Protein Design · Su et al., Westlake University
SaProt
A structure-aware protein language model predicting variant effects and generating embeddings using 3Di tokens.
What SaProt is
SaProt is a structure-aware protein language model developed by Su et al. at Westlake University. Unlike standard protein language models that operate exclusively on primary amino acid sequences, SaProt integrates tertiary structural information directly into its token stream by pairing sequence residues with Foldseek 3Di structural alphabet tokens.
By capturing explicit backbone geometry alongside sequence information, the model achieves higher performance than ESM2 across multiple downstream protein function prediction tasks when structural coordinates are available.
The model can predict the functional consequences of single or multi-residue mutations, evaluate amino acid probabilities across sequence positions, and generate whole-protein vector representations. When using structural models from AlphaFold 2, it accommodates predicted disorder or uncertainty by applying structural confidence masks.
How it works
SaProt tokenises proteins using combined sequence and structure tokens (SA tokens). Each token links an amino acid residue to its local 3D structural state derived from Foldseek's 3Di discrete alphabet. For structures generated by AlphaFold 2, residues situated in regions with a predicted local distance difference test (pLDDT) score below 70 are assigned a special mask token to account for low structural confidence.
The underlying architecture is built upon the ESM masked language model framework (EsmForMaskedLM) and contains 650 million parameters. In addition to outputting token-level logits for masked language modelling and variant scoring, the model computes protein representations by averaging the hidden states across the final layer.
Key facts
- Model architecture
- ESM masked language model (EsmForMaskedLM)
- Parameter count
- 650 million (SaProt_650M_AF2)
- Modalities
- Protein sequence and structure (Foldseek 3Di)
- Hardware requirement
- 16 GB GPU memory
- Version
- 1.0
What you can use it for
- Single-point mutational effect scoring
- Evaluating the functional impact of individual amino acid substitutions relative to the wild-type sequence and structure.
- Combinatorial mutation effect prediction
- Estimating the combined functional consequences of multiple simultaneous amino acid changes within a protein target.
- Structure-aware representation extraction
- Generating fixed-length protein embedding vectors derived from mean-pooled final-layer hidden states that incorporate both sequence and 3Di structural tokens.
- Position-wise amino acid probability profiling
- Calculating the full categorical probability distribution of all standard amino acids at any selected position within a protein.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
SaProt
Strengths and limits
Good at
- Outperforms ESM2 across most functional prediction benchmarks when structural inputs are supplied.
- Combines sequence and Foldseek 3Di structural states directly in its tokenisation scheme.
- Accommodates structural uncertainty by incorporating pLDDT masking for regions scoring below 70 in AlphaFold 2 structures.
- Supports direct scoring of both individual and combinatorial amino acid variants.
Watch out for
- Frozen embeddings do not support amino-acid-only sequences and strictly require structure-aware (SA) tokens.
- Amino-acid-only sequences require task-specific fine-tuning to function.
- Requires high-quality atomic coordinates or predicted structures to achieve optimal results.
- Requires a minimum of 16 GB of GPU memory for deployment.
Inputs and outputs
You provide
Protein sequence
Single-letter amino acid code. Whitespace and FASTA headers are stripped automatically.
Structure (PDB or entry code)optional
Optional, but it is the reason to use this model rather than ESM2.
Mutations to scoreoptional
Comma-separated, e.g. A123V, W45F. Leave empty for an embedding.
You get back
Structure-aware representation.
Licence
Commercial use permitted · MIT
The MIT licence allows use in commercial research and products.
Questions
Can I use SaProt commercially?
Commercial use permitted. The MIT licence allows use in commercial research and products.
How do I run SaProt without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 16GB, typically in 2.5s, and a run costs about 14 credits. Visitors without an account get 1,000 credits a month.
Can SaProt be used without structural coordinates?
SaProt can process amino-acid-only sequences, but it requires fine-tuning to do so. Pre-trained, frozen embeddings are supported exclusively when structure-aware (SA) tokens are supplied.
How are structural coordinates represented in the model?
Structures are converted into discrete structural tokens using Foldseek's 3Di alphabet, which are then combined with corresponding amino acid letters into unified sequence-structure tokens.
How does the model handle uncertain structural predictions from AlphaFold 2?
Residues with an AlphaFold 2 pLDDT score below 70 can be masked using a dedicated '#' token to mitigate unreliable structural configurations.
How are whole-protein embeddings extracted?
Embeddings are calculated by taking the mean reduction of the hidden states across the model's final transformer layer.
Written from SaProt's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.