Two-level Optimization and Agentic AI Workflow for Molecule Discovery

Autonomous molecular discovery requires both identifying chemically meaningful regions (populations) of a vast design space and efficiently finding optimal candidates within them. We propose a population-specific continuous space embedding, differentiable surrogate model optimization, and post-optimization projection that ensures population membership to efficiently search for optimal candidates within LLM-designed populations. We evaluate this encode-optimize-decode method retrospectively on populations of actinium chelators generated by the Chelatron agentic workflow. The findings show that numerical optimization guides molecule generation towards creating candidates with improved predicted binding energy for a fraction of the computational costs than the original diffusion-based generation done by Chelatron. These results demonstrate a significant improvement in target figures of merit when coupling semantic hypothesis generation with numerical optimization in agentic molecular discovery.