RNAflash: Fast and Sensitive Structure-aware RNA Homology Detection and Alignment

Abstract RNA homology search plays a crucial role in RNA structure prediction, facing a trade-off between sensitivity and speed. Here we present RNAflash, a two-stage framework that combines RNA language model-based retrieval with structure-informed reranking and alignment. A retrieval bi-encoder RNAflash-R trained through contrastive learning first retrieves candidate RNA sequences from a large pre-encoded database. Then, its cross-encoder RNAflash-A reranks and aligns the retrieved RNA candidates via jointly trained prediction heads, using predicted secondary structures to refine nucleotide correspondence scores. Across held-out Rfam families, RNAflash retrieved RNA homologs more sensitively than sequence-based methods, and produced alignments comparable to those of structure-based methods. In large-scale searches of RNAcentral, RNAflash generated multiple sequence alignments (MSAs) with stronger coupling and covariation signals than sequence-based and Infernal pipelines, while performing search and MSA construction 3.7-fold faster than non-iterative Riboseek and over 100-fold faster than Infernal and RNAcmap. The lightweight variant RNAflash-L achieves sub-second average search times per query while producing higher-quality MSAs than nhmmer and non-iterative Riboseek. The results show that RNAflash makes sensitive RNA homology search and MSA construction more scalable, and provides an efficient source of MSAs for downstream RNA structure prediction.