Knowledge-guided taxon selection aligns eDNA with morphology-based diatom indices

Water quality assessment of rivers still depends on benthic diatom indices. In these schemes, microscopists enumerate living cells and combine the counts with published sensitivity values to produce a score that maps onto a legal water-quality class. The calculation itself is transparent, but the identification step does not scale readily as taxonomic expertise declines. Environmental DNA metabarcoding can supply a community inventory with far less specialist labor. Read counts, however, are a biased proxy for cell abundance, because amplification is uneven, gene copy number varies among taxa, and remnant DNA persists after cells are gone. As a result, entering every assigned sequence into the conventional formula yields scores that drift from the morphological standard and, in waters judged clean by microscopy, tend to assign a poorer class. Here we show that the index equation can itself decide which molecular taxa to retain. Species are removed in turn until the weighted average computed from the remaining reads most closely matches the morphological score, and no auxiliary regression is fitted. Applied to 1407 paired European rbcL samples, this selection raised the squared Pearson correlation (R ²) from 0.569 to 0.700 and reduced the root-mean-square error from 2.45 to 1.60, largely by discarding pollution-tolerant taxa overrepresented in the sequence data. Tree-ensemble models achieved a closer numerical fit (R ² > 0.89), yet their feature attributions cancelled rather than reproduced the arithmetic of the index. The same procedure, run independently on smaller 18S datasets from Europe and Korea, improved on the unfiltered baseline, although the absolute correspondence remained too weak for operational use. The score that remains is therefore the established index, calculated on a shorter molecular species list.