ISMB/ECCB 2025 · Bioinformatics

VenusREM

Retrieval-enhanced fitness prediction that bridges high-throughput mutation evaluation and wet-lab directed evolution — combining structure/sequence protein language model signals with explicit MSA retrieval.

Yang Tan, Ruilin Wang, Banghao Wu, Liang Hong, Bingxin Zhou

Paper Code Dataset ProteinGym Leaderboard
VenusREM overview figure
VenusREM scores mutants by fusing PLM logits with retrieved MSA context for zero-shot fitness prediction.

What the model does

VenusREM performs zero-shot mutant scoring: given a wild-type sequence, structure, and optional substitution list, it estimates relative fitness without assay-specific fine-tuning. Structure and sequence protein language model (PLM) signals are combined with MSA retrieval — homology alignments prepared as EVCouplings a2m or ColabFold a3m files, released with the VenusREM dataset on Hugging Face.

The same inference path supports ProteinGym-scale evaluation and single-protein wet-lab campaigns: prepare sequences, structures, structure tokens, and alignments, then run compute_fitness.py.

Why retrieval helps

VenusREM sits in a family of zero-shot predictors from the same line of work, each emphasizing a different structural or evolutionary cue:

  • ProtSSN — geometrical encoding over amino-acid coordinates (no MSA required).
  • ProSST — local structure tokens with disentangled sequence–structure attention.
  • VenusREM — explicit MSA retrieval layered on structure-aware PLM scoring.

In practice the three are complementary: retrieval helps when deep, informative alignments are available; coordinate- or token-level structure models can be preferable when MSAs are shallow or costly to build. Setting --alpha 0 in VenusREM’s fitness script recovers a ProSST-style score without MSA reweighting, easing head-to-head comparison.

Results highlight

Ranked 1st on the ProteinGym Substitution leaderboard (April 2025), with strong standing also reported on VenusMutHub and related mutation-effect suites. See the leaderboard page for a compact family comparison; official live rankings remain on ProteinGym.

Wet-lab oriented workflow

For a custom protein, the README pipeline mirrors a typical directed-evolution prep:

  1. Homology search — JackHMMER / EVCouplings against UniRef100; select an a2m (or supply ColabFold a3m).
  2. Structure — high-quality PDB from AlphaFold DB/Server, ESMFold, or experiment.
  3. Structure sequence — get_struc_seq.py over the PDB directory.
  4. Mutant table — assay CSV or all single-site variants via get_sav.py.
  5. Score — compute_fitness.py --base_dir … --out_scores_dir ….

Preprocessed ProteinGym v1 bundles (alignments, PDBs, structure sequences, substitutions) are available from the VenusREM dataset.

Integration

VenusREM is integrated into VenusFactory2 (venusfactory.bio); open-source stack on GitHub: ai4protein/VenusFactory2.

Citation

If you use VenusREM, please cite:

@article{tan2025venusrem,
    author = {Tan, Yang and Wang, Ruilin and Wu, Banghao and Hong, Liang and Zhou, Bingxin},
    title = {From high-throughput evaluation to wet-lab studies: advancing mutation effect prediction with a retrieval-enhanced model},
    journal = {Bioinformatics},
    volume = {41},
    number = {Supplement_1},
    pages = {i401-i409},
    year = {2025},
    doi = {10.1093/bioinformatics/btaf189}
}