3O - Pre-clinical pan-cancer drug repurposing via deep learning

Background

Cancer is a complex disease by nature, with tumor heterogeneity and durable drug efficacy presenting major challenges to the development of effective cancer therapeutics. Recently, efforts have been made toward applying "big data" to cancer research and precision oncology. As a result, large amounts of data are now available that have the potential to drive influential breakthroughs in cancer drug development. Deep learning methods are especially well-suited for identifying patterns and generating predictions from large amounts of complex data.

Methods

Aggregated data from public sources and supervised learning were used to develop pan-cancer deep neural network models via Keras. CCLE cell line RNA sequencing TPM were downloaded from DepMap. L1000 level 4 compound perturbation data were downloaded from LINCS Data Portal and algorithmically processed to generate transcriptional consensus signatures (TCS) across available cell lines and doses. PharmacoDB experiments were downloaded and filtered to retain CCLE cell lines and L1000 compounds. Prospective validation was performed in vitro for eight prostate cancer (PC) cell lines.

Results

Finalized deep learning models showed >92% accuracy, >95% area under the receiver operating characteristic (ROC) curve (AUROC), and >80% area under the precision-recall curve (AUPR). Predictions were generated for PC cell lines following bulk cell line RNA sequencing. 26 compounds predicted to be effective in PC cell lines were tested in vitro and found to inhibit growth in at least one cell line. Table: 3O

Performance summary for finalized deep neural network models

Conclusions

The study results showed that pan-cancer deep neural network models leveraging gene expression input data are capable of accurately predicting drug sensitivity for individual cancer cell lines. A deep learning model trained on pan-cancer data was able to accurately identify existing drugs with preclinical efficacy in prostate cancer.

Editorial acknowledgement

Clinical trial identification

Legal entity responsible for the study

The authors.

Funding

National Institutes of Health by grants U54HL127624 (NHLBI, LINCS / BD2K), P30CA240139 (NCI, CCSG), R01LM013391 (NLM, Data Curation), and the State of Florida Biomedical Research Program, Bankhead Coley research and research infrastructure grants 9BC13 and 23B16.

Disclosure

All authors have declared no conflicts of interest.