Surrogate Machine Learning Models for ADME Prediction: From Synthetic Labels to Experimental Data

Optimization of absorption, distribution, metabolism, and excretion (ADME) properties is central to small-molecule drug discovery, yet is often constrained by limited access to large and consistent experimental data sets. The performance and generalizability of machine learning (ML) models for ADME property prediction rely on the composition of the training data. Surrogate data sets comprise compounds annotated with synthetic ADME labels (predictions) generated by ML models trained on experimental data. This can offer a consistent data augmentation or sharing strategy without requiring access to proprietary data sets; ML models trained on these data can be referred to as surrogate models. However, systematic strategies for developing, evaluating, and applying surrogate models remain underexplored, particularly in the context of transfer learning and validation on new experimental data. In this work, surrogate modeling strategies were investigated for ADME prediction using a data set of over 270,000 compounds with synthetic labels across 25 endpoints. ML models were built on such surrogate data set and evaluated on prospective in-house and public experimental ADME data. Surrogate models based on multitask graph neural networks were predictive under prospective in-house validation and across multiple properties on external data. Uncertainty estimates were informative and actionable, identifying predictions associated with higher or lower prediction errors. Moreover, fine-tuning on experimental data further improved performance for most property data sets. Overall, these results show that surrogate models, especially coupled with fine-tuning, provide a practically valuable approach for ADME prediction, supporting model development from consistently generated data and bringing new opportunities for ML models for property prediction.