DisoPatho: A Cross-View Feature-Adaptive Interaction Encoding Framework for Predicting Disease-Associated Variants in Intrinsically Disordered Regions.

Mutation-centric architecture using the variant site as an anchor, combining energy representations specific to intrinsically disordered regions with protein language model embeddings, without explicit structural descriptors or multiple-sequence alignments.

Intrinsically disordered regions (IDRs) largely escape pathogenicity predictors, lacking stable structural conformations and showing high sequence variability. DisoPatho is a deep-learning framework dedicated to these regions, with a mutation-centric architecture in which the variant site anchors feature construction, combining IDR-specific energy representations with embeddings from protein language models (xTrimoPGLM and Evolutionary Scale Modeling), without explicit structural descriptors, multiple-sequence alignments or hand-crafted conservation scores. In 5-fold cross-validation it reaches average AUCs of 0.899 and 0.840 and accuracies of 0.862 and 0.860 on two datasets. On a highly confounded independent test set, where phylogenetic constraints provide limited discriminative signal, it achieves a 50.2% relative improvement in Matthews correlation coefficient over AlphaMissense on their respective predictable variants, with broader prediction coverage. Code, datasets and predictions are available for academic use.

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

The angle is well chosen: disordered regions are precisely where structure-based predictors, AlphaMissense first among them, are least reliable, and broader coverage has practical value when a missense variant falls in a disordered segment. The 50.2% gain must nonetheless be read for what it is — a relative improvement in a correlation coefficient on a selected test set, restricted to the variants each of the two methods can predict — and not as a diagnostic gain. Without calibration into evidence weights or evaluation on routinely classified variants, the tool remains a hypothesis aid, not a classification criterion.

Analysis by Dr Thibaut Benquey