DUMMEST shrinks frameshift-aware alignment: 2–12× faster than BATH, better specificity on simulated reads
Synopsis
Long-read DNA often carries frameshift errors, and aligning it to protein family profiles has meant either translating six reading frames or correcting errors first, both of which cost sensitivity. DUMMEST follows DUMMER's length-bias-free alignment idea and adds nucleotide-emitting indel states to the match, insert and delete states, approximating frameshifts with combinations of nucleotides and codons, with GPU-accelerated MMseqs2 multistage prefiltering. On simulated long reads it detects more alignments with greater average length, matches BATH in sensitivity, improves specificity under simulated sequencing artifacts, and runs 2–12× faster depending on dataset.
Interpretation
DUMMEST aligns raw, frameshift-prone DNA long reads directly to protein family profile hidden Markov models, without six-frame translation or error correction first. Earlier approaches either translated sequences across six reading frames and ignored frameshifts, or relied on slow and unreliable error correction; BATH models frameshifts natively but emits quasi-codons of length 1–5, 1364 symbols per position versus 20 for a conventional protein model. On simulated long reads, DUMMEST detects more alignments with greater average alignment length than non-frameshift-aware tools; against BATH it achieves comparable sensitivity and alignment lengths and improved specificity under simulated sequencing artifacts.
The model writes frameshifts into the state structure itself rather than into an expanded symbol table. Each profile position keeps one match state and three insertion and three deletion states: the match state emits a codon, the insertion states emit a codon, one nucleotide and two nucleotides, and the deletion states skip a codon, emit two nucleotides and one nucleotide, covering insertions and deletions that are not multiples of three. On simulated metagenomic long reads at an E-value threshold of 10⁻⁵, BATH found more hits (42,280 versus 39,447) but also more decoy false positives; DUMMEST produced a single background hit with an E-value of 4.3×10⁻⁶, which overlaps a valid alignment but overextends.
Speed comes from two places: the simpler model and GPU-accelerated MMseqs2 prefiltering. Filtering runs in three stages: an ungapped filter on the six-frame translation, a gapped Smith–Waterman filter, and a backward-only frameshift-aware filter on extracted substrings, with the full forward-backward computation reserved for the most promising fragments. On 8 threads, 32 GB RAM and an RTX 5070 Mobile GPU, it is about 11.7× faster on the transmark dataset and about 2.4× on the wastewater dataset; without the prefilter the core alignment is still about 2.9× faster than BATH.
Perspective
The demonstrated setting is simulated nanopore long reads (356,582 reads at 5× depth) and 2.3 million real long reads from wastewater influent, the latter polished with Illumina short reads so that non-frameshift-aware tools could be compared; real data have no ground truth, so performance is measured by two proxies, unique (read id, protein family) pairs and cumulative aligned base pairs.
E-value reliability was checked only on two randomly chosen Pfam profiles with random sequences, where the authors report a scale of λ≈1 consistent with the conjecture of Yu and Hwa; frameshift parameters are currently approximated from the Badread nanopore2020 error model, and the authors suggest estimating them from the query sequence instead. The authors also note that no studies have compared null models for frameshift-aware alignment, and that DUMMEST has high memory requirements and relies on large temporary files to pass filtered results.