ScaffComb: A Phenotype-Based Framework for Drug Combination Virtual Screening in Large-scale Chemical Datasets

This repository contains the source codes, the example data and some analysis codes.

ScaffComb

Dependences

• python 3.6.8
• tensorflow-gpu 1.12.0
• tensorboardx 2.0
• pytorch 1.3.1
• numpy 1.17.0
• rdkit 2020.03.1b1.0
• scipy 1.2.1
• scikit-learn 0.21.2
• matplotlib 3.1.1

Demos

Training GSG

> from GSG import AttnDGenerator
> dt_L1000 = pickle.load(open('data/data_GSG_input_train_example.pkl','rb'))
> dt_chembl = pickle.load(open('data/data_GSG_input_pretrain_example.pkl', 'rb'))
> tokens = pickle.load(open('data/tokens.pkl','rb'))
> # Pre-training with chembl scaffolds
> Trainer0 = AttnDGenerator(tokens, dt_chembl, batch_size=128, lr = 0.0001,
                           device=device, path='rst/GSG_pretrained/')
> Trainer0.train(1000)
> # Train with L1000 scaffolds
> Trainer1 = AttnDGenerator(tokens, dt_L1000, batch_size=128, lr = 0.0001,
                           device=device, path='rst/GSG_trained/')
> Trainer1.load_model('rst/GSG_pretrained/', 999)
> Trainer1.train(1000)

Generate scaffolds with GSG

> scaf = Trainer1.G_sample(gene = gene)

Train DSP

> from DSP import trainer
> X, Y = pickle.load(open('data/data_DSP_input_example.pkl', 'rb'))
> X_ = X.copy()
> names = ['~'.join(sorted(x.split('~')[:2])) for x in X_.index.tolist()]
> X_.index = names

> X1 = X_.iloc[:, :1593]
> X2 = X_.iloc[:, 1593:3186]
> X3 = X_.iloc[:, 3186:]

> model = AugSSP(1593, 978, device=device)
> Trainer3 = trainer([X1, X2, X3, Y], model, batch_size=512, 
                    path = 'rst/DSP_trained/model')
> Trainer3.load_model()
> Trainer3.one_fold_train(500)

Predict SS with DSP

> y_pred = Trainer3.predict([X1, X2, X3])

Train SDSP

> from SDSP import trainer
> tokens = pickle.load(open('data/tokens.pkl','rb'))
> dt = pickle.load(open('data/data_SDSP_input_example.pkl', 'rb'))
> dt_train = [x[:40] for x in dt]
> dt_test = [x[40:] for x in dt]
> Trainer4 = trainer(tokens, dt_train, dt_test, latent_dim=512,
                    batch_size = 512, lr = 0.0001,
                    device=device, path='rst/SDSP_trained/model')
> Trainer4.train(500)

Predict SS with SDSP

> from SDSP import SDSP_predict
> y_pred, y_pred_rev = SDSP_predict(Trainer4, dt)

Citation