Algorithmic prediction of response to checkpoint inhibitors: Hyperprogressors versus responders.

Abstract

11565 Background: Predicting response to checkpoint inhibitors (CPIs) using biological knowledge-based decision processes with machine learning (ML) has a great potential to predict rapid progression in patients treated with checkpoint inhibitors (CPIs) (hyperprogressive disease (HPD)) as well as responders. ML models risk overfitting data and do not always evaluate the underlying biology, thus performing well in the initial training cohort but lack generalizability when extended to other cohorts. Biology-based decision may not perform as well initially due to limited understanding and a simplified rule set, but often perform equally well when extended to larger similar cohorts of patients. Methods: A custom NGS cancer immune gene expression assay compared 87 patients treated with CPIs classified as CR, PR, or SD versus 12 HPD. A ML-based polynomial regression model based on 54 immune-related genes combined with mutational burden was optimized for prediction of response. A biological 4-gene decision tree model was constructed independently based on ML. A second biological decision tree incorporated the weighted average relative rank of the expression of multiple genes in 4 different immune functions including immune cell infiltration, regulation, activation, and cytokine signaling. Bayesian model average (BMA) incorporated all three models’ results into the final prediction. Results: For87 patients classified as CR, PR, or SD the PPV >96% for responders and a NPV >90% for non-responders was achieved with the regression model, however with response indeterminate for 24% of the population. While the two biological decision tree models’ PPV were in the 70% range, they accurately revealed the critical genes’ roles in immune response with strong literature support. BMA process integrated these three models resulted in a PPV >96% and a NPV >90% and eliminated the indeterminate group. For HPD a unique biology related to priming of short term memory T-cells was identified. Conclusion: Prediction of response to CPIs is best attained by combining ML with biological knowledge. Decision tree models using a large panel of immune related genes in the context of archival samples from patients treated with CPIs can be used to better understand the biology of responders versus non-responders and provides new insights into HPD.