Awesome
SQUID :squid:
This repository contains the necessary scripts to train and evaluate the 3D generative model SQUID from the paper:
Equivariant Shape-Conditioned Generation of 3D Molecules for Ligand-Based Drug Design
Paper Link: https://arxiv.org/abs/2210.04893
<video src="https://user-images.githubusercontent.com/52709065/194377661-7a915fad-898e-405c-a474-640ac95ad708.mp4" controls="controls" style="max-width: 730px;"> </video>This video demonstrates how SQUID can generate chemically diverse molecules for arbitrary molecular shapes. Specifically, we took a large (>25 heavy atoms) molecule from our test set and used RDKit to generate multiple diverse conformations for this flexible drug-like molecule. For each conformation, we used SQUID to encode the conformer's shape, and then sampled up to 50 novel molecules from the prior. We display the most shape-similar generated molecule (in red), overlaid on the target shape (in blue). The generated conformers have not been post-processed or aligned; their 3D conformations and poses are those directly generated by SQUID.
Downloading Data
Before running any scripts, please download the necessary data (>50 GB when unzipped) from:
https://figshare.com/s/3d2f8fd57d9a65fe237e
This data includes:
- The original train/test sets taken from Molecular Sets (MOSES): A Benchmarking Platform for Molecular Generation Models
- Our fragment library, which is extracted from the training set (100 ring-containing fragments and 24 unique atom types)
- Our filtered train/val/test splits, which remove molecules with fragments not included in our fragment library
- The RDKit-generated 3D conformers for all data splits, including the test-set molecules used in the experiments. We include the conformers both before and after we fix acyclic bond distances/angles. 3D molecules with fixed bonding geometries are typically denoted as "artificial" molecules.
- The pre-processed training data required to (re-)train both the graph generator and the rotatable bond scorer.
You can find the scripts used to generate all these data, along with further instructions, in the directory dataset_generation/
The downloaded data also includes the SQUID-generated 3D molecules for our Shape-Conditioned Generation of Chemically Diverse Molecules experiment, along with the (encoded) target molecules. We include these encoded/decoded molecules for user convenience, as re-running this experiment requires access to an (academic) OpenEye license.
Dependencies
You will also need to create a new Python (conda) environment with the dependencies listed in environment.yml
. The core dependencies needed to run the scripts and notebooks in this repository are as follows:
- notebook (6.4.11) (for running the Jupter Notebook demonstration)
- python (3.10.4)
- cudatoolkit (11.3.1)
- networkx (2.7.1)
- numpy (1.22.3)
- torch (1.11.0) with (cuda 11.3)
- torch-geometric (2.0.4)
- torch-scatter (2.0.9)
- torch-sparse (0.6.13)
- torchaudio (0.11.0)
- torchvision (0.12.0)
- pandas (1.4.2)
- scipy (1.8.1)
- tqdm(4.64.0)
- rdkit (2022.03.2)
- openeye-toolkits (2022.1.1)
To use the openeye toolkits (for optimally aligning 3D molecules with ROCS and computing aligned shape similarities), you will need access to an OpenEye license. Please see https://www.eyesopen.com/academic-licensing for details. After alignment, we always compute and report shape similarity using our own implementation of the (Gaussian-based) Shape Tanimoto (Eq. 1 in our paper). Hence, any other 3D shape-alignment program (besides ROCS) could be substituted if you cannot obtain access to OpenEye. One option is Shaep (https://users.abo.fi/mivainio/shaep/index.php), which is free to use but is significantly slower than ROCS.
Directory Organization
This directory is organized as follows:
-
dataset_generation/
contains scripts to process user-provided csv files of SMILES strings into 3D conformers (for evaluations) and training data -
models/
contains the implementation of SQUID -
utils/
contains support functions used across training and generation -
trained_models/
contains trained models for the graph-generator and the rotatable bond scorer -
MO_virtual_screening/
contains scripts used for virtual screening (VS) in our Shape-Constrained Molecular Optimization experiment -
train_graph_generator.py
andtrain_scorer.py
contain scripts to train the graph generator and the scorer, respectively. -
shape_conditioned_generation_evaluations.py
contains the script to evaluate SQUID in our Shape-Conditioned Generation of Chemically Diverse Molecules experiment. -
shape_conditioned_generation_dataset_baseline.py
contains the script to compute the dataset baseline for our S_hape-Conditioned Generation of Chemically Diverse Molecules_ experiment. -
shape_constrained_optimization_evaluations.py
contains the script to run the genetic algorithm for our Shape-Constrained Molecular Optimization experiment.
Notebook Demonstration
RUN_ME.ipynb
provides a lightweight, interactive demonstration of how we can easily use SQUID to generate chemically diverse molecular analogues with high shape similarity to an encoded molecule.
(Re)-Training
After downloading the training data, you can train the graph generator and the scorer (with their default settings, on 1 gpu) by running:
python train_graph_generator.py
python train_scorer.py
Running Experiments
You can re-generate the SQUID-generated molecules analyzed in our Shape-Conditioned Generation of Chemically Diverse Molecules experiment by running:
python shape_conditioned_generation_evaluations.py {experiment_name} {lambda_interp} {stop_threshold}
For instance, to generate 50 samples for 1000 target molecular shapes in the test set using $\lambda = 1.0$ (e.g., sampling from the prior), as performed in our paper, run:
python shape_conditioned_generation_evaluations.py lambda10 1.0 0.01
This will create pickle files containing the generated molecules and the (encoded) target molecules. We include these files in paper_results/
, which can be downloaded along with the training data (from https://figshare.com/s/3d2f8fd57d9a65fe237e). These generated molecules can then be filtered by tanimoto similarity to the target and sampled to generate the histograms of shape similarity (vs. chemical similarity) in Figure 3 of our paper.
You can run the Shape-Constrained Molecular Optimization experiment by running:
python shape_constrained_optimization_evaluations.py {experiment_name} {objective} {mol_index}
The objective is one of GSK3B
, JNK3
, Osimertinib_MPO
, Sitagliptin_MPO
, Celecoxib_Rediscovery
, or Thiothixene_Rediscovery
. The mol_index
corresponds to the index of a seed molecule in the test set. In our paper, we use different molecules for each objective. In particular, we use:
GSK3B
: (99300, 142337, 94211, 13059, 138951, 67478, 128739, 70016)
JNK3
: (2775, 7994, 10770, 108203, 126430, 9126, 128739, 70016)
Osimertinib_MPO
: (78600, 81366, 46087, 76561, 87747, 91918, 128739, 70016)
Sitagliptin_MPO
: (118822, 132656, 130062, 113584, 115006, 140953, 128739, 70016)
Celecoxib_Rediscovery
: (33351, 14473, 101938, 6686, 1200, 69153, 128739, 70016)
Thiothixene_Rediscovery
: (25628, 25659, 56430, 137033, 48156, 68289, 128739, 70016)
Hence, you can optimize GSK3B
while restricting the shape similarity to molecule # 99300 from our test set by running:
python shape_constrained_optimization_evaluations.py GSK3B_99300 GSK3B 99300
MO_virtual_screening/
contains the scripts used for virtual screening (VS) baseline, which includes screening for 1) each objective and 2) shape similarity to each of the seed (target) molecules.
Notes
Finally, please note that this repository will be continuously updated to improve the usability of SQUID.