#3 · ML / AI methods

Graph and transformer models infer regulatory networks and disease-linked RNAs

The dominant thread is pairing a learned architecture with an external biological information source to rank candidates. TEA-GCN builds co-expression networks from public RNA-seq, and GlncDDR applies graph neural network embeddings of cancer expression to predict 1232 candidate DDR-associated lncRNAs. CASER integrates multi-omics features in semi-supervised tri-training to prioritize epigenetic regulator genes, while PSGRN uses self-training with synthetic gold standards on single-cell perturbation data. GRNPred pairs masked expression pretraining with fine-tuning for TF-target inference, and APMF combines random-walk sequences with transformer contrastive learning for disease-related miRNAs. Not every item shares an explicit prior; what is new is where explanatory leverage sits, in graph structure, expression embeddings, and multi-omics integration rather than split design. The available summaries do not establish held-out generalization, so transferability remains untested at this level.