forked from Wang-Lin-boop/GeminiMol
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathREADME.html
593 lines (569 loc) · 42.5 KB
/
README.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
<!DOCTYPE html>
<html>
<head>
<meta charset="UTF-8">
<title>GeminiMol: Molecular Representation Model Enhanced by Conformational Space Profile</title>
<style>
/* From extension vscode.github */
/*---------------------------------------------------------------------------------------------
* Copyright (c) Microsoft Corporation. All rights reserved.
* Licensed under the MIT License. See License.txt in the project root for license information.
*--------------------------------------------------------------------------------------------*/
.vscode-dark img[src$=\#gh-light-mode-only],
.vscode-light img[src$=\#gh-dark-mode-only],
.vscode-high-contrast:not(.vscode-high-contrast-light) img[src$=\#gh-light-mode-only],
.vscode-high-contrast-light img[src$=\#gh-dark-mode-only] {
display: none;
}
</style>
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex/dist/katex.min.css">
<link href="https://cdn.jsdelivr.net/npm/katex-copytex@latest/dist/katex-copytex.min.css" rel="stylesheet" type="text/css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/Microsoft/vscode/extensions/markdown-language-features/media/markdown.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/Microsoft/vscode/extensions/markdown-language-features/media/highlight.css">
<style>
.language-shell {
background-color: #f0f0f0; /* 底色 */
padding: 10px; /* 内边距 */
border-radius: 5px; /* 圆角 */
font-family: monospace; /* 等宽字体 */
}
</style>
<style>
body {
font-family: -apple-system, BlinkMacSystemFont, 'Segoe WPC', 'Segoe UI', system-ui, 'Ubuntu', 'Droid Sans', sans-serif;
font-size: 14px;
line-height: 1.6;
margin: 0; /* 清除默认边距 */
margin-left: 300px; /* 左侧缩进 */
margin-right: 300px; /* 右侧缩进 */
margin-top: 200px;
margin-bottom: 200px;
}
</style>
<style>
.task-list-item {
list-style-type: none;
}
.task-list-item-checkbox {
margin-left: -20px;
vertical-align: middle;
pointer-events: none;
}
</style>
<style>
:root {
--color-note: #0969da;
--color-tip: #1a7f37;
--color-warning: #9a6700;
--color-severe: #bc4c00;
--color-caution: #d1242f;
--color-important: #8250df;
}
</style>
<style>
@media (prefers-color-scheme: dark) {
:root {
--color-note: #2f81f7;
--color-tip: #3fb950;
--color-warning: #d29922;
--color-severe: #db6d28;
--color-caution: #f85149;
--color-important: #a371f7;
}
}
</style>
<style>
.markdown-alert {
padding: 0.5rem 1rem;
margin-bottom: 16px;
color: inherit;
border-left: .25em solid #888;
}
.markdown-alert>:first-child {
margin-top: 0
}
.markdown-alert>:last-child {
margin-bottom: 0
}
.markdown-alert .markdown-alert-title {
display: flex;
font-weight: 500;
align-items: center;
line-height: 1
}
.markdown-alert .markdown-alert-title .octicon {
margin-right: 0.5rem;
display: inline-block;
overflow: visible !important;
vertical-align: text-bottom;
fill: currentColor;
}
.markdown-alert.markdown-alert-note {
border-left-color: var(--color-note);
}
.markdown-alert.markdown-alert-note .markdown-alert-title {
color: var(--color-note);
}
.markdown-alert.markdown-alert-important {
border-left-color: var(--color-important);
}
.markdown-alert.markdown-alert-important .markdown-alert-title {
color: var(--color-important);
}
.markdown-alert.markdown-alert-warning {
border-left-color: var(--color-warning);
}
.markdown-alert.markdown-alert-warning .markdown-alert-title {
color: var(--color-warning);
}
.markdown-alert.markdown-alert-tip {
border-left-color: var(--color-tip);
}
.markdown-alert.markdown-alert-tip .markdown-alert-title {
color: var(--color-tip);
}
.markdown-alert.markdown-alert-caution {
border-left-color: var(--color-caution);
}
.markdown-alert.markdown-alert-caution .markdown-alert-title {
color: var(--color-caution);
}
</style>
</head>
<body class="vscode-body vscode-light">
<h2 align="center"> Molecular Representation Model Enhanced by Conformational Space Profile </h2>
<p align="center">
📃 <a href="https://onlinelibrary.wiley.com/doi/10.1002/advs.202403998" target="_blank">Paper</a> · 🤗 <a href="https://huggingface.co/AlphaMWang/GeminiMol" target="_blank">Model</a> · 📕 <a href="https://zenodo.org/records/10450788" target="_blank">Data</a><br>
</p>
<p align="center">
<img style="float: center" src="imgs\geminimol.png" alt="alt text" width="750px" align="center"/>
</p>
<p>This repository provides the official implementation of the GeminiMol model, training data, and utilities. In this work, we propose a hybrid contrastive learning framework, which conducts <strong>inter-molecular contrastive learning</strong> by multiple projection heads of <strong>conformational space similarities (CSS)</strong>. Please also refer to our <a href="https://doi.org/10.1101/2023.12.14.571629">paper</a> for a detailed description of GeminiMol.</p>
<h2 id="table-of-contents">Table of Contents</h2>
<ul>
<li><a href="#table-of-contents">Table of Contents</a></li>
<li><a href="#-motivation">💗 Motivation</a></li>
<li><a href="#-highlight">💡 Highlight</a></li>
<li><a href="#-news">🔔 News</a></li>
<li><a href="#-limitations">😫 Limitations</a></li>
<li><a href="#-installation">📕 Installation</a>
<ul>
<li><a href="#download-datasets-and-models">Download datasets and models</a></li>
<li><a href="#installing-the-dependency-packages">Installing the dependency packages</a></li>
</ul>
</li>
<li><a href="#-application">📓 Application</a>
<ul>
<li><a href="#virtual-screening-and-target-identification">Virtual Screening and Target Identification</a></li>
<li><a href="#molecular-proptery-modeling-qsar-and-admet">Molecular Proptery Modeling (QSAR and ADMET)</a></li>
<li><a href="#molecular-clustering">Molecular Clustering</a></li>
<li><a href="#extract-molecular-features-geminimol-encoding">Extract Molecular Features (GeminiMol Encoding)</a></li>
</ul>
</li>
<li><a href="#-reproducing">👐 Reproducing</a>
<ul>
<li><a href="#download-training-and-benchmark-datasets">Download Training and Benchmark Datasets</a></li>
<li><a href="#re-training-our-models">Re-training our models</a></li>
<li><a href="#benchmarking-the-fingerprints-and-our-models">Benchmarking the fingerprints and our models</a></li>
</ul>
</li>
<li><a href="#-citing-this-work">⭐ Citing This Work</a></li>
<li><a href="#-license">✅ License</a></li>
<li><a href="#-get-in-touch">💌 Get in Touch</a></li>
<li><a href="#-acknowledgements">😃 Acknowledgements</a></li>
</ul>
<h2 id="-motivation">💗 Motivation</h2>
<p>The <strong>molecular representation model</strong> is an emerging artificial intelligence technology for extracting features of small molecules. Inspired by the dynamics of small molecules in solution, introducing the <strong>conformational space profile</strong> into molecular representation models is a promising aim. The conformational space profile covers the heterogeneity of molecule properties, such as the multi-target mechanism of drug action, recognition of different biomolecules, dynamics in cytoplasm and membrane, which may facilitate further downstream application and generalization capability of molecular representation model.</p>
<h2 id="-highlight">💡 Highlight</h2>
<ul>
<li>GeminiMol exhibits the capability to <strong>identify molecular pairs with similar 3D active conformers</strong>, even in scenarios where their 2D structures exhibit significant differences.</li>
<li>GeminiMol was pre-trained on only 37,336 molecular structures, yet it can <strong>generalize</strong> to zero-shot and QSAR tasks involving millions of molecules.</li>
<li>GeminiMol shown the <strong>balanced performance</strong> across various applications, including virtual screening, target identification, and cellular phenotype-based property modeling.</li>
</ul>
<h2 id="-news">🔔 News</h2>
<ul>
<li>2023-12, our paper has been uploaded to BioRxiv, you can find it <a href="https://www.biorxiv.org/content/10.1101/2023.12.14.571629">here</a>.</li>
<li>2024-01, we have released <code>PharmProfiler.py</code>, which facilitates virtual screening and target identification.</li>
<li>2024-03, we have released <code>PropPredictor.py</code>, which facilitates the deployment and repurposing of QSAR and ADMET prediction models.</li>
<li>2024-08, our paper was published in <a href="https://onlinelibrary.wiley.com/doi/10.1002/advs.202403998"><em>Advanced Science</em></a>.</li>
<li>2024-09, utilizing GeminiMol, we identified a new inhibitor targeting GluN1/GluN3A (IC<sub>50</sub> = 0.98 μM), and this result get rewarded the first prize in <a href="https://competition.huaweicloud.com/information/1000042002/introduction"><em>the 2023 Shanghai International Computational Biology Innovation Competition</em></a>.</li>
</ul>
<h2 id="-limitations">😫 Limitations</h2>
<ul>
<li>Note that, the conformational space profile is <strong>not a panacea</strong> for drug discovery. For a portion of tasks, the 2D structure of a compound already contains sufficient information to establish structure-activity relationships, rendering the introduction of the conformational space profile inconsequential for these tasks.</li>
<li>The evaluation of intermolecular similarity is not limited to pharmacophore similarity in 3D conformational space and maximum common substructure similarity in 2D structures. By incorporating <strong>additional intermolecular similarity metrics</strong> during pre-training, we can further enrich the knowledge that the model can learn, such as molecular fingerprints and molecular surface potentials.</li>
<li>Due to computational resource limitations, we only included 39,290 molecules in our pre-training. It is foreseeable that incorporating <strong>more molecular structures</strong> during pre-training could further enhance the performance of GeminiMol, particularly when guided by drug-target relationships to obtain high-quality data.</li>
</ul>
<h2 id="-installation">📕 Installation</h2>
<p>GeminiMol is a pytorch-based AI model. To set up the GeminiMol model, we recommend using conda for Python environment configuration. If you encounter any problems with the installation, please feel free to post an issue or discussion it.</p>
<blockquote>
<p>Installing MiniConda (skip if conda was installed)</p>
</blockquote>
<pre><code class="language-shell"> wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
sh Miniconda3-latest-Linux-x86_64.sh
</code></pre>
<blockquote>
<p>Creating GeminiMol environment</p>
</blockquote>
<pre><code class="language-shell"> conda create -n GeminiMol python=3.8.16
conda activate GeminiMol
</code></pre>
<blockquote>
<p>Setting up GeminiMol PATH and configuration</p>
</blockquote>
<pre><code class="language-shell"> git clone https://github.com/Wang-Lin-boop/GeminiMol
cd GeminiMol/
echo "# GeminiMol" >> ~/.bashrc
echo "export PATH=\"${PWD}:\${PATH}\"" >> ~/.bashrc # optional, not required in the current version
echo "export GeminiMol=\"${PWD}\"" >> ~/.bashrc
source ~/.bashrc
echo "export geminimol_app=\"${GeminiMol}/geminimol\"" >> ~/.bashrc # geminimol applications
echo "export geminimol_lib=\"${GeminiMol}/models\"" >> ~/.bashrc # geminimol models
echo "export geminimol_data=\"${GeminiMol}/data\"" >> ~/.bashrc # compound library
source ~/.bashrc
</code></pre>
<h4 id="download-datasets-and-models">Download datasets and models</h4>
<p>In this repository, we provide the pre-trained GeminiMol and CrossEncoder models.</p>
<blockquote>
<p>Download model parameters and weights via <a href="https://drive.google.com/drive/folders/183WGytS-zy_POlLxEvijEtarow56zmnz?usp=drive_link">Google Driver</a> and <a href="https://huggingface.co/AlphaMWang">HuggingFace</a></p>
</blockquote>
<p>Here is an example of how to download a model from huggingface. Besides wget, you can also download the model directly from Google Cloud Drive or huggingface using your browser.</p>
<pre><code class="language-bash">git <span class="hljs-built_in">clone</span> https://huggingface.co/AlphaMWang/GeminiMol
</code></pre>
<p>Then, we need place the models to the <code>${GeminiMol}/models</code>.</p>
<blockquote>
<p>Download all chemical datasets via <a href="https://zenodo.org/records/10450788">Zenodo</a> for applications</p>
</blockquote>
<pre><code class="language-shell"> cd ${geminimol_data}
wget https://zenodo.org/records/10450788/files/ChemDiv.zip # compound library for virtual screening
wget https://zenodo.org/records/10450788/files/DTIDB.zip # DTI database for target identification
for i in Benchmark*.zip css*.zip Chem*.zip;do
mkdir ${i%%.zip}
unzip -d ${i%%.zip}/ $i
done
unzip -d compound_library/ ChemDiv.zip
unzip -d compound_library/ DTIDB.zip
</code></pre>
<p>The expected structure of GeminiMol path is:</p>
<pre><code>GeminiMol
├── geminimol # all code for GeminiMol and CrossEncoder
│ ├── model # code for GeminiMol and CrossEncoder models
│ │ ├── CrossEncoder.py # model and methods for CrossEncoder
│ │ ├── GeminiMol.py # model and methods for GeminiMol models
│ ├── utils # utils in this work
│ │ ├── chem.py # basic tools for cheminformatics
│ │ ├── dataset_split.py # tools for dataset split
│ │ ├── fingerprint.py # basic tools for molecular fingerprints
│ ├── Analyzer.py # analysis given molecules using presentation methods
│ ├── AutoQSAR.py # molecular proptery modeling by AutoGluon
│ ├── PropDecoder.py # molecular proptery modeling by our PropDcoder
│ ├── FineTuning.py # molecular proptery modeling by fine-tune GeminiMol models
│ ├── Screener.py # screening molecules by GeminiMol similarity
│ ├── CrossEncoder_Training.py # scripts for training the CrossEncoders.
│ ├── GeminiMol_Training.py # scripts for training the GeminiMol models.
│ ├── benchmark.py # benchmarking presentation methods on provide datasets
├── data # training and benchmark datasets in this work
│ ├── Benchmark_DUD-E # virtual screeening benchmark, optional
│ ├── Benchmark_LIT-PCBA # virtual screeening benchmark, optional
│ ├── Benchmark_TIBD # target identification benchmark, optional
│ ├── Benchmark_QSAR # QSAR and ADMET benchmarks, optional
│ ├── Chem_SmELECTRA # text backbone of chemical language, optional
│ ├── css_library # CSS training data, optional
│ ├── benchmark.json # dataset index for benchmark tasks, optional
│ ├── database.csv # molecular datasets in this work, optional
│ ├── compound_library # the compound librarys
│ │ ├── DTIDB.csv # dataset used in target identification
│ │ ├── ChemDiv.csv # library of common commercial compounds
│ │ ├── Specs.csv # library of common commercial compounds
├── models # CrossEncoder and GeminiMol models
│ ├── CrossEncoder # CrossEncoder, optional
│ ├── GeminiMol # GeminiMol, recommended for zero-shot tasks
</code></pre>
<h4 id="installing-the-dependency-packages">Installing the dependency packages</h4>
<p>Before running GeminiMol, you need to install the basic dependency packages.</p>
<blockquote>
<p>Installing the RDkit for generating fingerprints</p>
</blockquote>
<pre><code class="language-shell"> pip install rdkit
</code></pre>
<blockquote>
<p>Installing the statatics and plot packages</p>
</blockquote>
<pre><code class="language-shell"> pip install six
pip install oddt scikit-learn matplotlib scipy==1.10.1
</code></pre>
<blockquote>
<p>Installing the dependency packages of GeminiMol</p>
</blockquote>
<pre><code class="language-shell"> pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 \
--extra-index-url https://download.pytorch.org/whl/cu116
pip install dgl==1.1.1+cu116 -f https://data.dgl.ai/wheels/cu116/repo.html
pip install dglgo==0.0.2 -f https://data.dgl.ai/wheels-test/repo.html
pip install dgllife==0.3.2
</code></pre>
<p>If you intend to reproduce the benchmark results in our work, it is required to install the AutoGluon.</p>
<pre><code class="language-shell"> pip install autogluon==0.8.1 # requried for AutoQSAR
</code></pre>
<h2 id="-application">📓 Application</h2>
<p>As a molecular representation model, GeminiMol finds applications in <strong>ligand-based virtual screening, target identification, and quantitative structure-activity relationship (QSAR)</strong> modeling of small molecular drugs.</p>
<p><img src="imgs\benchmark.png" alt="benchmark" width="900"></p>
<p>We have provided Cross-Encoder and GeminiMol models that can be used directly for inference. Here, we demonstrate the utilization of GeminiMol for virtual screening, target identification, and molecular property modeling.</p>
<p>Please note that while molecular fingerprints are considered simple molecular representation methods, they are an indispensable baseline (see our <a href="https://onlinelibrary.wiley.com/doi/10.1002/advs.202403998">paper</a>). When conducting your drug development project, we recommend exploring ECFP4, CombineFP, and GeminiMol that are provided simultaneously in our <code>PharmProfiler.py</code> and various molecular property modeling scripts.</p>
<h4 id="virtual-screening-and-target-identification">Virtual Screening and Target Identification</h4>
<p>In concept, molecules share similar conformational space also share similar biological activities, allowing us to predict the similarity of biological activities between molecules by comparing the similarity of GeminiMol encodings.</p>
<p>Here, we introduce the <code>PharmProfiler.py</code>, a novel approach that employs the GeminiMol encoding to establish pharmacological profiles and facilitate the search for molecules with specific properties in chemical space.</p>
<p><code>PharmProfiler.py</code> offers the capability to conduct ligand-based virtual screening using commercially available compound libraries. Furthermore, it enables target identification through ligand similarity analysis by leveraging comprehensive drug-target relationship databases.</p>
<p>To support experimentation, we have included a collection of diverse commercial compound libraries and drug-target relationship databases, conveniently located in the <code>${geminimol_data}/compound_library/</code> directory.</p>
<blockquote>
<ol>
<li>Prepare the pharmacological profile and compound libraries</li>
</ol>
</blockquote>
<p>To define a pharmacological profile, you will need to input a <code>profile.csv</code> file, which should have the following format:</p>
<pre><code>SMILES,Label
C=CC(=O)N[C@@H]1CN(c2nc(Nc3cn(C)nc3OC)c3ncn(C)c3n2)C[C@H]1F,1.0
C=CC(=O)Nc1cccc(Nc2nc(Nc3ccc(N4CCN(C(C)=O)CC4)cc3OC)ncc2C(F)(F)F)c1,1.0
C#Cc1cccc(Nc2ncnc3cc(OCCOC)c(OCCOC)cc23)c1,1.0
COC(=O)CCC/N=C1\SCCN1Cc1ccccc1,0.4
C=C(C)[C@@H]1C[C@@H](CC2(CC=C(C)C)C(=O)C(C(CC(=O)O)c3ccccc3)=C3O[C@@H](C)[C@@H](C)C(=O)C3=C2O)C1(C)C,-0.8
C/C(=C\c1ncccc1C)[C@@H]1C[C@@H]2O[C@]2(C)CCC[C@H](C)[C@H](O)[C@@H](C)C(=O)C(C)(C)[C@@H](O)CC(=O)O1,-0.5
</code></pre>
<p>The "Label" column signifies the weight assigned to the reference compound. Positive values indicate that the selected compounds should bear resemblance to the reference compound, while negative values imply that the selected compounds should be dissimilar to the reference compound. Typically, positive values are assigned to <strong>active</strong> compounds, whereas negative values are assigned to <strong>inactive</strong> compounds or those causing <strong>side effects</strong>.</p>
<p>The compound libraries are also stored in CSV format in the <code>${geminimol_data}/compound_library/</code> directory. It is requried to maintain consistency between the SMILES column name in the <code>profile.csv</code> file and the compound library.</p>
<blockquote>
<ol start="2">
<li>Perform the PharmProfiler</li>
</ol>
</blockquote>
<p>To perform virtual screening, the following command can be used.</p>
<p>Here, <code>profile_set</code> represents the provided pharmacological profile by the user, <code>keep_top</code> indicates the number of compounds to be outputted in the end, and <code>probe_cluster</code> determines whether compounds with the same weight should be treated as a cluster. Compounds within the same cluster will be compared individually with the query mol, and the highest similarity score will be taken as the score of query mol.</p>
<p>We have provided a processed version of the commercial Specs and ChemDiv compound library at the <code>${geminimol_data}/compound_library/specs.csv</code> and <code>${geminimol_data}/compound_library/ChemDiv.csv</code>, which contained 335,212 and 1,755,930 purchasable compounds.</p>
<pre><code class="language-shell">export job_name="Virtual_Screening"
export profile_set="profile.csv" # SMILES (same to compound library) and Label (requried)
export compound_library="${geminimol_data}/compound_library/ChemDiv.csv"
export smiles_column="SMILES" # Specify the column name in the compound_library
export weight_column="Label" # weights for profiles
export keep_top=1000
export probe_cluster="Yes"
python -u ${geminimol_app}/PharmProfiler.py "${geminimol_lib}/GeminiMol" "${job_name}" "${smiles_column}" "${compound_library}" "${profile_set}:${weight_column}" "${keep_top}" "${probe_cluster}"
</code></pre>
<p>To perform target identification, the compound library can be replaced with the <code>${geminimol_data}/compound_library/DTIDB.csv</code>, which contains drug-target relationships. This is a processed version of the BindingDB database, which contains 2,159,221 target-ligand paris.</p>
<pre><code class="language-shell">export job_name="Target_Identification"
export profile_set="profile.csv" # Ligand_SMILES (same to compound library), and Label (requried)
export compound_library="${geminimol_data}/compound_library/DTIDB.csv"
export smiles_column="SMILES" # Specify the column name in the compound_library
export weight_column="Label" # weights for profiles
export keep_top=2000
export probe_cluster="No"
python -u ${geminimol_app}/PharmProfiler.py "${geminimol_lib}/GeminiMol" "${job_name}" "${smiles_column}" "${compound_library}" "${profile_set}:${weight_column}" "${keep_top}" "${probe_cluster}"
</code></pre>
<p>After the initial run of PharmProfiler, a extracted GeminiMol feature file will be generated in the <code>${geminimol_data}/compound_library/</code>. Subsequent screening tasks on the same compound library can benefit from PharmProfiler automatically reading the feature file, which helps to accelerate the running speed.</p>
<h4 id="molecular-proptery-modeling-qsar-and-admet">Molecular Proptery Modeling (QSAR and ADMET)</h4>
<blockquote>
<ol>
<li>Prepare your datasets</li>
</ol>
</blockquote>
<p>Before conducting molecular property modeling, it is crucial to carefully prepare your data, which includes compound structure pre-processing and dataset splitting.</p>
<p>Firstly, you need to clarify the chirality and protonation states of molecules in the dataset, which can be done using chemical informatics tools such as RDKit or Schrödinger software package. Typically, omitting pre-processing will not result in an error, but it may potentially impair the performance of GeminiMol.</p>
<p>The processed data should be saved in CSV file format, containing at least one column for <strong><code>SMILES</code></strong> and one column for <strong><code>Labels</code></strong>. Subsequently, utilize the following command for skeleton splitting. You can modify the script to change the splitting ratio, where by default, 70% of the dataset is used for training and 30% for validation and testing.</p>
<pre><code class="language-shell">export dataset_path="data.csv"
export dataset_name="My_QSAR"
export smiles_column="SMILES" # Specify the column name in datasets
export label_column="Label" # Specify the column name in datasets
python -u ${geminimol_app}/utils/dataset_split.py "${dataset_path}" "${dataset_name}" "${smiles_column}" "${label_column}"
mkdir ${dataset_name}
mv ${dataset_name}_scaffold_*.csv ${dataset_name}/
export task=${dataset_name}
</code></pre>
<blockquote>
<ol start="2">
<li>Training the molecular property prediction model</li>
</ol>
</blockquote>
<p>We have presented three approaches for molecular property modeling, namely AutoQSAR (broad applicability, slow speed), PropDecoder (fast speed), and FineTuning (optimal performance, moderate speed).</p>
<p>Given that you have enough experience with hyperparameter tuning, the attainment of optimal performance can be accomplished through the utilization of the FineTuning script to invoke GeminiMol. Also, AutoQSAR is recommended if you lack experience with hyperparameter tuning.</p>
<blockquote>
<p>2.1 Fine-Tuning on downstream task</p>
</blockquote>
<pre><code class="language-shell">export task="Your_Dataset" # Specify a path to your datasets (train, valid, and test)
export smiles_column="SMILES" # Specify the column name in datasets
export label_column="Label" # Specify the column name in datasets
python -u ${geminimol_app}/FineTuning.py "${task}" "${geminimol_lib}/GeminiMol" "${smiles_column}" "${label_column}" "${task}_GeminiMol"
</code></pre>
<blockquote>
<p>2.2 AutoQSAR (AutoGluon)</p>
</blockquote>
<p>It is recommended to try using AutoQSAR to call CombineFP or GeminiMol when you lack deep learning experience, which usually produces a model with good performance.</p>
<pre><code class="language-shell">export encoder_method="${geminimol_lib}/GeminiMol" # only GeminiMol
</code></pre>
<p>In our paper, we introduced a powerful joint molecular fingerprint baseline method named <a href="http://CombineFP.In">CombineFP.In</a> our experiments, the performance of CombineFP in molecular property modeling is very superior and we highly recommend trying CombineFP along with GeminiMol.</p>
<pre><code class="language-shell">export encoder_method="ECFP4:AtomPairs:TopologicalTorsion:FCFP6" # CombineFP
</code></pre>
<p>Having defined the encoder, you can train the model to convert the encoding of the molecule into properties using AutoQSAR. In fact, a potential advantage of this over FineTuning is that it can decode diverse molecular properties based on the fixed encoding, which will speed up the efficiency of chemical space searching.</p>
<pre><code class="language-shell">export task="Your_Dataset" # Specify a path to your datasets (train, valid, and test)
export smiles_column="SMILES" # Specify the column name in datasets
export label_column="Label" # Specify the column name in datasets
python -u ${geminimol_app}/AutoQSAR.py "${task}" "${encoder_method}" "${smiles_column}" "${label_column}" "" "${task}_GeminiMol"
</code></pre>
<p>If the integration of molecular fingerprints and a pre-trained GeminiMol model is desired for training a molecular property prediction model, either PropDecoder or AutoQSAR can be employed.</p>
<pre><code class="language-shell">export fingerprints="ECFP4:AtomPairs:TopologicalTorsion:FCFP6:MACCS" # CombineFP+MACCS
export encoder_method="${geminimol_lib}/GeminiMol:${fingerprints}" # CombineFP+MACCS+GeminiMol
export task="Your_Dataset" # Specify a path to your datasets (train, valid, and test)
export smiles_column="SMILES" # Specify the column name in datasets
export label_column="Label" # Specify the column name in datasets
python -u ${geminimol_app}/AutoQSAR.py "${task}" "${encoder_method}" "${smiles_column}" "${label_column}" "" "${task}_GMFP"
</code></pre>
<blockquote>
<p>2.3 PropDecoder</p>
</blockquote>
<p>For the most tasks, performing fine-tuning or using AutoQSAR will give pretty good performance in molecular property modeling, so you don't need to try PropDecoder unless the first two give poor performance.</p>
<pre><code class="language-shell">export task="Your_Dataset" # Specify a path to your datasets (train, valid, and test)
export smiles_column="SMILES" # Specify the column name in datasets
export label_column="Label" # Specify the column name in datasets
python -u ${geminimol_app}/PropDecoder.py "${task}" "${encoder_method}" "${smiles_column}" "${label_column}" "${task}_GeminiMol"
</code></pre>
<blockquote>
<ol start="3">
<li>Make predictions (only for AutoQSAR or fine-Tuned models)</li>
</ol>
</blockquote>
<p>Next, we can load the model trained based on <code>AutoQSAR</code> and <code>FineTuning</code> to predict molecular properties in a new dataset.</p>
<pre><code class="language-shell">export model_path="QSAR_GeminiMol" # ${task}_GeminiMol when your build QSAR model
export encoder_method="${geminimol_lib}/GeminiMol" # Match to the encoders selected during QSAR model training
export extrnal_data="dataset.csv" # must contain the ${smiles_column}
export smiles_column="SMILES" # Specify the column name in datasets
export model_type="FineTuning" # FineTuning, PropDecoder, ['LightGBM', 'LightGBMLarge', 'LightGBMXT', 'NeuralNetTorch'] for AutoQSAR
python -u ${geminimol_app}/PropPredictor.py "${model_path}" "${encoder_method}" "${extrnal_data}" "${smiles_column}" "${model_type}"
</code></pre>
<p>If you have constructed a regression model using AutoQSAR, refer to the following command.</p>
<pre><code class="language-shell">export model_path="QSAR_GeminiMol" # ${task}_GeminiMol when your build QSAR model
export encoder_method="${geminimol_lib}/GeminiMol" # Match to the encoders selected during QSAR model training
export extrnal_data="dataset.csv" # must contain the ${smiles_column}
export smiles_column="SMILES" # Specify the column name in datasets
export model_type="NeuralNetTorch" # ['LightGBM', 'LightGBMLarge', 'LightGBMXT', 'NeuralNetTorch'] for AutoQSAR
export task_type="regression"
python -u ${geminimol_app}/PropPredictor.py "${model_path}" "${encoder_method}" "${extrnal_data}" "${smiles_column}" "${model_type}" "${task_type}"
</code></pre>
<h4 id="molecular-clustering">Molecular Clustering</h4>
<p>You can use GeminiMol to cluster molecules just like molecular fingerprints!</p>
<pre><code class="language-shell">export encoder_method="${geminimol_lib}/GeminiMol" # Match to the encoders selected during QSAR model training
export data_table="dataset.csv" # must contain the ${smiles_column}
export smiles_column="SMILES" # Specify the column name in datasets
export output_fn="Cluster"
export cluster_num=10 # only for supervised clustering algorithm, such as K-Means
python -u ${geminimol_app}/Analyzer.py "${data_table}" "${encoder_method}" "${smiles_column}" "${output_fn}" "cluster:${cluster_num}"
</code></pre>
<h4 id="extract-molecular-features-geminimol-encoding">Extract Molecular Features (GeminiMol Encoding)</h4>
<p>You can use GeminiMol or molecular fingerprints to extract molecular features for further analysis.</p>
<pre><code class="language-shell">export encoder_method="${geminimol_lib}/GeminiMol" # Match to the encoders selected during QSAR model training
export data_table="dataset.csv" # must contain the ${smiles_column}
export smiles_column="SMILES" # Specify the column name in datasets
export output_fn="${data_table%%.*}_Encoding"
python -u ${geminimol_app}/Analyzer.py "${data_table}" "${encoder_method}" "${smiles_column}" "${output_fn}" "encode"
</code></pre>
<h2 id="-reproducing">👐 Reproducing</h2>
<p>Here, we present the reproducible code for training the Cross-Encoder and GeminiMol models based on the CSS descriptors of 39,290 molecules described in the paper.</p>
<h4 id="download-training-and-benchmark-datasets">Download Training and Benchmark Datasets</h4>
<blockquote>
<p>Download all datasets via <a href="https://zenodo.org/records/10450788">Zenodo</a> for training, test and benchmark</p>
</blockquote>
<pre><code class="language-shell"> cd ${geminimol_data}
wget https://zenodo.org/records/10450788/files/css_library.zip # only for reproducing GeminiMol training
wget https://zenodo.org/records/10450788/files/Benchmark_DUD-E.zip # only for reproducing benchmark
wget https://zenodo.org/records/10450788/files/Benchmark_LIT-PCBA.zip # only for reproducing benchmark
wget https://zenodo.org/records/10450788/files/Benchmark_QSAR.zip # only for reproducing benchmark
wget https://zenodo.org/records/10450788/files/Benchmark_TIBD.zip # only for reproducing benchmark
wget https://zenodo.org/records/10450788/files/Chem_SmELECTRA.zip # only for reproducing cross-encoder baseline
</code></pre>
<h4 id="re-training-our-models">Re-training our models</h4>
<blockquote>
<p>Training the Cross-Encoder</p>
</blockquote>
<pre><code class="language-shell">conda activate GeminiMol
export model_name="CrossEncoder"
export batch_size_per_gpu=200 # batch size = 200 (batch_size_per_gpu) * 4 (gpu number)
export epoch=20 # max epochs
export lr="1.0e-3" # learning rate
export label_list="MCMM1AM_MAX:LCMS2A1Q_MAX:MCMM1AM_MIN:LCMS2A1Q_MIN" # ShapeScore:ShapeAggregation:ShapeOverlap:CrossSim:CrossAggregation:CrossOverlap
CUDA_VISIBLE_DEVICES=0,1,2,3 python ${geminimol_app}/CrossEncoder_Training.py "${geminimol_data}/css_library/" "${geminimol_data}/Chem_SmELECTRA" "${epoch}" "${lr}" "${batch_size_per_gpu}" "${model_name}" "${geminimol_data}/benchmark.json" "${label_list}"
</code></pre>
<blockquote>
<p>Training the GeminiMol Encoder and Decoder of CSS descriptors</p>
</blockquote>
<pre><code class="language-shell">conda activate GeminiMol
export model_name="GeminiMol"
export batch_size=512
export epoch=20 # max epochs
export patience=50 # for early stoping
export GNN='WLN' # Weisfeiler-Lehman Network (WLN)
export network="MeanMLP:2048:4:2048:None:0:5:0"
export label_dict="ShapeScore:0.2,ShapeAggregation:0.2,ShapeOverlap:0.05,ShapeDistance:0.05,CrossSim:0.15,CrossAggregation:0.15,CrossDist:0.05,CrossOverlap:0.05,MCS:0.1"
CUDA_VISIBLE_DEVICES=0 python -u ${geminimol_app}/GeminiMol_Training.py "${geminimol_data}/css_library/" "${epoch}" "${batch_size}" "${GNN}" "${network}" "${label_dict}" "${model_name}" "${patience}" "${geminimol_data}/benchmark.json"
</code></pre>
<h4 id="benchmarking-the-fingerprints-and-our-models">Benchmarking the fingerprints and our models</h4>
<p>Additionally, benchmark test scripts were provided. With this code, the community can reproduce the results reported in the paper, explore different model architectures, even incorporate additional molecular similarity data to further enhance the performance of the models.</p>
<blockquote>
<p>Benchmarking molecular fingerprints and GeminiMol on virutual screening and target identification</p>
</blockquote>
<p>For each molecular fingerprint, we used all supported similarity metrics, including Tanimoto, Cosine, and Tversky. For the GeminiMol model, in addition to the projected heads used in pre-training, we introduced similarities between molecular representation vectors, including Cosine and Pearson. It is worth noting that in practice we cannot be sure which combination of molecular fingerprints and similarity metrics is optimal, and therefore each combination is considered an independent method in benchmarking.</p>
<pre><code class="language-shell">conda activate GeminiMol
<span class="hljs-meta prompt_"># </span><span class="language-bash">benchmarking Fixed GeminiMol models and Fingerprints</span>
for task in "DUDE" "LIT-PCBA" "TIBD" # zero-shot tasks
do
for model_name in "FCFP6" "MACCS" "RDK" "ECFP6" "FCFP4" \
"TopologicalTorsion" "AtomPairs" "ECFP4" \
"${geminimol_lib}/GeminiMol"
do
mkdir -p ${model_name}
CUDA_VISIBLE_DEVICES=0 python -u ${geminimol_app}/benchmark.py "${model_name}" "${geminimol_data}/benchmark.json" "${task}"
done
done
</code></pre>
<blockquote>
<p>Benchmarking molecular fingerprints and GeminiMol on molecular property modeling</p>
</blockquote>
<p>It is worth noting that different decoders exhibit varying performance on different tasks and encodings. Therefore, it is essential to select the appropriate decoder for each specific molecular encoder and task. In practice, we can determine when the model should stop-training and choose the optimal decoder architecture by dividing the training, validation and test sets. Consequently, all results should be merged using a data pivot table to analyze the optimal decoder for each encoder-task combination. In our work, the hyperparameters of the PropDecoder were chosen based on empirical experience and were not subjected to any hyperparameter tuning. Performing further hyperparameter tuning for each task may potentially yield improved performance.</p>
<pre><code class="language-shell">for task in "ADMET-C" "ADMET-R" \
"LIT-QSAR" "CELLS-QSAR" "ST-QSAR" "PW-QSAR" \
"PropDecoder-ADMET" "PropDecoder-QSAR" # fixed the molecular encoder
do
for model_name in "CombineFP" \
"FCFP6" "MACCS" "RDK" "ECFP6" "FCFP4" "TopologicalTorsion" "AtomPairs" "ECFP4" \
"${geminimol_lib}/GeminiMol"
do
mkdir -p ${model_name}
CUDA_VISIBLE_DEVICES=0 python -u ${geminimol_app}/benchmark.py "${model_name}" "${geminimol_data}/benchmark.json" "${task}"
done
done
for task in "FineTuning-ADMET" "FineTuning-QSAR"; do # benchmarking with FineTuning GeminiMol models
for model_name in "${geminimol_lib}/GeminiMol"; do
CUDA_VISIBLE_DEVICES=0 python -u ${geminimol_app}/benchmark.py "${model_name}" "${geminimol_data}/benchmark.json" "${task}"
done
done
</code></pre>
<h2 id="-citing-this-work">⭐ Citing This Work</h2>
<p><strong>Conformational Space Profiling Enhances Generic Molecular Representation for AI-Powered Ligand-Based Drug Discovery</strong><br>
Lin Wang, Shihang Wang, Hao Yang, Shiwei Li, Xinyu Wang, Yongqi Zhou, Siyuan Tian, Lu Liu, Fang Bai<br>
<em>Advanced Science</em>, 2024; doi: 10.1002/advs.202403998</p>
<h2 id="-license">✅ License</h2>
<p>GeminiMol is released under the Academic Free Licence, which permits academic use, modification and distribution free of charge. GeminiMol can be utilized in academic publications, open-source software projects, and open-source competitions (e.g. Kaggle competitions under the MIT Open Source license).</p>
<p>GeminiMol prohibits unauthorised commercial use, including commercial training and as part of a paid computational platform, which intended to prevent speculators from exploiting informational asymmetry for profit. Communication and authorization with <a href="[email protected]">our supervisor</a> is permitted for its application in pipeline development and research activities within pharmaceutical R&D.</p>
<h2 id="-get-in-touch">💌 Get in Touch</h2>
<p>We welcome community contributions of extension tools based on the GeminiMol model, etc. If you have any questions not covered in this overview, please contact the <a href="[email protected]">GeminiMol Developer Team</a>. We would like to hear your feedback and understand how GeminiMol has been useful in your research. Share your stories with <a href="[email protected]">us</a>.</p>
<h2 id="-acknowledgements">😃 Acknowledgements</h2>
<p>We appreciate the technical support provided by the engineers of the high-performance computing cluster of ShanghaiTech University. Lin Wang also thanks Jianxin Duan, Gaokeng Xiao, Quanwei Yu, Zheyuan Shen, Shenghao Dong, Huiqiong Li, Zongquan Li, and Fenglei Li for providing technical support, inspiration and help for this work. We express our gratitude to Dr. Zhongji Pu, Dr. Quanwei Yu for their invaluable assistance in third-party testing for model installation, reproducibility and application.</p>
<p>We also thank the developers and maintainers of MarcoModel and PhaseShape modules in the Schrödinger package. Besides, GeminiMol communicates with and/or references the following separate libraries and packages, we thank all their contributors and maintainers!</p>
<ul>
<li><a href="https://www.rdkit.org/"><em>RDKit</em></a></li>
<li><a href="https://pytorch.org/"><em>PyTorch</em></a></li>
<li><a href="https://auto.gluon.ai/stable/index.html"><em>AutoGluon</em></a></li>
<li><a href="https://lifesci.dgl.ai/"><em>DGL-Life</em></a></li>
<li><a href="https://oddt.readthedocs.io/en/latest/"><em>ODDT</em></a></li>
<li><a href="https://scipy.org/"><em>SciPy</em></a></li>
<li><a href="https://scikit-learn.org/stable/"><em>scikit-learn</em></a></li>
<li><a href="https://matplotlib.org/"><em>matplotlib</em></a></li>
</ul>
<script async src="https://cdn.jsdelivr.net/npm/katex-copytex@latest/dist/katex-copytex.min.js"></script>
</body>
</html>