Biologically informed deep neural network for prostate cancer discovery

Haitham A Elmarakeby,Justin Hwang,Rand Arafeh,Taylor E Arnoff,Jihye Park,Keyan Salari,David Liu,Jett Crowdis,Steven Kregel,Camden Richter,Saud H Aldubayan,Eliezer M Van Allen,Sydney Gang,William C Hahn

doi:10.1038/s41586-021-03922-4

Abstract

The determination of molecular features that mediate clinically aggressive phenotypes in prostate cancer remains a major biological and clinical challenge1,2. Recent advances in interpretability of machine learning models as applied to biomedical problems may enable discovery and prediction in clinical cancer genomics3–5. Here we developed P-NET—a biologically informed deep learning model—to stratify patients with prostate cancer by treatment-resistance state and evaluate molecular drivers of treatment resistance for therapeutic targeting through complete model interpretability. We demonstrate that P-NET can predict cancer state using molecular data with a performance that is superior to other modelling approaches. Moreover, the biological interpretability within P-NET revealed established and novel molecularly altered candidates, such as MDM4 and FGFR1, which were implicated in predicting advanced disease and validated in vitro. Broadly, biologically informed fully interpretable neural networks enable preclinical discovery and clinical prediction in prostate cancer and may have general applicability across cancer types.

Highlights

There is a wide range of potential approaches when developing a predictive model, each comes with trade-offs of accuracy and interpretability
We hypothesized that a biologically informed deep learning model built on advances in sparse deep learning architectures, encoding of biological information and incorporation of explainability algorithms would achieve superior predictive performance compared with established models and reveal novel patterns of treatment resistance in prostate cancer, with translational implications
We developed a deep-learning predictive model that incorporates previous biologically established hierarchical knowledge in a neural network language to predict cancer state in patients with prostate cancer on the basis of their genomic profiles

Summary

Introduction

There is a wide range of potential approaches when developing a predictive model, each comes with trade-offs of accuracy and interpretability. Interpretability of predictive models is critical, as properties that contribute to the predictive capabilities of the model may inform patient care, and provide insights into the underlying biological processes to prompt functional investigation and therapeutic targeting. Linear models such as logistic regression tend to have high interpretability with less accurate predictive performance, whereas deep learning models often have less interpretability but higher predictive performance[13,14]. We hypothesized that a biologically informed deep learning model built on advances in sparse deep learning architectures, encoding of biological information and incorporation of explainability algorithms would achieve superior predictive performance compared with established models and reveal novel patterns of treatment resistance in prostate cancer, with translational implications

Methods

Results

Conclusion