#36 · Structural bio & protein design

New tools predict RNA structure, binding sites and validate models from sequence

The dominant thread is computational prediction and assessment of RNA structure, binding, and interfaces. DUET predicts RNA-small molecule binding sites from sequence alone using a relational graph tower and frozen RNA language model convolutions. ScanFold3 accelerates RNA secondary structure scanning for structured targets in genomes and transcriptomes, reporting 38-129x gains over ScanFold2 on long sequences. A validation approach extends conkit-validate to RNA, comparing predicted contact patterns from deep-learning predictors against deposited structures. An AlphaFold 3 evaluation across more than 1,300 protein-RNA complexes and about 270 RNA-only systems finds fold and interface accuracy declining on novel RNAs. CryoID2 applies 3D U-Net analysis to cryoEM maps to identify protein and RNA candidates. What is new is the shift toward testing deep-learning RNA predictions against deposited structures and novel folds, not only generating them.