Cross-architecture ensembling of DNA foundation models improves the precision and stability of chimera detection in long-read metagenomic bins

Abstract

MotivationChimeric metagenome-assembled genomes (MAGs) that pool DNA from multiple organisms contaminate downstream analyses. Marker-gene tools such as CheckM2 miss low-level chimerism, and DNA foundation models have been proposed as a sequence-composition alternative, but whether large autoregressive models (Evo2, 7B parameters) outperform smaller contrastive models (DNABERT-S, 117M) has not been rigorously tested. ResultsOn 131 MAGs from CAMI2 Nanopore (21 samples, 32 true chimeras), an Evo2 embedding-distance detector achieved high recall (0.84) but low precision (0.34, F1 0.49 [95% CI 0.37-0.60]), producing 52 false positives. Systematic diagnosis revealed convergent multi-factor bias: cross-sample same-species redundancy (50%), low coverage (76% FP at 43.96 OR DNABERT-S > 1.80) achieved in-sample F1 0.65 [0.49-0.79] with 7 false positives, and held-out test F1 0.57 with lowest cross-validation variance ({sigma}=0.05 vs 0.09-0.11 for single models). Ensemble improvement over single models is most robust in precision (0.73 vs 0.34, non-overlapping CIs), while F1 gains are marginal given sample size. On real Nanopore data (ZymoBIOMICS D6331), the tuned threshold transferred without false positives, though well-defined mock communities bin cleanly and reveal a strain-level detection floor for sequence-embedding methods. Training objective and cross-architecture complementarity matter more than parameter scale. Availabilityhttps://github.com/sunsungkim04-sys/evo2-mag. Contact2023024947@knu.ac.kr