Coordinate- and Sequence-Based Features for a new Combined Annotation-Dependent Depletion Framework of Structural Variants (CADD-SV v2.0)

Unified Random Forest model scoring structural variant deleteriousness from coordinate-based annotations and sequence-derived signals (SegmentNT)

The authors present CADD-SV v2.0, an improved machine learning framework for scoring structural variant deleteriousness. This version introduces a unified Random Forest model trained on an expanded set of proxy-neutral and proxy-deleterious variants from human and non-human primate genomes, integrating updated genomic annotations (constraint metrics, regulatory elements, chromatin architecture). It scores deletions, insertions, duplications and inversions in a single framework accounting for the variant and its flanking regions. The authors additionally explore sequence-derived annotations via SegmentNT, a deep learning model providing nucleotide-resolution functional predictions. CADD-SV v2.0 outperforms its previous version and other tools for genome-wide deleterious variant prioritization.

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

A solid update to an already recognized framework for structural variant prioritization, with a useful addition of inversions and exploration of sequence-model signals. The evaluation is rigorous but remains a computational benchmark, without validation on a real diagnostic cohort, and preprint status warrants caution. The tool is nonetheless directly relevant to SV interpretation in genome data, a persistent blind spot in diagnostics.

Analysis by Dr Thibaut Benquey