A Graph-to-Image Representation Framework for Modeling Splice-Site Responses in SMN2 Exon 7 Variants

SVGI models treatment-induced splicing responses across 285 engineered SMN2 exon 7 5′ splice-site variants by combining ΔPSI, replicate-normalized responsiveness, co-splicing network connectivity, GAT-derived graph geometry, Gromov–Wasserstein grid mapping, and image-based classification. In sealed five-fold out-of-fold evaluation, ROC-AUC was 0.641 for branaplam and 0.717 for risdiplam, with average precision of 0.328 and 0.467.

Problem / scientific focus

Characterizing treatment-induced SMN2 exon 7 inclusion across 285 engineered 5′ splice-site variants
Modeling coordinated co-splicing relationships without allowing treatment-response information to define the baseline graph
Transforming learned graph geometry into a structured image representation while preserving relational neighborhoods
Predicting responsive and non-responsive splice-site variants under risdiplam and branaplam from a limited dataset

Objectives

Quantify treatment-induced exon 7 inclusion changes and replicate-normalized response strength for risdiplam and branaplam
Model coordinated baseline co-splicing structure and learn influential splice-site neighborhood relationships
Map graph-derived geometry to a fixed two-dimensional grid and encode ΔPSI, responsiveness, and weighted degree as complementary image channels
Evaluate held-out response classification with sealed five-fold out-of-fold predictions and uncertainty estimates

Methods

GEO GSE221868 PSI data for 285 engineered SMN2 exon 7 5′ splice-site variants
Treatment-induced ΔPSI and replicate-normalized responsiveness scoring
Drug-specific DMSO co-splicing networks using Spearman correlation and weighted degree
Fold-specific Graph Attention Network learning of splice-site neighborhood importance
Entropic Gromov–Wasserstein mapping to a 17 × 17 lattice with one-to-one Hungarian assignment
Target-specific 5 × 5 × 3 patches encoding ΔPSI, responsiveness, and weighted degree with leakage control
AlexNet-style CNN classification with sealed five-fold out-of-fold ROC-AUC, average precision, and bootstrap uncertainty

← Back to Research Projects