Predicting super-class drug mechanisms from large-scale IC50 cell-line sensitivity profiles

Abstract

Predicting mechanisms of action of drug compounds at the super-class level is indispensable for early drug discovery and functional assessment of compounds. The aim of this study is to perform Super-Class-level mechanism-of-action prediction from phenotypic IC50 cell-line sensitivity data under strict leakage control. In this work, we present a leakage-safe machine learning pipeline for Super-Class Mechanism-of-Action (Super-Class MoA) prediction using drug-cell line sensitivity profiles prepared at large scale from the GDSC dataset. The standardized features derived from IC50 were encoded into 32-dimensional latent embeddings using an autoencoder, while SMOTE was applied solely to the training embeddings of each outer fold to address class imbalance. The resulting embeddings were used to train a stacked neural-boosted ensemble consisting of a neural network and XGBoost, with validation embeddings kept completely untouched to ensure unbiased evaluation. The framework achieved a mean accuracy of 0.9182, a mean macro-F1 score of 0.9118, and a mean macro-AUROC of 0.9568 across five outer folds, demonstrating consistent performance across the three Super-Class MoA categories. These results indicate that latent representations derived from phenotypic drug-response profiles capture biologically meaningful variation, supporting robust Super-Class MoA prediction in early-stage screening settings.