Abstract 1637: Biologically informed deep neural network for genomic discovery and clinical classification in prostate cancer

Abstract Background: Dissection of the molecular properties that distinguish primary and metastatic cancers may reveal new underlying biological drivers of aggressive disease and inform clinical stratification. The rapid increase in the size of molecularly profiled patient cohorts creates an opportunity to develop machine learning algorithms that interrogate these data for discovery and clinical application. However, the superior performance of deep learning models typically comes with the downside of reduced interpretability mainly due to using arbitrary architectures with dense connections. We hypothesized that a biologically informed deep neural network could effectively classify localized or advanced disease (here, in prostate cancer (PrCa)) using molecular features while maintaining interpretability for genomic discovery. Methods: We introduce P-NET, an artificial neural network with biologically informed, parsimonious architecture that accurately predicts metastasis in PrCa patients based on their genomic profiles. In P-NET, each node encodes some biological entity and each edge represents a known relationship between the corresponding entities. P-NET can be used to simultaneously rank features, genes, and biological pathways based on their importance to the clinical classification. We applied P-NET to whole-exome sequencing data from 1012 primary and metastatic prostate cancers and validated our model on two independent validation sets of 130 primary samples and 95 metastatic samples. We compared P-NET performance to other models and visualized ranked features to generate novel biological hypotheses. Results: The trained P-NET outperforms other models including Support Vector Machine, Logistic Regression, and Decision Trees (average area under curve AUC=0.93, the area under precision-recall curve AUPRC= 0.91 for 5-fold cross-validation). In general, copy number variation (CNV) was more informative compared to mutations. Highly-ranked genes in P-NET include AR, PTEN, and TP53, which are known PrCa drivers. In addition, less expected genes such as I-kappa-B kinases and proteasome-related genes are suggested to play a role in predicting the outcome. P-NET selected a hierarchy of 17 pathways (out of 3047 pathways on which P-NET was trained) as significantly relevant to classification including post-translational modification pathways such as ubiquitination and SUMoylation. The model accuracy was 0.77 for the primary validation set and 0.83 for the metastatic set. Conclusion: P-NET, a biologically informed deep neural network, accurately classifies metastatic vs. primary prostate cancers. Visualizing the trained model generates novel hypotheses of mechanisms of metastasis. This represents a novel approach to integrating biology with machine-learning by building mechanistic predictive models, providing a platform for biological discovery. Citation Format: Haitham Elmarakeby, David Liu, Saud H. Aldubayan, Eliezer M. Van Allen. Biologically informed deep neural network for genomic discovery and clinical classification in prostate cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 1637.