AI Convergence 2026

Overcoming Data Scarcity and Cognitive Bottlenecks in Lead Optimization: A Physics-Based Generative AI Workflow for the Design of Selective Molecular Glues

Speaker:
Sathesh Bhat, Executive Director, Therapeutics Group, Schrödinger

Abstract:
Generative AI in lead optimization is frequently bottlenecked by two seemingly contradictory limitations: the experimental data scarcity required for multi-parameter optimization (MPO) of a novel chemical series, and the “cognitive bottleneck” preventing human chemists from processing the complex, non-linear structure-activity relationships (SAR) inherent in ultra-large chemical spaces. We present a workflow that overcomes both barriers by synergizing de novo design (AutoDesigner), physics-based simulations (WaterMap/FEP+) and reinforcement learning based generative AI. To address the stochastic variance and mode collapse typical of single-agent reinforcement learning in sparse reward landscapes, we introduced two methodological enhancements to the standard framework: a parallelized ensemble strategy to maximize trajectory diversity, and a dynamic experience replay buffer with periodic flushing to force iterative exploration. We applied this unified workflow to the design of potent and selective p38α/MK2 molecular glues, starting from a single reference compound with no established selectivity SAR. AutoDesigner first enumerated over 1 million idea molecules which served as high quality, human-mimetic project specific SAR. The generative AI was trained on the project specific AutoDesigner prior along with high-fidelity target specific datasets of FEP+ binding affinities and WaterMap scores. By leveraging the AutoDesigner prior, the generative AI functioned as an augmented medicinal chemistry design workflow to autonomously navigate the complex interplay between ligand structure, ligand binding affinity and site-specific hydration thermodynamics for selectivity. The model identified a novel, high-affinity design space comprising 10,500 candidates that met all the project criteria. None of these ideas were found in the initial Autodesigner enumeration. Furthermore, the workflow demonstrated a 57-fold increase in the number of generated candidates meeting all project criteria compared to traditional funnel-based filtering. Prospective experimental evaluation of three generative AI molecules yielded p38-MK2 inhibitory potencies (pIC50​) of 10.3, 8.8, and 7.9. These compounds also exhibited excellent experimental selectivity over p38α (451-, 255-, and 239-fold, respectively), significantly surpassing the selectivity profile of the starting reference lead. These results demonstrate that a generative AI framework driven by accurate physics-based scoring and domain-specific priors, can autonomously address complex interface specificity requirements and optimize selectivity in data-poor regimes. Finally, the observed steerability of the generative prior toward diverse physicochemical descriptors suggests the broader utility of this platform for multi-parameter lead optimization across varied therapeutic programs.