Convolutional networks
Can a 3D image network predict how mutations change protein stability?
Open access · cc by · source: Europe PMC
Training a 3D convolutional network on both each mutation and its reverse removed the usual bias toward predicting destabilisation, and removing related proteins from training showed earlier methods' scores were likely inflated.
Study at a glance
- Design
- Computational / modelling — Ensemble of ten 3D CNNs on voxelised wild-type and mutant structures, trained on homology-pruned data augmented with reverse mutations; compared with seventeen predictors on a balanced blind test set, plus ablations and case studies.
- N
- N=684 · 684 test mutations in S_sym (342 direct plus their reverses); training used 3,488 points (1,744 direct mutations plus reverses); case studies of 42 p53 and 134 myoglobin mutations; ClinVar variants for an application analysis.
- Population
- Experimentally measured single-point protein mutations and their structures
- Outcome
- Pearson correlation and RMSE between predicted and measured stability change (ΔΔG), and anti-symmetry bias between direct and reverse mutations
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
ThermoNet reached a correlation of about 0.47 with measured values for both direct and reverse mutations, with almost perfect anti-symmetry between the two (correlation of −0.96). A version trained only on direct mutations did similarly on direct mutations but essentially failed on reverse ones (r about −0.06). A version trained without homology pruning scored better on the same test (reverse r 0.59 vs 0.48), showing leakage inflates results. On ClinVar, ThermoNet placed 80.2% of benign variants in a near-neutral stability zone, versus 39.7% for FoldX.
Methodology
The authors built ThermoNet, which voxelises the 3D neighbourhood of a mutation site in both the original and a computer-modelled mutant structure, encodes atom properties as channels, and feeds this to an ensemble of ten 3D CNNs to predict the stability change. They pruned training proteins that were homologous to test proteins and doubled the data by adding each mutation's reverse, whose stability change is the negative. They tested on a balanced benchmark of direct and reverse mutations against seventeen other predictors, ran an ablation trained on direct mutations only, and applied the model to p53, myoglobin and ClinVar variants.
Limitations
Accuracy is moderate (correlations near 0.5), and after homology pruning two other methods, DDGun and DDGun3D, outperformed ThermoNet; the authors could not retrain competitors on the pruned data, so comparisons are not like-for-like. The method needs an experimental or modelled 3D structure and relies on Rosetta-generated mutant models rather than real mutant structures. The ClinVar analysis checks consistency with a theoretical model of protein evolution, not against measured stability values. Training data are small for a 3D CNN with over a hundred thousand parameters.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
Similar sequences in train and test sets make CNNs look better than they are.
A 3D CNN for protein stability changes reached correlations near 0.47, and a version trained without removing homologous proteins scored higher on the same test (reverse r 0.59 vs 0.48), showing leakage inflates results.
Evidence for the claim as stated.
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
Evidence for the claim as stated.
Open questions
Tensions this paper is part of
From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
Related papers in this topic
Same topic cluster — not a recommendation engine.