Prediction of chemo-response for serous ovarian cancer using DNA methylation patterns with deep machine learning (AI)

Objectives: Patients with high grade serous ovarian cancer (HGSC) that do not respond to initial chemotherapy have a poor prognosis. DNA methylation has been implicated in epigenetic regulation of chemotherapy response. We hypothesize that processing genomic data from HGSC patients with deep learning algorithms will improve the prediction performance of chemo-response. Our aims are to build a prediction model of chemo-response using deep machine learning (AI) and DNA methylation data, and to compare the performance of this model with models built with more traditional methods of statistical learning. Methods: Chemotherapy response was defined as progression-free survival (PFS) greater than 6 months (N=45). Non-response to chemotherapy was defined as stable or progressive disease during initial therapy, or recurrence within 6 months after completion of therapy (N=36). DNA was extracted from 81 patients with HGSC at the University of Iowa and methylation status was performed with IlluminaEPIC arrays. A prediction model of chemo-response using MetylNet, open source, deep learning (AI) network was created. This model was compared to a model built with lasso, a statistical learning regression method that includes k-fold cross validation. The performance on both models were measured with area under the curve (AUC) and their 95% confidence intervals (CI). Results: The AUC of the AI prediction model was 62.5% with a 95% CI of 33.9%, 91.1%. With a sensitivity of 90%, the calculated negative predictive value was 85.7%. The performance of the lasso model, measured in AUC, was 80% with a 95% CI of 72%, 89%. CIs of both AUCs’ models overlapped. Conclusions: Performance of neither model was superior, however the lasso method had more precise 95% confidence intervals. This pilot study shows that is possible to build prediction models of chemo-response with DNA methylation patterns using AI algorithms. To improve the performance of AI models we will need to increase sample size and optimize hyperparameter selection. The best model would ultimately require an independent data set for validation.