PharMolixFM

PharMolixFM is an all-atom generative framework for modeling and designing biomolecular structures, released by PharMolix Inc. with the Institute for AI Industry Research (AIR) at Tsinghua University. It decomposes any system — a small molecule, a protein, a peptide, or a complex of them — into atoms carrying 3D coordinates, atom types, and bond types, then treats every structural task as one denoising process run under a different prior. That formulation lets a single trained network dock a ligand into a pocket, invent a new ligand for that pocket, generate conformers, or fill in part of a biomolecular system, without a task-specific model for each.

The technical problem the paper attacks is that atomic data is genuinely multi-modal: each atom has a discrete identity and continuous coordinates, and generative models built for one do not transfer cleanly to the other. Rather than settle on a single answer, PharMolixFM implements three variants of the same architecture — multi-modal diffusion, multi-modal flow matching, and Bayesian flow networks — and compares them under identical training and sampling conditions, which makes the release as much a controlled study of generative objectives for structural biology as a model.

Against AlphaFold 3, the diffusion variant trades a few points of docking accuracy for a roughly fiftyfold reduction in inference time, which changes what is affordable when a campaign requires thousands of poses rather than one. The March 2025 preprint has not been peer reviewed.

#Key Features

• Per-atom noise control: A "Fix" term sets the noise scale independently for each atom and each modality, so holding a protein rigid while generating ligand coordinates, or freezing atom types while sampling positions, is a change of prior rather than a change of model.
• Three generative backends: Diffusion, flow matching, and Bayesian flow network variants share one network and one training recipe, isolating the effect of the generative formulation itself.
• Confidence-scored sampling: Prediction heads emit per-step confidence alongside structure, giving a self-ranking signal for choosing among repeated samples without ground truth.
• Inference scaling: Docking accuracy rises logarithmically with the number of sampling repeats, so additional compute at inference buys measurable accuracy on a hard case.
• Multi-task training mixture: Conformation prediction, pocket-conditioned generation, and partial-system completion are trained jointly, supporting transfer to fragment linking and peptide work.

#Technical Details

The network is a dual-branch SE(3)-equivariant graph neural network following the PocketXMol architecture: a protein branch encodes receptor atoms, a molecule branch captures protein-molecule interaction through those atoms, and independent heads predict coordinates, atom types, bond types, and confidence. Discrete modalities are corrupted with D3PM-style categorical transitions and coordinates with Gaussian noise, with independent schedules per modality. Training ran for 180K steps at batch size 40 on four 80GB A800 GPUs over PDBBind, Binding MOAD, CrossDocked2020, and PepBDB, filtered to molecules of 4-122 heavy atoms and peptides under 15 residues. On the PoseBusters benchmark of 428 complexes with a 10 Å pocket supplied, PharMolixFM-Diff places 83.4% of ligands within 2 Å RMSD under self-ranking at 50 sampling repeats, ahead of PocketXMol at 82.2% on the same budget, rising to 83.9% at 500 repeats and short of AlphaFold 3 at 90.4%. Docking takes about 4.6 seconds per pose against roughly 249 seconds for AlphaFold 3 on a single A800. Oracle ranking of the same samples reaches 98.1%, indicating headroom in the ranking function rather than the generator. On the CrossDocked structure-based design test set, the variants average Vina scores of -6.18 to -6.38 with QED around 0.50 and synthetic accessibility around 0.73.

#Applications

The framework fits virtual screening and hit-to-lead work where pose prediction volume matters: re-docking large ligand libraries against a known pocket, generating candidate molecules directly into a binding site, enumerating conformers, linking fragments, and designing peptides. Its speed advantage makes exhaustive sampling with confidence-based reranking practical for medicinal chemists rather than compute-bound.

#Impact

PharMolixFM's contribution is comparative as much as absolute: by holding architecture and data fixed while swapping the generative objective, it supplies evidence that diffusion outperforms flow matching and Bayesian flow networks on docking, and that the gap narrows on de novo design where data and architecture dominate. Its demonstration of an empirical inference scaling law for molecular generation points at test-time compute as a lever the field had underused. What was released is narrower than what was studied. A single diffusion checkpoint is published on Hugging Face under Apache-2.0, with no flow-matching or Bayesian flow network weights, so the controlled comparison anchoring the paper cannot be rerun from the release alone. The code lives under an unmodified MIT license in PharMolix's OpenBioMed repository, which hosts several unrelated models and labels this one a preview version: its docking and structure-based design entry points are inference-only, with training left as explicit stubs. The checkpoint ships with no model card, so practitioners work from the paper and source code. Evaluation is in-silico only, with no experimental validation.