Can self-supervised learning predict how mutations change binding?
A graph network pretrained to repair perturbed protein structures predicted mutation effects on binding better than energy-based and feature-engineered methods, even on structures it had not seen.
Source
Deep geometric representations for modeling effects of mutations on protein-protein binding affinity
Study at a glance
- Design
- Computational / modelling — Self-supervised GNN pretraining on unlabelled complexes, followed by a supervised predictor, evaluated on six benchmarks with split-by-structure cross-validation and an independent test set
- N
- There is no single N. Pretraining used 13590 unlabelled complexes, and six benchmark datasets were evaluated (for example S1131 and M1707, whose names give their data-point counts). The independent test set had 641 data points.
- Population
- Protein-protein complexes with experimentally measured binding-affinity changes upon mutation
- Outcome
- Pearson correlation and RMSE between predicted and measured binding-affinity changes (ΔΔG)
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors built GeoPPI. Its graph neural network encoder first learns, without labels, to reconstruct protein complexes whose side chains have been randomly twisted. A second model then uses these learned representations to predict how mutations change binding affinity. They tested it on six standard mutation datasets using a split-by-structure cross-validation, so that training and test folds shared no protein domain. They also tested it on an independent dataset and on SARS-CoV-2 antibody data.
What they found
After pretraining, the encoder's representations separated interface from non-interface residues and grouped amino acids by chemical property, even though it was never given these labels. GeoPPI beat every baseline on all single-mutation benchmarks, including a 45% gain in Pearson correlation over the previous best method on one dataset. On especially subtle, conservative mutations it reached a correlation of 0.66, compared with 0.21 for its competitor. Performance dropped for every method, including physics-based FoldX, on the independent test set, but GeoPPI still ranked highest.
The limits
What it doesn't show
All methods, GeoPPI included, performed poorly on the truly independent test set, so real-world generalisation remains limited. Evidence that the encoder learned meaningful structure comes mainly from t-SNE visualisations, which are qualitative. The antibody applications relied on homology-modelled and docked structures, and the designed mutations were predictions that were not validated in the lab within this paper.
Key terms
- Self-supervised learning
- Training on a task built from unlabelled data itself, here restoring perturbed protein structures, in order to learn useful representations.
- Graph neural network
- A network that passes information between connected nodes, here atoms linked by spatial proximity.
- ΔΔG (binding affinity change)
- How much a mutation changes the free energy of binding between two proteins.
- Split-by-structure cross-validation
- A cross-validation scheme in which training and test folds share no protein domain, which prevents inflated scores from near-duplicate structures.
- t-SNE
- A method that projects high-dimensional representations into two dimensions for visualisation.
Flashcards
0 of 9 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What does GeoPPI's self-supervised stage learn from?
Common questions
Why split data by structure rather than randomly?
Benchmark complexes are often closely related. With random splits a model can memorise near-duplicates, and its scores collapse on genuinely new proteins.
What is the pretraining task?
Side chains of residues are randomly rotated, and the encoder learns to recover the original structure. This teaches it what normal atomic interactions look like.
Does GeoPPI generalise well to new proteins?
Better than the baselines, but not well in absolute terms. Every method dropped sharply on the independent test set.
More on Graph neural networks