ProSST Mutation Effect Prediction
ProSST predicts protein mutation effects and functions by integrating sequence and structural data via quantized tokens and disentangled attention.
ProSST is a protein language model that integrates sequence and structural information for enhanced protein function prediction. It employs a structure quantization module to convert 3D protein structures into discrete tokens using a GVP encoder and k-means clustering, coupled with a Transformer architecture featuring sequence-structure disentangled attention. Pre-trained on 18.8 million protein structures, ProSST achieves state-of-the-art performance in zero-shot mutation effect prediction (e.g., Spearman correlation of 0.504 on ProteinGYM) and supervised tasks like protein localization (94.32% accuracy on DeepLoc). The model explicitly learns residue-structure relationships, enabling robust predictions even for stability, binding, and expression-related mutations.
Estimated resource cost: 0.0002945
Categories: Protein Design, Sequence Analysis & Annotation
Tags: Mutational Scanning Proteins
Key Capabilities
- Combines protein sequences and 3D structures via a Transformer with disentangled attention.
- Quantizes local residue structures into discrete tokens using a GVP encoder and k-means clustering.
- Outperforms SOTA models in zero-shot mutation prediction (Spearman: 0.504) and supervised tasks (e.g., 94.32% accuracy on DeepLoc).
- Pre-trained on 18.8 million protein structures from AlphaFoldDB, enabling broad applicability.
Runtime Statistics
| Metric | Value |
|---|---|
| runtime_mean | 158 |
| runtime_median | 12 |
| runtime_std | 373 |
| runtime_90th_percentile | 491 |
| runtime_max | 1987 |
Similar Tools
- BindFilter
- Efficient Evolution
- ProteinMPNN-ddG
- StaB-ddG
Ready to submit your job?
Review your configuration, then confirm the estimated credit cost before you run the job. Note that credit estimates are not guaranteed and runtime can vary depending on inputs and settings.
Estimated Credits: 0.0002945
Invite-only, limited-time access. Please contact ztang@getantibody.com.
Invite-only, limited-time access. Please contact ztang@getantibody.com.