Open-world fungal ITS embeddings improve higher-rank placement and cross-view retrieval, but percent identity is stronger on a matched ITS2 novelty benchmark

Authors

O'Brien, A., Gardette, A.

Abstract

1. Background. Fungal internal transcribed spacer (ITS) analyses usually infer taxonomy and novelty from similarity to homologous reference sequences. Percent identity is a strong signal in that setting, but it cannot directly compare non-homologous barcode views and does not explicitly optimize the higher-rank placement of genera absent from the reference collection. We asked whether a compact sequence encoder could learn that open-world structure and whether maximum embedding cosine similarity could improve novelty detection. 2. Design. From the UNITE dynamic release of 19 February 2025 we built ITS-core, ITS1 and ITS2 views and a genus-separated development, calibration and test design. A convolutional encoder mapped sequences to 256-dimensional L2-normalized embeddings. A controlled ladder compared genus-proxy supervision (M0), cross-view invariance (M1), hierarchical taxonomic geometry (M2), leave-one-genus family episodes (M3), and the combined objective (M4). M4 was frozen before CAL_KNOWN and TEST were opened. A secondary historical ITS2 benchmark was audited for leakage and scored only after a fixed-recipe, benchmark-safe retraining that excluded all held-out benchmark taxa. 3. Results. Cross-view supervision raised controlled ITS2 development AUROC from 0.688 to 0.717 and novel-family placement from 22.1% to 33.0%. Episodic training increased novel-family placement to 41.0% but reduced novelty discrimination at some views. The combined M4 model recovered both behaviours, reaching development AUROCs of 0.778, 0.755 and 0.751 for ITS-core, ITS1 and ITS2, with novel-family placement of 57.5%, 48.7% and 41.3%. On untouched TEST data, ITS2 reached AUROC 0.763 (95% CI 0.748-0.779) and placed unseen genera correctly at family, order and class for 52.0%, 81.1% and 92.9% of queries; development AUROCs lay above the TEST intervals at ITS-core and ITS1, indicating selection optimism of roughly 0.02-0.04 AUROC at those two views. At alpha = 0.05, conformal calibration gave a 4.7% observed false-novelty rate and detected 11.2% of novel-genus ITS2 queries. On the paired leakage-safe historical benchmark, percent identity remained stronger than M4 cosine: AUROC 0.770 versus 0.702 and 27.8% versus 8.6% novel-query detection at approximately a 5% false-novelty rate. 4. Interpretation. Structured fungal ITS embeddings add higher-rank placement and cross-view retrieval that are not available from a same-locus identity score alone. Maximum embedding cosine does not replace direct sequence identity as the novelty statistic when homologous ITS2 references are available; whether a different score read from the same embedding would narrow that gap remains untested.

Preprint server: bioRxiv
The authors list and abstract were imported from bioRxiv on 24 Sep 2026.

Stats

1-terrible, 9-excellent. How would you rate this preprint? Sign in in to submit your rating.

Post a comment

You need to be signed in to post comments. You can sign in here.

Comments

There are no comments yet.