#16 · ML / AI methods

Transformer and graph models learn regulatory links from DNA, RNA and protein sequences

The dominant thread is learned representations of biological sequences and graphs, with transformer or graph architectures applied to predict molecular relationships. Deep-Plant is a supervised foundation model that predicts chromatin state from plant DNA sequence across tissues and outperforms fine-tuned DNA language models. Graph-based methods infer regulatory structure from expression data: AutoGERN models explicit edge embeddings with architecture search for scRNA-seq networks, GRNFormer applies a graph transformer with TF-anchored subgraph sampling across species, and DeepLMI combines sequence pretraining, self-attention and a global-enhanced graph convolutional network for lncRNA-miRNA interaction. What is new is heterogeneous input. TFBindFormer fuses DNA sequence and chromatin features with protein sequence and structure representations for TF-DNA binding, and LncADeep 2.0 combines peptide, sequence and structure features with transfer learning for lncRNA annotation. Not every study follows this pattern; the hypergraph scRNA-seq study applies random-walk clustering to a different problem.