HemePLM-Diffuse

HemePLM-Diffuse is a generative transformer that simulates the motion of protein-ligand complexes instead of integrating equations of motion. Given a starting structure, it samples trajectories directly, fills in missing ligand atoms, interpolates between sparsely sampled frames, and generates transition paths between bound and unbound states. It was developed by Rakesh Thakur and Riya Gupta at the Amity Centre for Artificial Intelligence, Amity University, Noida, and released as a preprint in August 2025.

Classical molecular dynamics resolves these processes accurately but becomes prohibitively expensive for large systems and for rare events such as ligand dissociation, which unfold over microseconds to milliseconds. Learned alternatives have each addressed part of the problem: neural force fields such as TorchMD-Net still require numerical integration, MDGen generates backbone-level protein trajectories without ligands, and Uni-Mol pretrains on static 3D molecular structures with no temporal component. HemePLM-Diffuse targets the gap those leave — all-atom dynamics of a protein together with its bound ligand.

The model complements ensemble generators such as BioEmu-1, which samples equilibrium conformational ensembles of proteins alone. HemePLM-Diffuse instead models the time-ordered trajectory of a complex, with the ligand treated as a first-class part of the system rather than as context.

#Key Features

• SE(3)-invariant tokenization: Proteins are tokenized as Cα coordinates, backbone torsion angles, and auxiliary features including B-factor, solvent-accessible surface area, and binding-site scores; ligands as RDKit fragments with atom-wise descriptors, internal dihedrals, and top-k residue-ligand distances.
• Time-aware cross-attention: The diffusion backbone interleaves intra-molecule self-attention with ligand-protein cross-attention, conditioned on a sinusoidal time embedding, so atomic motion is modeled jointly across both partners.
• Ligand inpainting: Masked ligand atoms are recovered from the remaining fragments and the surrounding protein context, a task neither force-field nor trajectory-only models support.
• Transition path sampling: Conditioning on start and end conformations produces intermediate frames that preserve contact maps, ligand orientation, and energetic plausibility.
• Trajectory upsampling: Sparse molecular dynamics sequences can be interpolated into finer-grained trajectories, useful where frequent snapshots are impractical.

#Technical Details

The architecture is a diffusion transformer trained with a velocity-based denoising objective in the flow-matching style, using a multi-objective loss that combines a trajectory MSE term, a free-energy-surface term expressed as a KL divergence against reference molecular dynamics, a cross-entropy inpainting term on masked ligand fragments, and a transition-path term. Training data were curated from the Protein Data Bank, selecting systems of at least 10,000 atoms with bound ligands of variable topology and resolution of 2.5 Å or better. Reference dynamics were produced with GROMACS using the Amber99SB-ILDN force field.

Reported evaluation is on a single real-world system: PDB 3CQV, the Rev-erbβ nuclear receptor bound to heme, run on one NVIDIA A100. Masking 10 heme atoms, the model recovers their coordinates at 0.91 Å RMSD, against 1.67 Å for MDGen and 1.82 Å for TorchMD-Net on simpler systems. Generated trajectories deviate from the reference molecular dynamics by 1.03 Å per frame on average, compared with 1.90 Å for MDGen and 2.85 Å for Uni-Mol, and a transition-path quality score of 0.95 exceeds MDGen's 0.75 and Uni-Mol's 0.40. One nanosecond of equivalent dynamics takes about 12 minutes, versus roughly 4 hours for TorchMD-Net and about 20 hours for GROMACS on a single GPU.

#Applications

The intended uses are structure-based drug design and mechanistic structural biology: watching a ligand dock into a pocket, estimating how a bound pose relaxes, completing partially resolved ligand density, and enriching sparse simulation archives without rerunning the underlying dynamics. A speedup of this magnitude over explicit integration would make trajectory-level reasoning tractable inside screening and lead-optimization loops rather than reserved for a handful of targets.

#Impact

The work stakes out a task that existing generative dynamics models leave uncovered — all-atom protein-ligand trajectories, including the ligand's internal degrees of freedom — and pairs it with concrete auxiliary capabilities in inpainting and transition path sampling. Its evidence base is narrow: results come from one system, the heme-bound Rev-erbβ complex, and the authors describe fine-tuning on multiple diverse systems such as membrane proteins as ongoing work. Solvent molecules are not explicitly simulated, and integrating density-functional-theory-based active learning to improve force accuracy is listed as future development. The preprint has not been peer reviewed, and it links no code repository or trained weights.