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

Sequence-only predictors can score RNAs without experimental structures, but continuously updated training pools can reabsorb benchmark chains. Test overlap and within-training redundancy are often handled together, although they affect evaluation differently. DUET combines a relational graph tower with two convolutional sequence towers built on frozen RNA language models. We removed 98 of 388 training and validation records reaching 80% identity to canonical test chains, retained 110 sequence-identical records, and isolated 115 homology components across folds. Across 10 paired seeds, DUET reached 0.240 +/- 0.013 MCC versus 0.192 +/- 0.031 for CoBRA-M3 on the same data (Delta = +0.047; Holm p = 0.0073), with parallel gains in AUC and AUPRC. DUET and GRASSP were statistically indistinguishable on external sets (0.246 versus 0.243 MCC). Real three-dimensional topology improved MCC by +0.037 over shuffled edges and by +0.042 over predicted secondary structure. Increasing the training pool from 80 to 290 records raised DUET by +0.135 MCC. Benchmark evaluation therefore requires separate controls for test overlap and within-training redundancy.