Abstract C006: DrugCell: A visible neural network to guide precision medicine

Abstract

Abstract Many potential cancer therapies fail during clinical trials, in part due to difficulty in predicting how a particular population of cancer cells will respond to a given drug. Here we develop DrugCell, a “visible” neural network that predicts anti-cancer drug responses by modeling the hierarchical organization of a human cancer cell. We devised a branched visible neural network combining two heterogeneous neural networks. The structure of the first neural network mirrors the hierarchical organization of cellular subsystems (e.g., complexes, pathways and organelles). During training, this neural network learns which cellular subsystems contribute to therapeutic responses based on the mutational profile of the genome. The second neural network is fully connected and models the chemical structure of each drug. These two neural networks connect to a series of small, fully connected layers that simulate a cellular drug treatment conditioned on a genetic mutation landscape. By this design, DrugCell can be used to predict how any cancer cell would respond to any small molecule inhibitor. DrugCell outperformed Elastic Net (rho = 0.80 versus 0.74) while maintaining performance similar to that of a fully connected (black box) neural network (rho = 0.82). When making predictions using DrugCell, genotypes and chemical compound structures induce differential patterns of subsystem activity, enabling in silico investigations of the molecular mechanisms underlying cancer drug response. We therefore applied DrugCell to highlight pathways mediating sensitivity to multiple FDA-approved therapies. Top pathways regulating paclitaxel sensitivity included G2 DNA damage checkpoint, consistent with previous studies showing BRCA1 mediates G2/M arrest in response to paclitaxel, and regulation of TOR signaling, which has previously been described to mediate paclitaxel resistance in multiple cancers. Accordingly, combination of paclitaxel with the mTOR inhibitor, MK8669, was highly synergistic across multiple cell lines. Similarly, pathways mediating sensitivity to topoisomerase II inhibitors included cholinergic receptor signaling and regulation of histone deacetylation. In support of these mechanisms, combination of etoposide, a topoisomerase II inhibitor, with AKT or MEK inhibitors (downstream of the cholinergic receptor) or the HDAC inhibitor vorinostat were highly synergistic across multiple cell lines. Similar results were observed upon CRISPR/Cas9 mediated knockdown of key genes in these pathways. Finally, we used DrugCell to predict effective therapeutic combinations for acute myeloid leukemia patients, leading to identification of drug combinations currently in clinical trials including lenalidomide + DNMT inhibitors and tandutinib + PI3K/MTOR inhibitors. Armed with interpretability and generalizability, DrugCell serves as an important step towards a next generation of intelligent systems in drug discovery and precision medicine. Citation Format: Brent M Kuenzi, Jisoo Park, Samson Fong, Jason Kreisberg, Trey Ideker. DrugCell: A visible neural network to guide precision medicine [abstract]. In: Proceedings of the AACR-NCI-EORTC International Conference on Molecular Targets and Cancer Therapeutics; 2019 Oct 26-30; Boston, MA. Philadelphia (PA): AACR; Mol Cancer Ther 2019;18(12 Suppl):Abstract nr C006. doi:10.1158/1535-7163.TARG-19-C006