Use SPURS Online
Commercially Available SPURS No-Code Web Server
SPURS: Rapid, Generalizable Protein Stability Prediction (ΔΔG) for Point Mutations
SPURS is a deep learning framework for predicting how amino acid substitutions change protein thermostability (ΔΔG). It was developed by Ziang Li and Yunan Luo at the School of Computational Science and Engineering, Georgia Tech, and is described in Nature Communications. The authors release the code under the MIT license.
SPURS combines a protein language model with an inverse folding model, fine-tuned on mega-scale stability data, so that one forward pass returns predictions for every possible single substitution.
Stability is central to protein engineering: stabilizing mutations improve expression and shelf life, while destabilizing ones are often linked to loss of function and disease. Experimentally scanning every substitution is slow and costly, so accurate computational predictors are valuable.
The authors note that stability datasets are biased toward destabilizing variants and that prior predictors can struggle to generalize across proteins. SPURS addresses this by rewiring pretrained generative models rather than training a large model from scratch.
Item | Detail |
|---|---|
Task | Predict ΔΔG for amino acid substitutions (single and higher-order mutants) |
Core models | ESM2 protein language model plus ProteinMPNN structure encoder, joined by a lightweight Adapter |
Training data | Megascale stability dataset (272,721 single mutants across 298 proteins) |
Speed | All single mutants for 118 ProteinGym proteins (mean length 492) in under 20 seconds on one NVIDIA A40 GPU |
Venue | Nature Communications (2026), open access |
License | MIT (code) |
How SPURS Predicts Protein Stability Changes
The method keeps ESM2 frozen and injects structure-derived features into its sequence embeddings through an Adapter module. Only the Adapter and ProteinMPNN parameters are trained, which reduces trainable parameters by 98.5% compared with fully fine-tuning ESM.
Structure input: an experimental structure is used when available, otherwise an AlphaFold-predicted structure.
Whole-protein scan: ΔΔG for all possible point mutations is produced in one forward pass.
Higher-order mutants: effects are combined by summing single-mutation effects plus a learned epistasis term.
SPURS Benchmark Results for ΔΔG Prediction
The authors evaluated SPURS on 12 datasets covering ΔΔG and melting temperature, including eight independent test sets and the Human Domainome dataset.
Evaluation | Result |
|---|---|
Megascale test set (28 proteins, 28,312 mutants) | Median Spearman 0.83 vs 0.77 for ThermoMPNN; better in 24 of 28 proteins |
Human Domainome (563,534 variants, 522 proteins) | Spearman 0.54 vs 0.49 for the best baseline (ThermoMPNN) |
Eight independent test sets | Significantly higher or comparable correlations; significant gains in 7 of 8 |
Functional site identification (239 human domains) | Average AUROC 0.69 with no training on functional labels |
Low-N fitness (141 ProteinGym datasets) | SPURS-augmented DeepSequence beat the original in 115 (82%); about 15% overall Spearman gain |
Beyond Stability: Functional Residues and Fitness
The same representations help flag functionally important residues and improve low-N fitness prediction, and the paper also studies stability and pathogenicity in human disease.
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 SPURS on Tamarind Bio
Stabilizing mutation search: rank all substitutions in a protein to shortlist candidates that lower ΔΔG.
Variant effect triage: screen large variant sets, such as human disease variants, for stability impact.
Functional site discovery: use stability signals to highlight residues likely important for function.
Low-N engineering: add SPURS features to fitness models when few labeled variants exist.
How to Use SPURS on Tamarind Bio
Log in and open the tool: go to tamarind.bio and select SPURS from the tool list.
Provide a structure: upload a PDB file and, if exposed, specify the chain ID, as the SPURS code expects.
Choose the mutation scope: single-mutation scanning returns a per-position matrix over the 20 amino acids; if exposed, supply specific mutation lists for multi-mutation scoring.
Run the job: submit and wait for the prediction to finish.
Download outputs: the single-mutation output is a ΔΔG matrix normalized so wild-type entries are zero.
Interpret: compare per-substitution ΔΔG values to identify stabilizing and destabilizing candidates, keeping the benchmark caveats below in mind.
Use downstream: select top candidates for experimental validation or further design.
Things to Keep in Mind
The README reports Python 3.7 and PyTorch 1.12 with CUDA 11.3 for the original setup, which is one reason a hosted run is convenient.
Stability training data are biased toward destabilizing variants, according to the paper.
Structure quality matters: when no experimental structure exists, AlphaFold predictions are used.
Functional site identification has moderate accuracy (average AUROC 0.69).
The low-N augmented model is an add-on to competitive predictors, not a standalone fitness predictor.
Predictions are computational estimates and should be validated experimentally.
Source: Ziang Li and Yunan Luo, "Generalizable and scalable protein stability prediction with rewired protein generative models," Nature Communications 17, article 891 (2026), DOI 10.1038/s41467-025-67609-4. Code: https://github.com/luo-group/SPURS.