eLife 2025 · Sequence–structure · Zero-shot

ProtSSN

Semantical and geometrical protein encoding toward enhanced bioactivity and thermostability. Fuse sequence semantics with geometric structure—no MSA required—for zero-shot mutant scoring and directed evolution.

Yang Tan, Bingxin Zhou, Lirong Zheng, Guisheng Fan, Liang Hong

Paper Code Models Leaderboard ProteinGym

ProtSSN overview: sequence–structure fusion for zero-shot protein engineering — ProtSSN couples an ESM2 sequence encoder with an EGNN structural adapter for end-to-end zero-shot fitness prediction.

Method

ProtSSN treats structure-aware modeling as an adapter on a pretrained protein language model. Residue embeddings from ESM2 are refined by an EGNN that operates on a local Cα graph (neighborhood size k ∈ {10, 20, 30}), so geometry and sequence semantics are fused in one forward pass.

Pretraining follows a denoising objective on CATH domains: residues are corrupted and the network learns to recover amino-acid identity conditioned on both sequence context and 3D neighbors. At inference, mutant scores come from log-likelihood ratios—no task-specific fine-tuning and no MSA.

Nine checkpoints span k10/20/30 × h512/768/1280. For a single model, prefer k20_h512; ensembling all nine further improves ProteinGym correlation at modest cost.

Zero-shot usage

Inputs are a wild-type PDB (experimental or predicted) and a mutant table. Fold with AlphaFold / ESMFold or pull structures from the AlphaFold Database when wet-lab coordinates are unavailable.

Download checkpoints from Hugging Face (e.g. protssn_k20_h512.pt).
Run zeroshot_predict.py for single-model or --use_ensemble scoring.
For proteins without assay data, build single-site saturation libraries (src/build_sav.py) and rank variants for directed evolution.
Fine-tuning and embedding extraction scripts are included for downstream labels and representation analysis.

Beyond ProteinGym DMS assays, ProtSSN was evaluated on constructed DTm and DDG stability sets, where structural context is especially informative.

When to use ProtSSN vs ProSST / VenusREM

ProtSSN — Best when you have (or can fold) a structure, want a compact EGNN adapter on ESM2, and prefer an MSA-free pipeline for zero-shot fitness or early directed-evolution screens.
ProSST — A full sequence–structure Transformer with quantized structure tokens and disentangled attention; use when you want a dedicated structure-conditioned PLM rather than an EGNN adapter.
VenusREM — Prefer when deep MSAs are available and you need top ProteinGym Substitution accuracy via retrieval-enhanced scoring; VenusREM later leads that board relative to ProtSSN’s 2023 ensemble result.

Applications

ProtSSN is used in wet-lab redesign and community evaluation stacks:

Zheng et al., Nature 2026 (CytoTape) — guided single-site mutagenesis during monomer redesign.
Zhang et al., Acta Pharm. Sin. B 2025 (VenusMutHub) — structure-aware zero-shot mutant scoring across hundreds of small-scale assays.
VenusFactory2 / web server — zero-shot mutation scoring in an open engineering toolkit.

See the leaderboard page for model variants, ProteinGym context, and the applications table.

Citation

Tan et al. Semantical and geometrical protein encoding toward enhanced bioactivity and thermostability. eLife 2025;13:RP98033. doi:10.7554/eLife.98033.

@article{tan2025protssn,
  article_type = {journal},
  title = {Semantical and geometrical protein encoding toward enhanced bioactivity and thermostability},
  author = {Tan, Yang and Zhou, Bingxin and Zheng, Lirong and Fan, Guisheng and Hong, Liang},
  editor = {Koo, Peter and Cui, Qiang},
  volume = 13,
  year = 2025,
  month = {may},
  pub_date = {2025-05-02},
  pages = {RP98033},
  citation = {eLife 2025;13:RP98033},
  doi = {10.7554/eLife.98033},
  url = {https://doi.org/10.7554/eLife.98033},
  journal = {eLife},
  issn = {2050-084X},
  publisher = {eLife Sciences Publications, Ltd},
}