Use SPURS Online

SPURS: Generalizable and Scalable Protein Stability Prediction

Predict changes in protein thermostability (ΔΔG) for all single mutations in a single forward pass using a rewired, multimodal protein foundation model framework.

What is SPURS?

SPURS (Stability Prediction Using a Rewired Strategy) is a state-of-the-art deep learning framework developed to predict changes in protein thermostability caused by amino acid substitutions. Efficiently modeling protein stability is a cornerstone of understanding human disease mechanisms and engineering robust proteins for industrial and therapeutic applications.

Traditional machine learning predictors are often trained on small datasets and struggle to generalize to unseen proteins or rare mutations. SPURS overcomes these limitations by seamlessly fusing the complementary data modalities of two leading foundation models: ESM2 (a protein language model trained on massive sequence data) and ProteinMPNN (a structure-based inverse folding model).

By embedding a lightweight cross-attention Adapter module into the network architecture, SPURS integrates evolutionary sequence priors with 3D structural geometric features. It undergoes parameter-efficient fine-tuning on a mega-scale dataset of over 270,000 mutational stability measurements.

Key Features & Capabilities

• All-Mutants-Per-Pass Scalability: Traditional methods require separate computational forward passes for every individual mutant sequence (O(L x 20) computational cost). SPURS processes the wild-type sequence and structure to score all possible point mutations simultaneously in a single forward pass, executing site-saturation mutagenesis for an average protein in under 20 seconds.

• State-of-the-Art Accuracy: Outperforms leading predictors like ThermoMPNN across 12 diverse independent validation benchmarks measuring both free energy changes (ΔΔG) and melting temperature alterations (ΔTm).

• Unbiased Stabilizing Mutation Discovery: Stability datasets are heavily biased toward destabilizing variants, causing most AI models to perform poorly when identifying stabilizing mutations. SPURS breaks this bottleneck, demonstrating significantly higher precision and recall for prioritizing stabilizing variants (ΔΔG < -0.5 kcal/mol).

• Higher-Order Epistasis Decoding: Beyond single mutations, SPURS incorporates a dedicated non-additive epistatic decoder to capture complex combinatorial effects, reliably predicting stability modifications for double and higher-order mutants.

• Broad Downstream Biological Applications: Deep mutational stability outputs from SPURS can scale to identify functional residues, guide low-N sequence fitness predictions, and map stability-pathogenicity landscapes across the human proteome.

How SPURS Works Under the Hood

   [Wild-Type Sequence] --------> ESM2 (Frozen PLM) ---------\
                                                              \--> [Cross-Attention Adapter] --> Structure-Enhanced Features --> All Single Mutants (1 Pass)
   [3D Structure (PDB/AF2)] ----> ProteinMPNN (Structure Encoder)

   [Wild-Type Sequence] --------> ESM2 (Frozen PLM) ---------\
                                                              \--> [Cross-Attention Adapter] --> Structure-Enhanced Features --> All Single Mutants (1 Pass)
   [3D Structure (PDB/AF2)] ----> ProteinMPNN (Structure Encoder)

   [Wild-Type Sequence] --------> ESM2 (Frozen PLM) ---------\
                                                              \--> [Cross-Attention Adapter] --> Structure-Enhanced Features --> All Single Mutants (1 Pass)
   [3D Structure (PDB/AF2)] ----> ProteinMPNN (Structure Encoder)

1. Multimodal Representation: The system receives the wild-type protein sequence alongside its 3D atomic coordinates. If an experimental structure is unavailable, AlphaFold2-predicted structures are seamlessly generated.

2. Neural Network Rewiring: Rather than blending outputs heuristically, SPURS's cross-attention Adapter injects structural constraints learned from ProteinMPNN directly into the intermediate layers of the frozen ESM2 text transformer.

3. Parameter Efficiency: During training, 650 million parameters of the base PLM remain frozen. Updating only the structural encoder and the Adapter reduces trainable parameters by 98.5%, effectively safeguarding the model from overfitting.

4. Shared Decoder Matrix: A shared Multi-Layer Perceptron (MLP) projects per-residue latent representations to an L x 20 matrix, allowing rapid arithmetic retrieval of thermodynamic shifts for all substitutions simultaneously.

What is Tamarind Bio?

Tamarind Bio is a pioneering no-code bioinformatics platform built to democratize access to powerful computational biology tools for life scientists, chemists, and researchers. Recognizing that advanced machine learning models are frequently gatekept by command-line interfaces, local hardware constraints, and delicate software dependencies, Tamarind Bio provides a fully abstracted, secure web-based workspace.

The platform seamlessly manages the underlying high-performance computing, GPU orchestration, and complex parallelization workflows. This architecture empowers laboratory scientists to upload experimental biological data, execute deep learning simulations at a massive scale, and secure their enterprise data or proprietary IP under a secure cloud environment.

How to Use SPURS on Tamarind Bio

Tamarind Bio completely removes the technical friction of deploying large models like ESM2 and ProteinMPNN, compressing sitewide proteome profiling into an intuitive workflow:

1. Access the Tool: Log in to the platform interface at tamarind.bio and select the SPURS tool.

2. Input Wild-Type Sequence: Paste the raw amino acid sequence of your target protein into the input field or import multiple sequences using a standard file type.

3. Provide or Generate Structure: Provide a 4-character Protein Data Bank (PDB) ID or upload a structural .pdb coordinate file. If your protein structure has not been experimentally determined, select the AlphaFold2 Auto-Predict toggle to construct a high-fidelity structural backbone natively on the platform.

4. Configure Mutational Depth: Choose between generating a complete site-saturation mutagenesis heatmap (all single mutations calculated simultaneously in one job submission) or input a custom string list of specified high-order combinatorial variants to execute the epistasis decoding model.

5. Run Inference & Export Results: Click Submit Job. Tamarind Bio spins up scalable cloud graphics hardware to deliver results within seconds. Download interactive stability heatmaps, view structure-enhanced evolutionary features, or immediately pipeline the data into other sitting platform options, such as the Tamarind Bio Assay Portal, to order wet-lab validation assays.

Source