A MODULAR AGENTIC FRAMEWORK FOR SYNTHETICALLY CONSTRAINED MULTI-OBJECTIVE HIT-TO-LEAD OPTIMIZATION

Hit-to-lead optimization in drug discovery requires medicinal chemists to iteratively design and evaluate molecular analogs while balancing competing objectives such as binding affinity, selectivity, physicochemical properties, and synthetic tractability. Exhaustive evaluation of all possible chemical modifications is computationally impossible due to the enormous size of accessible chemical space. To address this challenge, the authors developed SABLE (Synthetically-accessible Agentic Bayesian Ligand Exploration), an open-source framework that combines large language model (LLM) task orchestration with specialized cheminformatics tools. SABLE operates as a computational counterpart to the design-make-test-analyze cycle. Starting from a user-defined seed molecule and target, an LLM interprets the optimization goal and coordinates specialized modules that perform reaction-templated analog enumeration (using HEALER), ADMET and physicochemical property prediction, structure-based affinity scoring (using Boltz-2), and Bayesian optimization.

SABLE uses a stateful workflow implemented in LangGraph to manage campaigns across single and multiple objectives. In single-objective tasks, such as optimizing binding affinity for CAMKK2, Bayesian optimization efficiently samples the reaction-constrained search space to find improved candidates while evaluating only a tiny fraction of the library. For multi-objective campaigns—such as simultaneously optimizing affinity, quantitative estimate of drug-likeness (QED), and CNS activity—the framework employs acquisition functions like qLogNEHVI to map out the Pareto frontier of nondominated solutions. This allows researchers to inspect trade-offs between competing properties and select balanced candidates rather than relying on a single rigid optimum. The modular architecture also allows users to swap out individual prediction backends or tools by editing a simple configuration file without altering the core operational logic.