#26 · Genomics

Alignment-free k-mer features resolve variant and class differences in genomic sequences

The dominant thread is estimating discriminative genomic features directly from sequence composition. A k-mer random forest reaches 93% accuracy separating pre- and post-Omicron Delta spike sequences and identifies persistent mutations, while PORT-EK-v2 recovers over-represented HIV-1 sequence regions that distinguish DNA from RNA across group M subtypes. What is new in these k-mer studies is that variant-level and class-level features are learned from sequence content without reference-based alignment. The remaining items address adjacent problems rather than this approach: a benchmark of six pangenome-graph tools on E. coli O157:H7 shows that representation strategy and assembly completeness drive graph size, topology and runtime; ITSxRust extracts ITS regions from long-read amplicons using HMMER with partial-chain recovery; RUMINA improves low-frequency SNV detection through UMI-aware deduplication; and Kun-peng provides memory-efficient pan-domain metagenomic classification.