saprot

is a Data Source.

SaProt is a structure-aware protein language model from Westlake University. It uses a structure-aware vocabulary that combines each amino acid with a Foldseek 3Di structure token, and was pre-trained on about 40 million AlphaFold2-predicted protein structures (UniRef50 clusters), with a 650M model further trained on PDB structures and a 1.3B model also trained on metagenomic and NCBI sequences. SaProt supports zero-shot prediction of mutational effects (including missense variant pathogenicity, benchmarked on ClinVar and ProteinGym) and fine-tuning for protein property prediction. The authors distribute the trained models (35M, 650M and 1.3B parameters) and the AF2_UniRef50 pre-training dataset on Hugging Face and the code on GitHub under the MIT license, along with ColabSaprot and the SaprotHub model-sharing platform. The authors do not distribute precomputed variant-effect scores.

License

MIT

Homepage

saprot

Repository

GitHub

External IDs

No external IDs available

Product Summary

Products

From this Resource
ID Name URL Category Format Description
saprot.software SaProt Software SaProt Process python SaProt source code (Python, PyTorch) ...
saprot.models SaProt Models SaProt_650M_AF2 Product mixed Trained SaProt model checkpoint (650M...
saprot.af2-uniref50 SaProt AF2_UniRef50 Pre-training Dataset AF2_UniRef50 Product mixed AF2_UniRef50, the SaProt pre-training...
saprot.saprothub SaprotHub and ColabSaprot SaprotHub Process python SaprotHub and ColabSaprot, a platform...
From other Resources
ID Name URL Category Format Relation Description
biobtree.api BioBTree REST API api Programming Interface http had primary source REST API for searching identifiers an...

Details

SaProt

SaProt is a protein language model developed by Fajie Yuan’s group at Westlake University. Each residue is represented by a structure-aware token that pairs the amino acid with a Foldseek 3Di token describing its local structure. The model was pre-trained on about 40 million AlphaFold2-predicted structures and was presented at ICLR 2024 (spotlight). It can score the effect of amino acid substitutions without task-specific training and can be fine-tuned for protein property prediction.

Access

  • Models: 35M, 650M and 1.3B parameter checkpoints in the westlake-repl Hugging Face organization (MIT).
  • Pre-training data: the AF2_UniRef50 dataset on Hugging Face.
  • Software: the SaProt GitHub repository, with code for zero-shot mutational effect prediction and fine-tuning.
  • SaprotHub / ColabSaprot: tools for training and sharing fine-tuned models, published in Nature Biotechnology (2025).

The authors do not distribute precomputed variant-effect scores.

Is this information incorrect or incomplete? Request an update.

Created: November 12, 2026 | Last modified: November 12, 2026