This code is used to benchmark model performance for high throughput cell line screens.

Data Folder Requirements

• drugfeats.pkl contains mOrdred descritpors without 3D features, that are scaled and imputed
• imputer.pkl contains a dictionary {imputer, scaler} to be used for on the fly feature generation
• rnaseq.pkl contains columns with label combo_auc.DRUG, combo_auc.AUC, combo_auc.CELL
• extended_combined_smiles which matches smiles to combo_auc.DRUG
• cellpickle.pkl contains RNAseq data frame with label lincs.CELL for merging
• extended_combined_smiles contains smiles matching combo_auc.DRUG
• extended_combined_mordred_descriptors contains precomputed mordred descriptors for combo_auc.DRUG
• testsmiles.smi is a SMILES formated file with 100k random sample smiles for testing

Training Instructions

Training is done either with --mode [graph, desc, image] (RNN SMILES coming soon). Use python train.py -h for options.

For this benchmark the following commands were used:

python train.py --mode graph -o saved_models/graph_model.pt -w 32 -s cell
python train.py --mode desc -o saved_models/desc_model.pt -w 32 -s cell
python train.py --mode image -o saved_models/image_model.pt -w 32 -s cell

Throughput Benchmarking Instructions

Again use python infer.py -h to see all options.

For this benchmark the following commands were used:

python infer.py --mode graph -o saved_models/graph_model.pt -w 32 -g 2 --smiles_file data/testsmiles.smi --output_file saved_models/graph_infers.txt
python infer.py --mode desc -o saved_models/desac_model.pt -w 32 -g 2 --smiles_file data/testsmiles.smi --output_file saved_models/desc_infers.txt
python infer.py --mode image -o saved_models/image_model.pt -w 32 -g 2 --smiles_file data/testsmiles.smi --output_file saved_models/image_infers.txt

Results