An extensible deep learning framework for multi-type chromatin interaction prediction through hierarchical adaptation strategies.

UniLoopNet predicts multiple types of 3D chromatin interactions from DNA sequences, achieving AUROCs above 0.90 in most tasks and 0.8934–0.9366 under stringent tests across multiple cell lines.

• Why it matters: Specialized single-task models limit accurate, scalable prediction of diverse interaction types, and robust evaluation under realistic constraints is needed to support sequence-to-3D genome applications.
• What they did: UniLoopNet uses a shared encoder-decoder with hierarchical adaptation via multi-task, transfer, and continual learning based on task semantic similarity, then benchmarks across chromosome-matched, distance-matched, imbalanced, and locus-disjoint settings.
• The result: In realistic 1:5 class-imbalance, AUPRCs are 0.7956 (enhancer-promoter) and 0.6935 (enhancer-enhancer), and cross-cell-line transfer via supervised fine-tuning yields performance comparable to within-cell-line training.