#14 · ML / AI methods

Sequence models predict viral evolution, RNA structure and protein mutational effects

The dominant thread is learned representations of biological sequences, using pretrained transformers, attention, or diffusion objectives, applied to distinct prediction targets. PETRA pretrains an evolutionary transformer to forecast SARS-CoV-2 mutations, while USP-ddG uses a dual-channel structural model with mixture-of-experts to estimate binding free energy changes from protein-protein interface mutations. SCISOR applies discrete diffusion to delete residues from protein sequences and fits evolutionary data competitively. In RNA, a PDB-curated benchmark tests whether deep learning secondary structure predictors generalize, and HydraRNA combines bidirectional state space and multi-head attention in a full-length RNA foundation model pretrained on mRNAs and ncRNAs. What is new is the breadth of targets: viral escape, protein sequence compression, structural generalization, and interface energetics are addressed within one family of sequence-learning approaches rather than a single shared task.