Use ThermoMPNN Online
Commercially Available ThermoMPNN No-Code Web Server
ThermoMPNN: Predicting Protein Stability Changes (ΔΔG) from Single Amino Acid Mutations
ThermoMPNN is a graph neural network that predicts changes in protein stability for point mutations using transfer learning from ProteinMPNN. It was developed by Henry Dieckhaus, Michael Brocidiacono, Nicholas Z. Randolph and Brian Kuhlman and published in PNAS in 2024. The code is released under the MIT license.
From a single wild-type structure, ThermoMPNN can run site-saturation mutagenesis and rank substitutions by their predicted effect on stability.
Computational methods for predicting how mutations perturb stability are of great interest for protein engineering, but experimental stability datasets have historically been small. The recent Megascale dataset is much larger, which makes data-hungry deep learning practical.
ThermoMPNN reuses the representations of a pretrained inverse folding model rather than learning structure features from scratch.
Item | Detail |
|---|---|
Task | Predict ΔΔG for single point mutants |
Architecture | Frozen ProteinMPNN embeddings, light attention block and small MLP |
Training data (preprint) | Megascale (272,712 mutations, 298 proteins) and curated Fireprot (3,438 mutations, 100 proteins) |
Speed (preprint) | Site-saturation mutagenesis in about 2 seconds for a 105-residue protein; about 8 seconds for 693 residues on one GPU |
Venue | PNAS 121(6), e2314853121 (2024) |
License | MIT (code) |
How ThermoMPNN Uses Transfer Learning for Stability Prediction
Structure encoding: the wild-type structure is passed through a frozen, pretrained ProteinMPNN.
Embedding extraction: decoder embeddings at the mutated position are combined with the wild-type residue embedding.
Prediction head: a light attention block and small MLP output the stability change.
Scope: single point mutants; double mutants are covered by a separate ThermoMPNN-D repository.
ThermoMPNN Accuracy on Megascale, Fireprot and Other Benchmarks
The figures below are from the bioRxiv preprint version of the work.
Benchmark | Result |
|---|---|
Megascale test | Pearson 0.754; Spearman 0.725 with transfer learning |
Unmodified ProteinMPNN baseline (Megascale) | Spearman 0.487 |
Fireprot | Spearman 0.657 vs 0.50 for ProteinMPNN |
SSYM direct / inverse | Pearson 0.72 / 0.60 |
S669 | Pearson 0.43; some methods outperformed ThermoMPNN |
vs Rosetta, RaSP, PROSTATA | Pearson 0.04-0.05 higher on each dataset |
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 ThermoMPNN on Tamarind Bio
Stabilizing mutation discovery: scan every position of an enzyme or binder for stability-enhancing substitutions.
Triage of design candidates: filter candidate mutations before experimental testing.
Site-saturation mutagenesis: predict a full mutational landscape from one PDB file.
Pipeline integration: run fast scans as part of larger protein engineering workflows.
How to Use ThermoMPNN on Tamarind Bio
Log in and open the tool: go to tamarind.bio and select ThermoMPNN.
Provide a structure: upload a PDB file, as the repository's custom inference expects.
Select scope: site-saturation mutagenesis over all positions, or specific chains or positions, if exposed.
Run the job and wait for predictions.
Download outputs: predicted stability changes per mutation.
Interpret: rank substitutions by predicted ΔΔG and avoid over-reading surface cysteine results (see below).
Use downstream: choose a small set of mutations to validate experimentally.
Things to Keep in Mind
Megascale covers a limited dynamic range (about -3 to 5 kcal/mol); the model struggles with large-effect outliers.
Error tends to rise for larger proteins, and Megascale contains only small proteins (under 75 residues).
Surface cysteine mutations look stabilizing in Megascale because of an assay artifact and should be excluded from predictions.
Fireprot is skewed toward alanine mutations, and generalization on S669 was weaker.
Covers single point mutants; the indel model (ThermoMPNN-I) has limited in silico validation only.
Source: Henry Dieckhaus, Michael Brocidiacono, Nicholas Z. Randolph and Brian Kuhlman, "Transfer learning to leverage larger datasets for improved prediction of protein stability changes," PNAS 121(6), e2314853121 (2024), DOI 10.1073/pnas.2314853121; preprint DOI 10.1101/2023.07.27.550881. Code: https://github.com/Kuhlman-Lab/ThermoMPNN.