ProteinGym · Substitution

Leaderboard context

Official rankings live on the ProteinGym site. This page summarizes which ProSST checkpoints to use and how they relate to the MSA-enhanced follow-up VenusREM.

Official ProteinGym board

ProSST is evaluated on the ProteinGym substitution (DMS) benchmark for zero-shot fitness prediction. Always treat proteingym.org/benchmarks as the source of truth for Spearman correlations and relative ranking — numbers there are updated by the benchmark maintainers.

Reproduce ProteinGym scoring with the repo script zero_shot/proteingym_benchmark.py and a matching structure-token directory (vocab size must match the checkpoint).

Structure vocabulary sizes

Released Masked LM checkpoints differ only in the structure-token vocabulary size used at quantization and pretraining. Larger vocabs give finer geometric codes; 2048 is the recommended default in the project README.

Checkpoint Structure vocab Hugging Face Notes
ProSST-20 20 AI4Protein/ProSST-20 Coarse structure codes
ProSST-128 128 AI4Protein/ProSST-128
ProSST-512 512 AI4Protein/ProSST-512
ProSST-1024 1024 AI4Protein/ProSST-1024
ProSST-2048 2048 AI4Protein/ProSST-2048 Recommended default
ProSST-4096 4096 AI4Protein/ProSST-4096 Finest released vocab

Relation to VenusREM

VenusREM is an MSA-enhanced follow-up that builds on ProSST’s structure-aware modeling. On ProteinGym’s Substitution track it has been reported at approximately 0.518 Spearman’s ρ in the ProSST README project news (approximate / as stated there — confirm against the live ProteinGym board).

In short: use ProSST when you want a structure-token Masked LM without MSA retrieval; use VenusREM when MSA retrieval is available and you care about top Substitution-benchmark performance. Both are integrated into VenusFactory2.