Machine learning assessment of molecular trophic diatom index (mTDI) using planktonic eDNA for water quality in freshwater ecosystems

Water quality indices using diatoms, such as the Trophic Diatom Index (TDI), are commonly used in freshwater environmental monitoring; however, their taxonomic identification is labor-intensive and ambiguous. In the present study, we applied planktonic environmental DNA (eDNA) metabarcoding to the molecular TDI (mTDI) and evaluated its feasibility as a rapid assessment tool compared with microscopic TDI. We collected surface water samples for eDNA metabarcoding and benthic diatom samples for morphological identification from the Han and Nakdong Rivers in Korea. The 18S rDNA metabarcoding determined the planktonic diatom community, and sequence read counts were used to calculate an mTDI. Canonical Correspondence Analysis (CCA) showed that eDNA community has a stronger statistical correlation with environmental factors. However, the predicted water quality exhibited a clear discrepancy between methods; planktonic eDNA–derived mTDI consistently produced lower estimates. Decomposition analysis of contribution revealed that this discrepancy was driven by the over-representation of pollution-tolerant diatoms, particularly Aulacoseira granulata and Cyclotella meneghiniana, in the eDNA dataset. SIMPER analysis identified these species as the principal contributors to the dissimilarity between the eDNA and microscopy diatom communities. Permutation importance was used to interpret (not predict) variable effects in the Random Forest model, with A. granulata and C. meneghiniana showing the highest %IncMSE (13.3% and 11.0%). These findings suggest that benthic microscopic and surface water eDNA samples may offer complementary insights into water quality. However, directly using eDNA read counts is potentially biased, given the structural inconsistency across sample types.