An official website of the United States government
Predicts combinations of drug responses under different experimental configurations.
Data scientists interested in bioinformatics, computational cancer biology, drug discovery, and machine learning.
Enables predictions of drug responses under different experimental configurations.
Background
The National Cancer Institute drug pair screening effort against 60 well-characterized human tumor cell lines (NCI-60) presents an unprecedented resource for modeling combinational drug activity.
Results
The authors present a computational model for predicting cell line response to a subset of drug pairs in the NCI-ALMANAC database. Based on residual neural networks for encoding features as well as predicting tumor growth, the model explains 94% of the response variance. While the best result is achieved with a combination of molecular feature types (gene expression, microRNA and proteome), the authors show that most of the predictive power comes from drug descriptors. To further demonstrate value in detecting anticancer therapy, the authors rank the drug pairs for each cell line based on model predicted combination effect and recover 80% of the top pairs with enhanced activity.
Conclusions
The authors present promising results in applying deep learning to predicting combinational drug response. The authors' feature analysis indicates screening data involving more cell lines are needed for the models to make better use of molecular features.
The objective was to create a model that combines the activity of anticancer drugs with deep neural networks for identifying any synergistic interactions of anticancer drugs that can overcome the inherent heterogeneity of tumors and prevent emergence of drug resistance in cell subpopulations.
This resource is related to other drug response models Unified Drug Response Predictor and Single Drug Response Predictor. E-COXEN could be used for feature selection of genes.
Data scientists can use multiple machine learning techniques to predict drug response. The general rule is that classical methods like random forests would perform better for small datasets, while neural network approaches like Combo would perform better for relatively larger datasets.
Data scientists can train the provided untrained model with the shared preprocessed data or with their own preprocessed data, or can use the trained model to predict the drug response from the NCI-ALMANAC study. The provided scripts use data that have been downloaded from NCI-ALMANAC and normalized.
The following components are in the Model and Data Clearinghouse (MoDaC):
- The Cancer Drug Response Prediction Dataset contains processed training and test data.
- The Combination Drug Response Predictor (Combo) asset contains the untrained model topology, and the trained model weights to be used in inference.
- Type of data required: Genomic expression, dose response, and drug feature data from multiple sources
- Source of data required:
- Combo drug response screening results from NCI-ALMANAC (public data);
- 5-platform normalized expression, microRNA expression, and proteome abundance data from the NCI;
- Dragon7 generated drug descriptors based on 2D chemical structures from NCI
- Public vs. Restricted: Public
Using cell line response data for a subset of drug pairs in the NCI-ALMANAC database, the model explains 94% of the response variance. While the best result is achieved with a combination of molecular feature types (gene expression, microRNA, and proteome), most of the predictive power comes from drug descriptors. To further demonstrate value in detecting anticancer therapy, drug pairs are ranked for each cell line based on the model's predicted combination effect. Over 80% of the top pairs with enhanced activity were recovered.
Accuracy of dose-response predictions