Novel Machine Learning Model Uncovers Key Prostate Cancer Pathways

Determining molecular and genetic features that define clinically relevant traits of cancers is critical not only for enhancing understanding of the biological underpinnings of disease but also for predicting patient outcomes. Computational approaches can enable detailed and sophisticated analyses of genetic and molecular characteristics of interest; however, each approach comes with shortcomings. Elmarakeby and colleagues considered the limitations of several existing computational approaches used to predict or interpret biological information as it relates to key aspects of prostate cancer, such as treatment resistance. For example, traditional linear regression modeling yields highly interpretable results, but they are generally not as accurate as results from deep learning–based modeling, which itself often comes with the detriment of low interpretability. Use of a biologically informed sparse neural network dubbed pathway-aware multi-layered hierarchical network (P-NET) enabled many of the limitations of existing approaches to discern biologically and clinically relevant information from the molecular profiles of prostate cancer samples to be overcome. Informed by 3,007 human-curated biological pathways, P-NET was trained and tested using data from 1,013 prostate cancers (333 metastatic castration-resistant prostate cancers and 680 nonmetastatic prostate cancers). P-NET outperformed traditional machine learning models along with dense neural network–based models and, using an external validation dataset, correctly classified 73% of nonmetastatic tumors and 80% of metastatic tumors. Interestingly, patients with tumors incorrectly classified by P-NET as treatment resistant had worse outcomes than patients correctly classified into the same group. The biological interpretability of P-NET made it possible to determine which genes and pathways contributed to the computational model's predictions, revealing some expected and some surprising contributors. For example, known prostate cancer driver mutations, such as those in TP53, were detected, but so were mutations in genes not established as strong drivers of prostate cancer, such as those in MDM4, a finding supported by follow-up in vitro studies. In summary, this work demonstrates the utility of P-NET in discovering biological features of both clinical and basic science relevance in cancer.

Elmarakeby HA, Hwang J, Arafeh R, Crowdis J, Gang S, Liu D, et al. Biologically informed deep neural network for prostate cancer discovery. Nature 2021 Sep 22 [Epub ahead of print].

Note: Research Watch is written by Cancer Discovery editorial staff. Readers are encouraged to consult the original articles for full details. For more Research Watch, visit Cancer Discovery online at http://cancerdiscovery.aacrjournals.org/CDNews.