Use ProteinMPNN-ddG Online
Commercially Available ProteinMPNN-ddG No-Code Web Server
ProteinMPNN-ddG: Fast Unsupervised Prediction of Protein Stability Changes from Point Mutations
ProteinMPNN-ddG is an unsupervised method that predicts how single-point mutations change protein stability (ddG) by adjusting the log-odds output of the ProteinMPNN inverse folding model. It needs no experimental stability data and no retraining or fine-tuning. The same recipe is also applied to ESMIF, giving ESMif-ddG.
Estimating the stability effect of a mutation is central to protein engineering, but physics-based tools such as Rosetta and FoldX are slow, and supervised predictors depend on experimental training data. Inverse folding models already know which residues fit a backbone, yet using their raw scores for ddG is biased by amino acid frequency and backbone geometry.
ProteinMPNN-ddG corrects these biases and adds a cheaper decoding scheme, making proteome-scale saturation mutagenesis practical.
Item | Detail from the paper |
|---|---|
Task | Zero-shot ddG prediction for point mutations |
Base models | ProteinMPNN and ESMIF |
Training on stability data | None |
Throughput | Up to 10,000 residues per second on one V100 16 GB GPU |
Benchmarks | Tsuboyama (164 proteins), S2648 and S669 |
How ProteinMPNN-ddG Works
Mask and score: for a mutation X to Y at position i, the wild-type residue is masked and the model scores the change from all other sequence and backbone context.
Decode position i last: instead of a random autoregressive order, the authors generate orders where residue i is decoded last so it sees full context. This raised sequence recovery on the validation set by nearly 4% (375,445 residues from 1,464 clusters).
Subtract a single-residue term: a correction computed from the residue's backbone atoms alone removes the part of the score that reflects backbone geometry rather than stability, analogous to the unfolded-state term in classical free-energy calculations.
Average over structures: the correction is averaged over 5,615,050 residue geometries from 23,349 training clusters.
Tied decoding: sharing partial decoding cuts the extra cost for a 1,024-residue protein from over 200-fold to under 4-fold relative to one ProteinMPNN pass.
Benchmark Results and Proteome-Scale Speed
The authors report Success@10, the fraction of a protein's top-10 predicted mutations that are more stable than wild type. They compare against supervised predictors (RaSP, DDGun, ACDC-NN) even though their own method is unsupervised.
Model | Success@10 before | Success@10 after |
|---|---|---|
ProteinMPNN to ProteinMPNN-ddG | 66% | 77% |
ESMif to ESMif-ddG | 64% | 69% |
Speed
Saturation mutagenesis for all 23,391 AlphaFold2-predicted human proteome structures took about 30 minutes on one V100 16 GB GPU. The paper gives approximate speedups of a million-fold over Rosetta and FoldX and about 400-fold over RaSP, using comparison figures from another paper.
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 ProteinMPNN-ddG on Tamarind Bio
Stabilizing mutations: select the top-scoring candidates for experimental testing in protein engineering.
Proteome-wide screening: scan saturation mutagenesis across many structures.
Binding (exploratory): the authors note that logit differences with and without a binding partner might flag mutations that strengthen binding, though this was not benchmarked.
How to Use ProteinMPNN-ddG on Tamarind Bio
Open the tool: log in to tamarind.bio and open ProteinMPNN-ddG.
Provide a structure: upload the backbone structure of the protein, since the method conditions on backbone coordinates.
Choose mutations: specify positions or a full saturation scan, if exposed.
Pick the base model: select ProteinMPNN or ESMIF scoring, if exposed.
Run the job: submit and wait for scores.
Download outputs: retrieve predicted ddG scores per mutation.
Rank candidates: focus on the top-ranked mutations, as in the Success@10 metric.
Validate: test shortlisted variants experimentally.
Things to Keep in Mind
ESMIF decodes N to C terminus, so it cannot use residues after position i.
The ESMIF methionine correction is an empirical fit that uses ProteinMPNN as its reference.
For S2648 and S669, metrics were computed over all proteins combined because per-protein data are sparse.
The binding-affinity extension has not been benchmarked.
Tied decoding still costs up to about 4-fold more than a single ProteinMPNN pass for long proteins.
Source: ProteinMPNN-ddG preprint (authors and title not shown on the fetched page), bioRxiv, identifier 10.1101/2024.06.15.599145 (v3): biorxiv.org/content/10.1101/2024.06.15.599145v3.