bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelTermsPrivacyContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / fampnn
Protein
Stanford UniversityArc InstituteReleased February 2025

FAMPNN

Fixed-backbone protein sequence design that co-generates amino acid identity and sidechain conformation, with 49.7% sequence recovery on CATH 4.2.

The short version

  • —Designs a sequence and packs its sidechains together instead of in two separate stages
  • —Scores mutations for stability and binding affinity without training on fitness data
  • —Per-atom predicted sidechain error flags which rotamers in a design are trustworthy
  • —Ten iterative unmasking steps replace one decoding step per residue
68Openness

Where to run it

No providers recorded yet. Browse all providers

FAMPNN (Full-Atom MPNN) is a fixed-backbone protein sequence design method that models what its predecessors left implicit: where the sidechain atoms actually go. Given a protein backbone, established inverse folding models predict amino acid identities from backbone geometry alone, reasoning about van der Waals contacts, hydrogen bonds and pi-stacking only indirectly. FAMPNN instead learns a joint per-residue distribution over the discrete amino acid label and the continuous sidechain conformation, training a single network end to end with a combined categorical cross-entropy and diffusion loss.

The model comes from Talal Widatalla, Richard Shuai, Po-Ssu Huang and Brian Hie at Stanford University and the Arc Institute, posted as a preprint in February 2025 and presented at ICML 2025. It builds directly on the encoder design of ProteinMPNN, extending it with geometric vector perceptron layers and a full-atom graph, and it is positioned against LigandMPNN, which can accept sidechain atoms as context but still packs them with a separate model after the sequence is fixed.

Making sequence and structure a single generative problem turns out to be synergistic rather than merely convenient: joint training improves sequence recovery and simultaneously delivers strong sidechain packing, and the full-atom context transfers to zero-shot prediction of experimental stability and binding measurements.

#Key Features

  • Joint sequence and sidechain generation: A dual masked-language-modeling scheme co-generates amino acid identity and sidechain coordinates, so each design decision is conditioned on the three-dimensional arrangement of its neighbours.
  • Cartesian diffusion for sidechains: Coordinates are denoised directly in atom37 representation rather than through torsion angles, avoiding the lever-arm effect where small chi-angle errors produce large positional deviations.
  • Predicted sidechain error as a confidence metric: A confidence module estimates per-atom packing error in Ångströms, correlating with true error at Spearman ρ = 0.843 per atom and 0.780 per residue on CASP15 crystal structures.
  • Iterative parallel decoding: Sampling starts fully masked and unmasks subsets of positions each step, so ten steps suffice for high self-consistency instead of one step per residue.
  • Flexible partial conditioning: Sequence and sidechain masking are independent, allowing packing given a known sequence, design given partial sidechain context, or any mixture — useful at protein-protein interfaces.

#Technical Details

The full-atom encoder is a hybrid MPNN-GVP graph neural network over a k-nearest-neighbour residue graph with 48 neighbours, comprising an invariant backbone encoder, an invariant full-atom encoder and an equivariant full-atom encoder. Residues use the atom37 format, with absent atom types collapsed to "ghost atoms" at the Cα position so that masked sidechains cannot leak amino acid identity. Two checkpoints are released, trained with 0.0 Å and 0.3 Å of Gaussian coordinate noise, alongside a CATH-only variant; PDB training used the AlphaFold3 reproduced splits with a 2021-09-30 cutoff, chain-level clustering at 40% homology and interface-based clustering, over 300k steps on four H100 GPUs. On the CATH 4.2 test set FAMPNN reaches 49.66% median sequence recovery at 4.99 perplexity, above ProteinMPNN's 45.96% and the GVP module's 39.47%. Sidechain packing on CASP15 gives 0.690 Å all-atom RMSD for the 0.0 Å model and 0.785 Å for the 0.3 Å model, ahead of LigandMPNN (0.788 Å) and FlowPacker (0.765 Å) on core residues. On self-consistency over RFdiffusion backbones from length 100 to 500, the 0.3 Å model tracks ProteinMPNN closely. Zero-shot, FAMPNN outperforms unsupervised baselines on SKEMPIv2 binding ΔΔG and improves on ProteinMPNN and ESM-IF across S669, Megascale and FireProtDB stability datasets. Inference costs 0.03 s for a single step and 0.11 s for five steps per sample on one H100.

#Applications

FAMPNN slots into the same position as any inverse folding model — sequencing backbones from generative design tools such as RFdiffusion — but its full-atom awareness matters most where sidechain packing determines success: redesigning protein-protein interfaces, stabilizing therapeutic proteins, and prioritizing point mutations for expression or affinity. Because it scores mutations by conditional likelihood with the mutated position and its sidechain masked, it can rank variant libraries without any fitness training data, which is directly useful in antibody affinity maturation. The predicted sidechain error gives designers an atom-level confidence signal for deciding which parts of a model to trust before committing to synthesis.

#Impact

FAMPNN establishes that explicit sidechain modelling is not a separate downstream task but a training signal that improves sequence design itself — an argument the ablations make directly, with both full-atom conditioning and the packing objective independently raising sequence accuracy. The code and all pretrained checkpoints are released under the MIT license, lowering the barrier to adoption in existing design pipelines. Limitations are honest ones: sequence recovery, while strong, trails specialized methods such as PiFold; the antibody-antigen zero-shot benchmark did not improve with full-atom conditioning, which the authors attribute to inaccurate input structures; and self-consistency on long de novo backbones remains comparable to rather than better than ProteinMPNN.

At a glance

Released
February 2025
Category
Protein
Organizations
Stanford University / Arc Institute

Links

GitHub RepositoryResearch PaperbioRxiv Preprint

Tags

diffusiongraph_neural_networksequence_designsidechain_packingzero_shot

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.