Systematic enrichment of protein stability datasets
MSc thesis code for rebalancing protein-stability datasets to reduce bias in ΔΔG predictors (JCIM 2022)
Overview
gRoR is the code accompanying my MSc work on flattening the ΔΔG curve in protein-stability datasets. Most public stability datasets (S2648, PON-tstab, Sʸᵐ) are dominated by neutral mutations concentrated near ΔΔG ≈ 0, which biases every predictor trained on them. gRoR implements a systematic under-sampling strategy that groups mutations by biochemical and/or structural similarity and then evenly samples across the ΔΔG range, producing subsets with less peaked distributions and more balanced amino-acid frequencies.
- Repository: github.com/naroddaldal/gRoR
- Paper: Kebabci, Timucin & Timucin, J. Chem. Inf. Model. 62, 1345–1355 (2022) — 10.1021/acs.jcim.2c00054
- Datasets: curated PON-tstab and five enriched subsets on OSF
What it does
- Curates the PON-tstab data set down to 1,451 mutations across 89 proteins by resolving repetitions, residue-numbering mismatches, and missing PDB information.
- Introduces a ternary labelling (neutral, destabilising, stabilising) with neutral defined as ΔΔG ∈ [−0.5, 0.5] kcal/mol, so the dense neutral zone can be diagnosed as a bias rather than merged into destabilising mutations.
- Encodes mutations with either the 20-letter alphabet or a 4-letter reduced alphabet grouped by side-chain biochemistry (aliphatic / aromatic / polar / charged), and optionally sub-groups them by secondary structure (helix / sheet / loop, from DSSP) and three bins of relative solvent-accessible surface area.
- From every 2 kcal/mol window of each subgroup, selects three mutations (min, median, max ΔΔG), systematically diluting the neutral peak.
- Produces five enriched subsets (20L, 4L, 4L/SS, 4L/ASA, 4L/SS/ASA) with reduced kurtosis and skewness and more balanced amino-acid frequencies than the parent PON-tstab.
Why it matters
Benchmarking 11 stability predictors (DeepDDG, mCSM, INPS-3D, I-Mutant2.0/3.0, SDM, MAESTRO, PoPMuSiC, DUET, iStable, iDeepDDG) on the curated PON-tstab shows that every predictor makes its smallest errors in the dense neutral zone, systematically under-estimates destabilising mutations, and over-estimates stabilising ones — regressing toward ΔΔG ≈ 0. The paper argues that peakedness (kurtosis) of the ΔΔG distribution is a data-set bias in its own right, distinct from the well-known destabilising/stabilising asymmetry (skewness). Errors on the enriched subsets were higher than on the parent set, confirming that the workflow concentrates the difficult-to-predict mutations — the ones with extreme ΔΔG that actually matter for disease variants.