Neural Network-Driven Lead Optimization with Embedded ADMET Compliance.
Abstract
The predominant failure mode in small-molecule drug development is not insufficient potency — it is unanticipated toxicity and poor pharmacokinetics discovered late in the development cycle. Conventional practice treats ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) evaluation as a sequential gate applied after molecular design, creating costly iterative loops that consume resources without guaranteed convergence toward a viable candidate.
This white paper presents a deep learning framework for lead optimization that encodes multi-endpoint ADMET compliance directly into the molecular generation objective. The architecture couples a dual-mode generative neural network — a sequence-based recurrent model for exploration and a graph neural network for exploitation — with an ensemble of task-specific ADMET scoring models. Reward signals derived from predicted ADMET scores across five endpoints are aggregated into a differentiable weighted sum and used to update generative model parameters via policy gradient reinforcement learning. Generated candidates are iteratively scored in silico, with high-scoring compounds advanced to confirmatory wet laboratory assays.
Across three internal benchmark programs, ADMET-constrained generation achieved a 74% reduction in predicted hepatotoxicity incidence, a 61% reduction in cardiac ion channel liability (hERG IC50 < 1 µM), and a 2.1-fold improvement in mean predicted metabolic clearance (CLint) relative to unconstrained baselines. Prospective wet laboratory confirmation of 48 compounds demonstrated 81% concordance with in silico predictions across all three endpoints, establishing empirical validity for deployment in active lead optimization pipelines.
Download '.pdf' Format of the whitepaper.
