Introduction

CaDRReS-Sc is an AI/ML framework for robust cancer drug response prediction based on single-cell RNA-sequencing (scRNA-seq) data. It extends an existing recommender system model (CaDRReS, Suphavilai et al., 2018) with new features calibrated for diverse cell types, and accounting for tumor heterogeneity (Suphavilai et al., 2020). In addition to monotherapy response, CaDRReS-Sc can also predict response to combinatorial therapy targeting tumor sub-clones.

Key features:

• Predicts cancer drug response from transcriptomic profiles for (a) single-cells, (b) cell clusters (e.g. using Scanpy, see tutorial notebook), and (c) from bulk analysis.
• Includes a pre-trained model based on the GDSC database.
• Reports half-maximal inhibitory concentrations (IC50) as well as predicted cell death percentages for a specific drug dosage (see tutorial).
• Robust to gene expression measurements across diverse platforms, including microarrays, RNA-seq and scRNA-seq (Suphavilai et al., 2018, Suphavilai et al., 2020).
• Interpretable model based on a latent "pharmacogenomic space" (Suphavilai et al., 2018, Suphavilai et al., 2020).
• Flexibility to train a new model based on other drug response datasets (e.g. CTRP, see model training tutorial).

Usage

git clone https://github.com/CSB5/CaDRReS-Sc.git

CaDRReS-Sc is based on Python 3.x

Required packages

• Pandas
• Numpy
• TensorFlow 1.14

Optional package

• Scanpy (for single-cell clustering)

Usage examples

Please refer to our tutorial notebooks. Below are snippets from the notebooks:

Model training (tutorial)

model_dict, output_dict = model.train_model_logistic_weight(
    Y_train, X_train,                 # Y = drug response; X = kernel features
    Y_test, X_test, 
    sample_weights_logistic_x0_df,    # Sample weight w.r.t. maximum drug dosage
    indication_weight_df,             # High weight for specific tissue types
    10, 0.0, 100000, 0.01,            # Hyperparamters
    model_spec_name=model_spec_name,  # Select objective function
    save_interval=5000,
    output_dir=output_dir)

Predicting monotherapy response based on a pre-trained model (tutorial)

cadrres_model = model.load_model(model_file)
pred_df, P_df = model.predict_from_model(cadrres_model, X_test, model_spec_name)

pred_df.head() # Predicted drug response (log2 IC50)

	D1	D1001	D1003	D1004	...
906826	3.86	11.11	-5.71	-5.56	...
687983	7.00	11.52	-4.12	-4.48	...
910927	1.74	10.84	-7.32	-6.77	...
1240138	3.55	11.42	-4.79	-4.82	...
1240139	2.79	10.70	-7.89	-7.42	...

P_df.head() # A latent vector of each sample in the 10D pharmacogenomic space

	1	2	3	4	5	6	7	8	9	10
906826	0.38	-1.39	-1.26	-0.15	-0.37	-1.35	1.09	0.04	-0.78	0.37
687983	0.26	-0.68	0.47	1.31	0.61	0.93	-0.09	-0.77	-2.20	-0.42
910927	-0.52	0.47	0.12	-0.10	-1.56	-2.99	1.15	-0.06	-0.58	-1.21
1240138	0.72	0.51	-1.16	-0.34	1.56	-1.21	1.06	-0.59	0.08	-0.53
1240139	-0.08	-0.45	0.45	-0.19	-0.75	-2.93	0.50	0.67	-0.11	0.27

Predicting response to drug combinations at specific dosages (tutorial)

Inputs:

freq_df Proportion of each cell cluster in each sample

cluster_gene_exp_df Cluster-specific gene expression profiles

drug_df Drug dosage information

Example outputs:

drug_combi_pred_df.head() # Predicted cell death percentage at specific dosage

patient	drug_id_A	drug_id_B	cell_death_A	cell_death_B	cell_death_combi	improve
HN120	D1007	D133	8.52	75.63	77.19	1.56
HN120	D1007	D201	8.52	60.39	63.04	2.65
HN120	D1007	D1010	8.52	15.31	22.39	7.08
HN120	D1007	D182	8.52	64.66	67.22	2.56
HN120	D1007	D301	8.52	63.39	66.39	3.00

Citation

Suphavilai C, Chia S, Sharma A et al. Predicting heterogeneity in clone-specific therapeutic vulnerabilities using single-cell transcriptomic signatures. Genome Medicine 13, Article number: 189 (2021)

Suphavilai C, Bertrand D, Nagarajan N. Predicting cancer drug response using a recommender system. Bioinformatics 2018 Nov 15;34(22):3907-14.

Contact information

For additional information, help and bug reports please email Chayaporn Suphavilai (suphavilaic@gis.a-star.edu.sg)