Run-periodicity features for alignment-free DNA analysis: evaluating incremental predictive value in genomic windows

This study examines explicit alignment-free DNA representations and the loss of run-order information after feature extraction. We tested whether a proposed 29-coordinate run-periodicity feature vector has out-of-sample predictive value beyond both 69 aggregate run features and a nearest simple order-sensitive control comprising 11 exact lag-match rates. The study combined a classification of seven feature families by five retained properties, 1,200 matched synthetic pairs, and five-fold grouped cross-validation on 58,124 non-overlapping 8-kb windows from 129 RefSeq/PGAP accessions. Operational labels were obtained independently from TRF, DUSTMasker, and longdust intervals. Metrics were first computed within each accession containing both classes and then averaged, preventing longer sequences from dominating the final estimate. Adding the run-periodicity vector to the 80-feature control increased mean accession-level average precision by 0.0664 for TRF, 0.0035 for DUSTMasker, and 0.1101 for longdust. The three original 95% accession-bootstrap intervals excluded zero, and the paired permutation tests remained significant after Holm correction. The TRF and longdust effects remained positive at a 2% interval-coverage threshold and under taxid-grouped cross-validation, whereas the small DUSTMasker effect was not stable in these two checks. The contribution is therefore not a general information-theoretic quantity or a universal repeat detector. It is a compact representation whose incremental predictive value survives a direct simple-order comparator for two distinct operational label sources. The run-periodicity vector can supplement a window-level feature description but does not replace specialized interval annotation.