Window-level seeding cost in long-read alignment: an exact FM-index evaluation of tiered k-mer selection

Seeding dominates the cost of long-read alignment, and several schemes have been proposed to reduce it by adapting seed length to local sequence complexity. The costs of these schemes are usually reported as reductions in cumulative suffix-array interval mass, a quantity that, as we show, bears little relation the work an exact FM-index actually performs. We built an exact C++ FM-index (BWT via libsais, 64-symbol-block sampled occurrence table, LF-mapping locate) over GRCh38 chromosome 22 and measured where seeding cost accrues for 5,000 simulated ONT reads: the backward-search (count) phase requires O(k) steps per seed and costs 12–14 ms per 1,000 reads, whereas the locate phase and all downstream anchor processing scale with the number of occurrences retained by a frequency threshold. Under this honest accounting, window-tiered k-mer selection guided by an error-robust seeding-mappability statistic reduces retained occurrences by 35% (R10.4.1-like error) and 68% (R9.4.1-like error), against 26% for a random-tier null and 31% for the intra-window repetitiveness statistic itself; fixed k = 19 reduces occurrences by 49% and 74%. A (k, w) grid scan shows the window parameter, not k, governs load, and that chain length tracks anchor mass, removing earlier chain-based claims. Mapping accuracy is statistically equivalent across all schemes (TOST, pre-specified margin 1 pp). On real ONT reads, the scheme reduces retained occurrences by 55% (E. coli, 5,000 reads) and 60% (human chr22, 1,079 reads), the latter without any truth information. We release the exact-index implementation, all pipelines, and per-read results.