Preprint about Benchmarking Study on Hormone-related Toxicity Prediciton

We are happy to announce that the preprint of our paper “A Comprehensive Evaluation of Machine Learning Pipelines for Toxic Endpoint Prediction” is finally live on Research Square: https://www.researchsquare.com/article/rs-10470980/v1. The paper presents the first comprehensive benchmark of ML models, molecular representations, and DR methods for hormone receptor activity assays in ToxCast. In contrast to prior work, we systematically evaluate complete modelling pipelines with rigorous hyperparameter tuning and chemically-informed data splitting, supported by extensive statistical analysis. Our main finding is that simple approaches often outperform more complex alternatives. In fact, the combination of random forest with physicochemical properties and no dimension reduction emerged as the strongest and most consistent performer. Notably, it also outperforms deep neural networks and learned compound representations. Moreover, our analysis revealed that activity cliffs remain one of the major challenges for machine learning-based toxicity prediction.