#25 · ML / AI methods

Language-model architectures for RNA binding prediction and mRNA sequence design

The dominant thread is the application of transformer and language-model architectures to RNA sequence problems. DLRNA-BERTa combines RNA-BERTa and ChemBERTa-v2 to predict molecule-RNA binding affinities, circMAC localizes miRNA-conditioned binding sites on full-length circRNA isoforms, and GRASSP fuses RNA language model embeddings with graph attention over predicted structure to predict RNA-small molecule binding sites. A second strand concerns sequence design. VeloHARMO pairs organism-specific models trained on Ribo-seq data with a genetic algorithm to design synonymous coding sequences, and ID3 applies differentiable optimization over codon choices while preserving amino acids. ATSFormer applies attention-guided token pruning to long genomic sequences for RNA modification analysis, and the UTR study links 5' and 3' UTR sequences using pretrained embeddings and contrastive learning. What is new is the reuse of pretrained language-model representations together with learned or differentiable optimization for both prediction and design.