Large language models for molecular design: bridging the gap between chemical syntax and biological semantics

Large language models (LLMs) are bridging the gap between chemical syntax and biological semantics in early drug discovery. This review traces the evolution from sequence-based generative models to autonomous discovery systems. We analyze diverse paradigms, including de novo hit identification and multi-objective lead optimization. Crucially, we discuss structure-aware and multimodal frameworks that integrate 3D geometric constraints to overcome the limitations of 1D representations. Furthermore, we explore autonomous agents capable of orchestrating closed-loop design-make-test-analyze (DMTA) cycles. Finally, we critically assess the persistent disparity between computational benchmarks and experimental reality, proposing a roadmap toward hybrid neuro-symbolic architectures and unified foundation models for robust, autonomous molecular design.