L2DiffM: LLM-Guided Variational Expert Diffusion for Target-Aware 3D Molecular Generation

Target-aware three-dimensional (3D) molecular generation aims to produce candidate molecules with chemical validity, conformational plausibility, and target compatibility under protein binding pocket constraints. Existing pocket conditioned diffusion models rely primarily on local geometric and atom-level state modeling, leaving high-level medicinal chemistry semantics underutilized; moreover, single-path de noising and fixed-step sampling strategies fail to accommo date the heterogeneity of diffusion states and inter-sample structural diversity, limiting generation quality and structural convergence stability. To address these limitations, we propose L2DiffM, a large language model (LLM)-guided variational expert diffusion framework for target-aware 3D molecular generation. Specifically, L2DiffM establishes a semantic transfer pathway from a language teacher to a protein pocket conditioned structural prior by learning high-level medicinal chemistry patterns in 1D molecular representations, providing pharmacochemical constraints during late-stage denoising. Expert selection is formulated as latent variable inference, yielding a variational self-routed mixture-of-experts (VSRD MoE) denoising mechanism that simultaneously adapts to diffusion-state heterogeneity and inter-sample structural variability via energy- and uncertainty-guided responsibility al location, mitigating the averaging bias of shared denoisers. We further propose Bond-Critical Adaptive Sampling (BCAS), which jointly evaluates bond-topology uncertainty, local geo metric uncertainty, and pocket interaction sensitivity to adap tively reallocate the sampling budget and apply sparse local corrections at high-risk boundary states. Experiments on the PDBbind and CrossDocked benchmarks demonstrate consistent improvements in molecular validity, structural stability, binding affinity, and conformational quality, confirming that L2DiffM provides an effective modeling paradigm for target aware 3D molecular generation.