, which is inferred from `nbr_idxs`. 374 | random_state 375 | Random seed for randomly subsampling neighborhoods. 376 | 377 | Returns 378 | ------- 379 | Updates `adata.layers` with the following fields: 380 | 381 | `smooth`: 382 | Smoothed data matrix of shape `n_obs` × `n_vars` 383 | `n_samples`: 384 | Number of random neighbor samples used for averaging 385 | """ 386 | # Randomly subsample each cell's neighborhood 387 | if not n_samples: 388 | n_samples = adata.obsm['nbr_idxs'].shape[1] // 3 389 | np.random.seed(random_state) 390 | nbr_idxs_sampled = np.array([ 391 | np.random.choice(idxs, size=n_samples, replace=False) 392 | for idxs in adata.obsm['nbr_idxs'] 393 | ]) 394 | 395 | # Set each cell's representation to the average of its subsampled neighborhood 396 | X_smooth = np.zeros(adata.X.shape) 397 | for nth_nbrs in np.array(nbr_idxs_sampled).T: 398 | X_smooth += adata.X[nth_nbrs] / n_samples 399 | 400 | # Save metadata 401 | adata.obsm['nbr_idxs_sampled'] = nbr_idxs_sampled 402 | adata.layers['smooth'] = np.array(X_smooth) # in case adata.X is sparse 403 | adata.uns['n_samples'] = n_samples 404 | --------------------------------------------------------------------------------