QSAR-Complete (QComp)

A robust, interpretable, non-iterative imputation framework for sparse datasets in drug discovery.

This is the repo for the manuscript: https://arxiv.org/pdf/2405.11703

Install dependencies

On Linux:

conda create --name qcomp
conda activate qcomp
conda install pytorch torchvision torchaudio cpuonly -c pytorch
pip install deepchem
pip install tensorflow ## deepchem requires tensorflow
pip install matplotlib

Run

python main.py

Public ADMET datasets and QSAR results

The ADMET data compiled from various public sources and the corresponding Chemprop multitask model predictions are located under public_data_results. The dataset is randomly split to 80 % training and 20 % test sets using 5-folds. Details on the files under public_data_results:

• all_data: public_admet_data_all.csv contains all data for 25 different ADMET assays along with SMILES strings and molecular weights of compounds. data_count_name_unit_info.csv contains detailed information and unit of each ADMET assay. The dataset is very sparse. data_overlap_count_between_prop.csv shows the number of compound overlaps between each pair of assays. spearman_corr_heatmap.pdf shows the Spearman correlation heatmap generated for the assay pair that has at least 10 overlapping compounds.
• random_split_data_results: contains training and test sets for each fold. In each fold, chemprop_multitask_pred folder contains the predictions from Chemprop multitask model (e.g. public_admet_data_random_fold_0_test_set_model_pred.csv) and the ensemble variance of the model predictions (e.g. public_admet_data_random_fold_0_test_set_model_ensemble_variance.csv).
• result_figs: pred_comparison_RF_Chemprop_single_multitask.pdf shows the comparison among Random Forest (RF), Chemprop single-task, and Chemprop multi-task models. The RF model uses the Morgan Fingerprints and MOE2D descriptors. The errors are evaluated over 5-fold cross validation on the random split, and the error bars represent the standard deviations among the 5-folds.