DUMMEST: A simpler and faster model for frameshift-aware alignment of a DNA sequence to a protein family

Motivation: We address the problem of aligning DNA sequences containing frameshift errors to protein profile hidden Markov models (pHMMs), a task relevant to annotating protein-coding regions in long-read genomic, metagenomic, and transcriptomic sequence data. Existing approaches either translate sequences across six reading frames, ignoring frameshifts, or rely on slow and unreliable error correction. BATH natively models frameshifts but uses a more complex architecture than conventional pHMMs. Results: We propose a simpler frameshift-aware pHMM based on DUMMER, incorporating modern heuristics, including GPU-accelerated MMseqs2 prefiltering. Evaluations on simulated and real long reads show that our method detects more alignments, with greater average alignment lengths, than non-frameshift-aware tools. Compared with BATH, our method achieves comparable sensitivity and alignment lengths, improved specificity under simulated sequencing artifacts, and is 2--12x faster depending on dataset. Availability and Implementation: Source code and container image are freely available for download at: https://github.com/bioinfodlsu/dummest