Scalable Generative Modeling of Protein Ligand Trajectories via Graph Neural Diffusion Networks

ICLR 2026 Conference Submission20390 Authors

Keywords: Protein–ligand dynamics, Drug discovery, Graph Neural Networks, Cross-attention, Diffusion models, SE(3)-invariance, Generative modeling, Molecular simulation, Trajectory prediction, Transition path sampling

TL;DR: A scalable generative framework combining protein language models and diffusion models to simulate protein-ligand dynamics in large biomolecular systems.

Abstract: Modeling protein–ligand dynamics over long timescales is essential for drug discovery yet remains challenging in large biomolecular systems. We propose HemePLM-Diffuse, a generative framework that combines Graph Neural Networks, cross-attention, and diffusion models to simulate protein–ligand interactions with atomic-level fidelity. The method employs SE(3)-invariant graph representations to preserve molecular geometry and a time-aware cross-attention mechanism to capture context-dependent interactions between proteins and ligands. A diffusion-based generative process models stochastic motion, enabling trajectory forecasting and ligand fragment inpainting.HemePLM-Diffuse scales efficiently to systems exceeding 10,000 atoms while maintaining structural accuracy. On the 3CQV HEME system, it surpasses leading methods, including TorchMD-Net, MDGen, and Uni-Mol, in trajectory prediction and transition path sampling . By integrating geometry-aware graph learning with generative diffusion, HemePLM-Diffuse provides a scalable alternative to molecular dynamics, advancing data-driven approaches for drug design and protein function analysis.

Primary Area: generative models

Submission Number: 20390