#21 · ML / AI methods

Deep learning predicts RNA structure, IRES activity, and codon design from sequence

Across these three items, deep learning operates directly on RNA sequence for distinct tasks: end-to-end 3D structure prediction for RNA and single-stranded DNA (NUMonomer, preprint), internal ribosome entry site prediction in mRNAs (DB-IRES), and codon design (RNARL). DB-IRES combines densely connected 1D CNNs, BiGRUs, and self-attention in an ensemble. NUMonomer is reported to outperform leading methods using sequence alone. RNARL applies reinforcement learning to unify generation and multi-objective optimization and is reported to handle sequences over 3,900 nt. The dominant thread is task-specific sequence-to-property prediction and design with purpose-built architectures, rather than reuse of frozen pretrained representations. What is new is sequence-only tertiary structure prediction for nucleic acids and reinforcement-driven generative design for long RNA sequences.