This project focuses on developing Machine Learning models using both the SMILES molecular descriptors and 3D atomic representations of drug-like compounds targeting SARS-CoV-2 for rapid screening of binding affinity. Work was developed as part of the Lawrence Livermore National Laboratory Data Science Summer Institute 2022 Challenge Problem led by Dr. Hyojin Kim.
To clone this repository:
$ git clone https://github.com/marinadunn/covid19-drug-screening-ML.git COVID19-ML
$ cd COVID19-ML
It's recommended to first create a virtual Python environment, then install dependencies from the file requirements.txt
:
$ python3 -m venv my_env
$ source my_env/bin/activate
$ pip install --upgrade pip
$ pip install -r requirements.txt
-
Exploratory Data Analysis
-
Task 1: Use SMILES molecular descriptors of ligand data to predict binding affinity (Binary Classification)
-
Task 2: Use 3D atomic representation of ligand data to predict binding affinity (Binary Classification)
All development was done with Python 3.10. Some development was done using Jupyter notebooks, while others were done using Python scripts. Model training for Task 2 was performed on Livermore Compute HPCs.
Code for development & testing is located in the respective folders for team members.
Task 1 original data is located in the file mpro_exp_data2_rdkit_feat.csv
, with cleaned datasets located in their respective data folders for each individual.
Task 2 original train, test, and validation datasets can be found in the files postera_protease2_pos_neg_train.hdf5
, postera_protease2_pos_neg_test.hdf5
, and postera_protease2_pos_neg_val.hdf5
respectively, with cleaned datasets once again located in their respective data folders for each individual.
Task 1: - Marina: 11 classifier models tested initially using scikit-learn. Tested for both dataset where outliers greater than 4-sigma were removed, as well as dataset where they were not removed. Best performing models: - KNN: 89% Accuracy, 90% Precision (wgtd.), 89% Recall (wgtd.), 89% F1 (wgtd.) - Gaussian Process: 88% Accuracy, 89% Precision (wgtd.), 88% Recall (wgtd.), 88% F1 (wgtd.) - Neural Network: 89% Accuracy, 89% Precision (wgtd.), 89% Recall (wgtd.), 89% F1 (wgtd.) - Robert: - Logistic Regression Classifier: 60% Accuracy (poor generalization) - Feature Subset with Multi-layer Perceptron: >75% for some models (good generaization), 60%> for others (poor) - Ensemble Approach: 80.5% Accuracy, 74% Precision (wgtd.), 74% Recall (wgtd.), 81.2% F1 (wgtd.)
Ultimately, could not explore further given time constraints. Future considerations: look at additional metrics (i.e. ROC AUC), try model fusion, additional feature analysis
Task 2: - Robert & Lance: - Graph Neural Network: ~80% Accuracy - Graph Neural Network Ensemble: 81.25% Accuracy - Marina: 3D Convolutional Neural Network
Ultimately, could not explore further given time constraints. Future considerations: test more 3D CNN models, explore using uncertainty quantification, do comprehensive performance & behavior analysis, weight "binding" examples more than "non-binding" in GNN, test other kinds of pooling with GNN, try additional ensemble methods, explore additional methods of translating ligand data to graph in GNN, implement fusion model with GNN & 3D CNN
All code was developed as part of the Lawrence Livermore National Laboratory Data Science Summer Institute 2022 Challenge Problem, testing various ML approaches for screening molecular inhibitors for SARS-CoV-2 protein targets. More information can be found here.
Authors: Marina M. Dunn, Lance Fletcher, Robert Stephany
Please send us an email for any support questions or inquiries:
-
Marina M. Dunn (mdunn014@ucr.edu)
-
Lance Fletcher (lance.g.fletcher@gmail.com)
-
Robert Stephany (rrs254@cornell.edu)