Enhancing missense variant classification in predicted intrinsically disordered regions.

Machine learning combining global disordered-region conformation features (ALBATROSS), phase separation features and ProtTransBertBFD protein embeddings of wild-type and mutant sequences, with disordered-region boundaries defined by AlphaFold-RSA predictions.

Over 25% of known deleterious variants fall in intrinsically disordered regions, where in silico missense predictors generally perform worse than in ordered regions of the protein. The authors propose a machine learning methodology integrating global disordered-region conformation features from ALBATROSS, phase separation features computed with BioPython and 1,024-dimensional protein embeddings produced by ProtTransBertBFD for both wild-type and mutant sequences, with disordered-region boundaries defined by AlphaFold-RSA predictions. Using ClinVar classifications as ground truth, AlphaMissense, EVE and ESM1b are the highest-scoring unsupervised predictors for these variants, and the baseline model built on disordered-region features alone reaches a PR-AUC of 0.817 on the held-out test set. Combined with these predictors, the disorder features raise AlphaMissense from 0.807 to 0.919 PR-AUC, ESM1b from 0.679 to 0.845 and EVE from 0.591 to 0.910.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The problem is real and well framed: disordered regions concentrate a large share of deleterious variants while being the weak point of structure-based predictors, and the reported PR-AUC gains are substantial, especially for EVE. The main limitation is the ground truth: ClinVar labels partly depend on those same predictors, which raises a circularity concern, and validation stops at a held-out test set with no diagnostic cohort. Useful as a complementary signal within an interpretation file, not as a standalone argument, until an evaluation on variants classified by functional or segregation evidence is published.

Analysis by Dr Thibaut Benquey