What question did this study set out to answer?

The aim is to evaluate whether complex-aware structural features improve the classification of TP53 variants as dominant-negative or loss-of-function.

April 12, 2026Open Access

Pre-registered evaluation of complex-aware structural features for TP53 dominant-negative versus loss-of-function classification: no detectable gain over an AlphaMissense plus monomer FoldX baseline

Key Points

The aim is to evaluate whether complex-aware structural features improve the classification of TP53 variants as dominant-negative or loss-of-function.
Pre-registered assessment of complex-aware features including FoldX interface calculations.
Used 105 TP53 variants, split into dominant-negative and loss-of-function groups.
Established criteria for evaluating model performance based on AUROC and importance of complex features.
The full model achieved a 5-fold AUROC of 0.904, marginally better than the baseline's 0.892.
Failed to meet three of the four pre-registered criteria for significance.
Positional data alone produced an AUROC of 0.946, indicating position significantly impacts classification.

Abstract

Distinguishing dominant-negative (DN) from loss-of-function (LOF) missense variants is clinically meaningful but not yet reliably solved at the per-variant level. AlphaMissense (AM) reports pathogenicity but not mechanism; the gene-level mLOF tool of Badonyi and Marsh reports mechanism propensity but not per-variant labels. We pre-registered a test of whether complex-aware structural features—FoldX interface ΔΔG computed on biological assemblies of the TP53 DBD tetramer (PDB 2AC0), together with intramolecular distance features—provide independent predictive signal beyond a monomer-only baseline (AM pathogenicity score plus FoldX monomer ΔΔG) for DN versus LOF classification. The training set comprised 105 strict variants from Giacomelli et al. 2018 (45 DN-only and 60 LOF-only). Stop/go criteria were fixed before model fitting and required all four of: AUROC delta > +0.03, bootstrap 95% CI excluding zero, at least one complex feature among the top five by importance, and fold-consistent improvement in at least 7 of 10 cross-validation repeats. The full model achieved a 5-fold cross-validated AUROC of 0.904 versus 0.892 for the baseline (delta +0.013, 95% CI −0.025, +0.055, 6/10 fold-consistent), failing three of the four pre-registered criteria; the fourth criterion could not be evaluated due to a downstream script failure but the verdict was already determined. Two parallel observations carry independent value for the field. First, the AlphaMissense + monomer baseline reached AUROC 0.892 on a mechanism task it was not trained for, suggesting that AM may already encode substantial mechanism-correlated signal in TP53 through evolutionary constraint, structural context, or both. Second, residue position alone, fit by a random forest, reached AUROC 0.946: in this strict Giacomelli-derived TP53 set, positional segregation between DN-enriched (DBD-resident) and LOF-enriched (largely TET-resident) variants dominates the classification task, substantially limiting the measurable marginal value of additional complex-aware FoldX features. We report this as a pre-registered negative result and recommend that future work in TP53 mechanism prediction adopt AM + monomer as the minimum baseline and report position-only random forest performance as a sanity check.

Pre-registered evaluation of complex-aware structural features for TP53 dominant-negative versus loss-of-function classification: no detectable gain over an AlphaMissense plus monomer FoldX baseline

Key Points

Abstract

Cite This Study