Use SaProt Online
Commercially Available SaProt No-Code Web Server
SaProt: Structure-Aware Protein Language Model for Variant Effect and Function Prediction
SaProt is a general-purpose protein language model that combines amino acid sequence with 3D structure information through a structure-aware vocabulary. It was developed by Jin Su, Chenchen Han, Yuyang Zhou, Junjie Shan, Xibin Zhou and Fajie Yuan at Westlake University. The authors report that SaProt outperforms established baselines on 10 downstream tasks. The code is released under the MIT license on GitHub, with pretrained weights hosted on Hugging Face.
Large protein language models such as the ESM family perform well on structure and function tasks, but they are trained only on residue sequences and do not explicitly use structure. Because function is closely tied to 3D structure, a model that sees both can in principle learn richer representations.
SaProt addresses this by pairing each residue with a structure token, so that a single sequence carries both identity and local geometry. The result is a model that can score mutations, produce embeddings and be fine-tuned for supervised protein tasks.
Property | Detail |
|---|---|
Model type | Protein language model with a structure-aware vocabulary |
Structure tokens | Foldseek 3Di, 20 local structure states |
Vocabulary size | 441 tokens (21 x 21, including mask symbols) |
Architecture | Same as the 650M-parameter ESM-2, with a larger embedding layer |
Pretraining data | About 40 million AlphaFold2 structures and sequences |
Released variants | 35M, 650M and 1.3B parameter checkpoints, plus a PDB-tuned 650M model |
How SaProt's Structure-Aware Vocabulary Works
The core idea is to encode each residue's local 3D geometry as one of 20 Foldseek structure states and fuse it with the amino acid letter. Any protein sequence plus structure can therefore be written as a sequence of combined tokens and fed to a standard transformer.
Encode structure: Foldseek converts each residue's local environment into a 3Di structure token, using its default setup.
Fuse tokens: each residue token is combined with its structure token, and adding a mask symbol on each side gives 441 tokens in total.
Pretrain with masking: a BERT-style masked language modeling objective hides residue tokens while structure tokens stay visible, so the model learns to predict residue identity. The authors say this worked better than predicting structure tokens.
Handle low confidence: for AlphaFold2 residues with pLDDT below 70, only the residue identity is visible, and such positions are predicted only when selected for masking.
Training setup
The authors trained with the AdamW optimizer, sequences truncated to 1,024 tokens, batch size 512 and about 3 million steps. They report roughly 3 months on 64 NVIDIA A100 80GB GPUs. A further model, SaProt-PDB, was pretrained on 60,000 experimental PDB structures.
SaProt Benchmark Results for Mutation Effect Prediction and Protein Tasks
Zero-shot mutational effect prediction: on ClinVar and ProteinGym, SaProt outperforms ESM-1b, ESM-1v, ESM-2 650M, Tranception L, MIF-ST and ESM-IF. The repository reports a ClinVar AUC of 0.909 for the 650M model versus 0.862 for ESM-2 650M.
Supervised tasks: SaProt beats ESM-2 on all protein-level tasks tested, with the largest gains on thermostability, HumanPPI, metal ion binding and DeepLoc. Tasks also include EC and GO prediction.
Contact prediction: with a frozen backbone, SaProt outperforms ESM-2, which the authors read as evidence of better structural information.
With MSA information: adding MSAs improves all models, and SaProt still scores highest in the authors' comparison.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with SaProt on Tamarind Bio
Variant interpretation: score the likely effect of point mutations, as in the ClinVar and ProteinGym evaluations.
Protein engineering: use fitness-landscape predictions to rank candidate mutations, for example for thermostability.
Function annotation: fine-tuned SaProt models address tasks such as metal ion binding, EC and GO prediction and subcellular localization.
Interaction studies: the authors evaluate protein-protein interaction prediction on HumanPPI.
How to Use SaProt on Tamarind Bio
Open the tool: log in to tamarind.bio and open the SaProt tool page.
Provide your protein: supply the sequence and, as the model expects, a 3D structure for it (for example an experimental or AlphaFold2 structure), if the interface exposes this input.
Choose the task: select mutational effect scoring or embedding generation, if exposed.
Set parameters: the source recommends masking low-confidence AlphaFold2 residues (pLDDT below 70), if exposed.
Run the job: submit and wait for the job to complete.
Download outputs: retrieve the scores or embeddings from the results page.
Interpret results: compare mutation scores across variants to prioritize candidates.
Use downstream: feed rankings or embeddings into your own analysis or experimental design.
Things to Keep in Mind
Performance depends heavily on the accuracy of Foldseek's structure encoding, which the authors say has room to improve.
The 650M model may not have reached its maximum capacity because of computational limits.
When structure tokens are masked, SaProt performs on par with ESM-1b and ESM-2 rather than better.
The repository notes that the 35M and 650M models need structural (SA token) input for best performance.
Applications such as complex structure prediction and structure-constrained generation remain unexplored in the paper.
Source: Jin Su, Chenchen Han, Yuyang Zhou, Junjie Shan, Xibin Zhou, Fajie Yuan. SaProt: Protein Language Modeling with Structure-aware Vocabulary. bioRxiv preprint, 2023, DOI 10.1101/2023.10.01.560349. Code: https://github.com/westlake-repl/SaProt