FoldX interface ΔΔG depends on which structure it starts from, and a learned predictor does not: a cautionary analysis on glutaryl-CoA dehydrogenase
- 1Independent researcher
- 2Yıldız Technical University
Abstract
FoldX and related methods are widely used to interpret missense variants, and their sensitivity to the input structure is documented: repair of the coordinates is mandatory, averaging over structural ensembles improves agreement with experiment, and the uncertainty of a FoldX folding ΔΔG has been put at roughly ±3 kcal/mol, with binding stability worse. What has not been asked is whether that uncertainty is spread evenly over a protein. We took 258 ClinVar missense variants in glutaryl-CoA dehydrogenase, a homotetramer, verified every reference residue against the 1SIQ structure and computed ΔΔG for the 124 classified pathogenic. Eighteen variants were then recomputed on three molecular dynamics snapshots of the same complex: six interface variants selected on their monomer-to-tetramer difference, six interface variants drawn at random, and six at least 10 Å from any contact. The displacement between the crystal-based value and the relaxed-frame mean had a median of 1.10, 1.38 and 0.20 kcal/mol in the three groups (random interface against distant, Mann-Whitney p = 0.0130), while frame-to-frame scatter was indistinguishable between them (Kruskal-Wallis p = 0.385). The variation is therefore positional, and it is not explained by noise in the mean. SAAMBE-3D, a learned interface predictor, run on the same eighteen variants and the same four structures, moved by a median of 0.075 kcal/mol at the interface and by nothing at all away from it, so the instability is largely a property of FoldX. It also depends on magnitude: G171R, scored at +22.53 kcal/mol, moved 2.7%, whereas A421V moved from +0.18 kcal/mol, which reads as tolerated, to +4.37 ± 2.22. Along the way we found that ThermoMPNN given a chain selector on a tetramer returns values identical to the isolated monomer at all 7,800 scanned positions. For variant interpretation the consequence is specific: at a subunit interface, in the range where a classification is actually in doubt, a single starting structure does not support a single number.
Related
- Cross-clade limits of melanosome-based color prediction: a cautionary machine learning study
- Campo de Dalías and the “Almeria miracle”: a cautionary tale of the value of water in agriculture
- Gate-Based Feature Importance in Simulation-Based Inference for EEG: A Cautionary Tale
- Hexose-6-phosphate dehydrogenase deficiency disrupts hepatic fatty acid homeostasis and induces triglyceride accumulation
- Reductive carboxylation via isocitrate dehydrogenase 1 supports cardiac metabolic adaptation during oncometabolic stress.
BibTeX
@misc{ekmekci2026foldx,
title = {{FoldX interface ΔΔG depends on which structure it starts from, and a learned predictor does not: a cautionary analysis on glutaryl-CoA dehydrogenase}},
author = {Cumhur Gökhan Ekmekci and Ceren Aleyna Cengiz},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010007/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010007/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - FoldX interface ΔΔG depends on which structure it starts from, and a learned predictor does not: a cautionary analysis on glutaryl-CoA dehydrogenase
AU - Cumhur Gökhan Ekmekci
AU - Ceren Aleyna Cengiz
PY - 2026
DA - 2026/10/06
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010007/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010007/v1
AB - FoldX and related methods are widely used to interpret missense variants, and their sensitivity to the input structure is documented: repair of the coordinates is mandatory, averaging over structural ensembles improves agreement with experiment, and the uncertainty of a FoldX folding ΔΔG has been put at roughly ±3 kcal/mol, with binding stability worse. What has not been asked is whether that uncertainty is spread evenly over a protein. We took 258 ClinVar missense variants in glutaryl-CoA dehydrogenase, a homotetramer, verified every reference residue against the 1SIQ structure and computed ΔΔG for the 124 classified pathogenic. Eighteen variants were then recomputed on three molecular dynamics snapshots of the same complex: six interface variants selected on their monomer-to-tetramer difference, six interface variants drawn at random, and six at least 10 Å from any contact. The displacement between the crystal-based value and the relaxed-frame mean had a median of 1.10, 1.38 and 0.20 kcal/mol in the three groups (random interface against distant, Mann-Whitney p = 0.0130), while frame-to-frame scatter was indistinguishable between them (Kruskal-Wallis p = 0.385). The variation is therefore positional, and it is not explained by noise in the mean. SAAMBE-3D, a learned interface predictor, run on the same eighteen variants and the same four structures, moved by a median of 0.075 kcal/mol at the interface and by nothing at all away from it, so the instability is largely a property of FoldX. It also depends on magnitude: G171R, scored at +22.53 kcal/mol, moved 2.7%, whereas A421V moved from +0.18 kcal/mol, which reads as tolerated, to +4.37 ± 2.22. Along the way we found that ThermoMPNN given a chain selector on a tetramer returns values identical to the isolated monomer at all 7,800 scanned positions. For variant interpretation the consequence is specific: at a subunit interface, in the range where a classification is actually in doubt, a single starting structure does not support a single number.
ER -