DUET: a sequence-only ensemble for RNA-small molecule binding-site prediction

Abstract

Sequence-only predictors can score RNA-small-molecule binding sites without the need to obtain experimental RNA structures, but reported performance can be confounded by test-homologous training records and by inconsistent treatment of within-training redundancy. We introduce DUET, a sequence-only ensemble, and evaluate it with a test-clean protocol that controls these effects separately. Test-homologous records are removed, whereas sequence-identical training records are grouped into homology components and size-balanced to prevent family dominance. Across 10 paired seeds, DUET reached 0.240 {+/-} 0.013 MCC, compared with 0.192 {+/-} 0.031 for CoBRA-M3 trained on the same data (Delta = +0.047). Increasing the training pool from 80 to 290 records raised DUET by +0.135 MCC. DUET also reached 0.246 MCC on two independent SMARTPocket sets, supporting transport beyond the benchmark source. DUET is implemented in Python and PyTorch. Source code, manifests, processed graphs, trained checkpoints and scripts that regenerate the results are available at https://github.com/xiong-0212/DUET and archived at Zenodo (10.5281/zenodo.23635381).