Experiments with data-augmented modeling of ADME and potency endpoints in the ASAP-Polaris-OpenADMET Antiviral Challenge

We participated in the ASAP-Polaris-OpenADMET Antiviral Challenge for predictive modeling of Potency and ADME endpoints. We augmented each challenge with domain-specific publicly available datasets and used cheminformatics-based feature selection and engineering to build tree-based models, graph-based neural networks and transformer neural networks. For modeling ADME endpoints such as KSOL, LogD, HLM, MLM, and MDR1-MDCKII permeability, we found that XGBoost models trained on a feature space of 55 descriptors outperformed other tree-based models and GNNs on four out of the five endpoints, based on blind evaluation with the ADME testset. We also pretrained a fusion AttFP-GNN and transformer neural network potency model on a curated ChEMBL IC50 dataset of protein-ligand pairs, and fine-tuned it with the Potency train dataset. A fine-tuning variant, named gradual partial fine-tuning model, obtained an aggregate MAE ~0.79 on blind evaluation with the Potency testset. Our work serves as a set of experiments in the context of the ASAP Challenge, whose results yield insights into the efficacy of specific feature spaces and models for ADME endpoints and potency, and suggest scope for improvements.