DUET: sequence-only prediction of RNA-small molecule binding sites

DUET predicts RNA–small molecule binding sites with 0.240 +/- 0.013 MCC on the benchmark dataset, outperforming CoBRA-M3 at 0.192 +/- 0.031 under matched overlap controls.

• Why it matters: Sequence-only RNA predictors can be inflated when training pools overlap with benchmark chains, and overlap versus within-training redundancy can distort evaluation differently.
• What they did: DUET uses a relational graph tower plus two convolutional sequence towers on frozen RNA language models, with training/validation filtered to remove 98 of 388 records at 80% identity and retain 110 identical records while isolating 115 homology components across folds.
• The result: DUET’s MCC is statistically indistinguishable from GRASSP on external sets (0.246 vs 0.243) and improves with real 3D topology (+0.037 over shuffled edges; +0.042 over predicted secondary structure) and larger training pools (+0.135 MCC from 80 to 290 records).

AI summary of the abstract · may contain errors · not medical advice · read the paper.