Pretrained deep-learning ITS classifiers read the flanking regions, not the ITS2 barcode, and so fail on the amplicon that environmental fungal surveys sequence

Deep-learning classifiers for the fungal internal transcribed spacer (ITS) report accuracies above 90% and are increasingly proposed for environmental metabarcoding. We benchmark two pretrained models, a convolutional network and a transformer sharing a training corpus of 5.23M sequences, against two established k-mer methods on 5,222 identical queries, one per genus, evaluated on both full-length ITS and the ITS2 subregion that dominates environmental sequencing. The design favours the classifiers: queries are drawn from the same public dataset they were trained on and stratified by whether a querys genus lies in their own label space, recovered from the distributed models, while the reference the k-mer methods consult mirrors that label space and excludes the queries themselves. Even so, on full-length ITS both classifiers are outperformed by both classical methods at every rank and in both strata: SINTAX recovers the correct family for 92.0% of seen-genus and 70.3% of novel-genus queries and best-hit alignment against a 56,327-sequence reference for 92.3% and 67.2%, against 77.9% and 57.7% for the transformer and 76.5% and 54.3% for the convolutional network. A hierarchical logistic regression on k-mer counts, fitted in ten minutes to 1.07% of the MycoAI training corpus, also exceeds both and places novel genera better than either search method, and refitted on ITS2 it recovers 89.2% of seen-genus families on that amplicon against 88.6% for best-hit alignment, so neither learned classification nor the amplicon is what fails. Restricting the identical records to ITS2 costs the k-mer methods 3.5 and 3.7 percentage points of seen-genus family accuracy but costs the classifiers 49.6 and 58.0, reducing them to 28.3% and 18.5%. An ablation identifies the cause. Grafting each querys unaltered ITS2 between the flanking regions of a donor record from a different phylum returns the donors family for 34.4% of queries against the querys own for 4.2% in the convolutional model, and 63.7% against 0.8% in the transformer, from a baseline of 0.1% where no donor sequence is present. The models therefore read taxonomy principally from the flanking regions rather than from the ITS2 barcode, which explains the collapse and predicts the same failure for any subregion amplicon. Compounding this, on ITS2 the classifiers output probability all but ceases to separate novel from known genera (AUROC 0.541 and 0.503, the latter at chance, against 0.866 for alignment identity and 0.785 for the SINTAX bootstrap), so the failure is not detectable from the models own output. We recommend that reported accuracies for such models specify the amplicon region of the evaluation, state the length distribution of the training corpus, and include a same-query classical baseline.