Multitask Deep Learning for Solubility, Permeability, Protein Binding, and Clearance Prediction Download PDF

Solubility, permeability, protein binding, and clearance together define the ADME profile of a drug candidate. These endpoints are often predicted by separate models, even though they depend on overlapping chemical determinants. Single-task ADME models do not fully exploit the information contained in correlated endpoints. They can also produce fragmented molecular profiles that are difficult to reconcile during lead optimization. This article describes a multitask deep neural network that learns a common molecular representation and simultaneously predicts aqueous solubility, membrane permeability, plasma protein binding, and metabolic clearance. Each output is paired with a task-specific uncertainty estimate to support model-informed decision-making. A molecular graph encoder, such as a graph attention network or graph isomorphism network, produces a shared fixed-length. embedding for each compound. Four task-specific prediction heads map this representation to solubility, permeability, fraction unbound, and clearance outputs using heteroscedastic uncertainty-weighted learning.
Conceptually, the multitask model would be expected to improve information sharing across related ADME endpoints relative to equivalent single-task models. It could provide a coherent ADME profile from one molecular input without requiring separate model calls for each endpoint. A multitask approach to ADME prediction could simplify early pharmacokinetic screening by unifying several core developability endpoints in a single model. Such a framework would be especially useful for multi-parameter optimization, uncertainty-aware compound triage, and prospective medicinal chemistry design.