Kinome-Wide Virtual Screening by Multi-Task Deep Learning

Modeling and Interpretability Study of the Structure-Activity Relationship for Multigeneration EGFR Inhibitors

The fourth-generation EGFR inhibitors targeting L858R/T790M/C797S mutations are in clinical trials mostly, and it is necessary to develop new inhibitors. In this study, an internal data set containing 2302 multitarget EGFR inhibitors targeting the wild type (83%) and the L858R (92%), L858R/T790M (96%), and L858R/T790M/C797S (60%) mutations was collected. We established a structure-activity relationship model for predicting the bioactivities of multigeneration EGFR inhibitors by a multitask deep neural network (MT-DNN). We also constructed four single-task models on 1384 L858R/T790M/C797S (60%) mutation inhibitors by support vector machine (SVM), random forest (RF), XGBoost (XGB), and single-target neural network (ST-DNN), respectively. The MT-DNN model significantly outperformed single-task models on the external data set of 304 fourth-generation EGFR inhibitors. Furthermore, the combined application of MT-DNN and SHAP/delta-SHAP value interpretability analysis offers rigorous structural information from a global perspective. With SHAP/delta-SHAP methods, the MT-DNN model can mine the core scaffold and important fragments of multigeneration EGFR inhibitors and provide valuable information from a structure-activity relationship perspective to address the resistant mutation problem.

Pan-Cancer Drug Sensitivity Prediction from Gene Expression using Deep Learning

Cancer is a group of complex diseases, with tumor heterogeneity, durable drug efficacy, emerging resistance, and host toxicity presenting major challenges to the development of effective cancer therapeutics. While traditionally used methods have remained limited in their capacity to overcome these challenges in cancer drug development, efforts have been made in recent years toward applying "big data" to cancer research and precision oncology. By curating, standardizing, and integrating data from various databases, we developed deep learning architectures that use perturbation and baseline transcriptional signatures to predict efficacious small molecule compounds and genetic dependencies in cancer. A series of internal validations followed by prospective validation in prostate cancer cell lines were performed to ensure consistent performance and model applicability. We report SensitivitySeq, a novel bioinformatics tool for prioritizing small molecule compounds and gene dependencies in silico to drive the development of targeted therapies for cancer. To the best of our knowledge, this is the first supervised deep learning approach, validated in vitro, to predict drug sensitivity using baseline cancer cell line gene expression alongside cell line-independent perturbation-response consensus signatures.