On the Efficiency and Transferability of Inductive Mondrian Conformal Predictors for Drug-Drug Synergy

We propose a principled approach to quantifying prediction uncertainty in machine learning for drug-drug synergy, a burgeoning subfield within drug discovery where human decision-makers require a clear understanding of the errors associated with predictions. To address the limitations of traditional point prediction models typically outputting a single value (for regression settings) or a single label (for classification settings) without any measure of uncertainty, we introduce Mondrian inductive conformal prediction for drug-drug synergy with probabilistic guarantees on the accuracy of each prediction. By providing statistically valid prediction regions at predefined confidence levels, inductive Mondrian conformal predictors enhance the interpretability and reliability of computational drug-drug synergy models, with observed unconfidence and fuzziness scores of $0.13 \pm 0.02$.