1249-GRM UMCA-Net: Uncertainty-Aware Multi-Stage Cross-Attention for Cost-Aware Multi-Omics Data Classification
Multi-omics data can improve disease classification, but profiling every molecular layer for every patient is costly, and most models cannot signal when a prediction is unreliable. We present UMCA-Net, a Transformer framework that acquires additional omics layers only when current evidence is insufficient. Global Joint Cross-Attention aligns heterogeneous layers, while evidential learning with Dempster–Shafer fusion quantifies uncertainty at each decision stage. On four benchmark cohorts (ROSMAP, LGG, BRCA, KIPAN), UMCA-Net exceeded the accuracy of all 21 baseline methods. On BRCA, 90.5% of samples were classified with mRNA alone, at 0.901 overall accuracy. UMCA-Net offers a practical path toward accurate, trustworthy, and cost-aware multi-omics diagnosis.
Authors
- Huan Huang (ORCID: https://orcid.org/0000-0002-3063-2960)
- Yehong Huang (ORCID: https://orcid.org/0009-0002-2758-8126)
Publication Details
- Journal
- DigitalCommons - Kennesaw State University (Kennesaw State University)
- Published
- 2026-11-20
- Primary Topic
- Bioinformatics and Genomic Networks
- Type
- article
- Field-Weighted Citation Impact
- 0.00