Leakage-Safe ITS1 Identification of Fasciola hepatica and Fasciola gigantica with Reverse-Complement-Invariant CNN Inference.

backgroundAccurate discrimination between Fasciola hepatica and Fasciola gigantica is essential for surveillance and control of fascioliasis in endemic areas. ITS1 (internal transcribed spacer 1) is a widely used molecular marker, but computational approaches to ITS-based identification must account for data quality issues in public repositories and ensure reliable decision support. PURPOSE: We developed a leakage-safe, decision-oriented pipeline for ITS1-based species identification. A curated GenBank dataset of 527 ITS1 sequences was standardized to 1,000 bp, sanitized (A/C/G/T/N only), and clustered to identify near-duplicates. Cluster-aware train/validation/test splits prevented similarity leakage. We implemented ITSformer-RC, a compact reverse-complement-invariant 1D convolutional neural network paired with a lightweight ensemble, and compared it against BLAST and k-mer logistic regression baselines. Reliability evaluation included post-hoc calibration, robustness testing under realistic perturbations (substitutions, indels, masked spans), and conformal prediction for principled abstention.

methodsWe developed a leakage-safe, decision-oriented pipeline for ITS1-based species identification. A curated GenBank dataset of 527 ITS1 sequences was standardized to 1,000 bp, sanitized (A/C/G/T/N only), and clustered to identify near-duplicates. Cluster-aware train/validation/test splits prevented similarity leakage. We implemented ITSformer-RC, a compact reverse-complement-invariant 1D convolutional neural network paired with a lightweight ensemble, and compared it against BLAST and k-mer logistic regression baselines. Reliability evaluation included post-hoc calibration, robustness testing under realistic perturbations (substitutions, indels, masked spans), and conformal prediction for principled abstention.

resultsBoth ITSformer-RC and k-mer logistic regression achieved 100% accuracy (95% Wilson CI 0.953–1.000) on the held-out cluster-aware test split (n=78), with ROC-AUC and average precision of 1.00. Forced BLAST achieved 97.4% accuracy, while strict BLAST achieved 35.9% coverage with 100% conditional accuracy among non-abstention cases. Vector/classwise temperature scaling improved probability calibration (NLL 0.00399, ECE 1.6×10⁻⁸). Under controlled perturbations, the ensemble demonstrated robust performance across substitutions (AURC 0.9981), indels (AURC 1.0000), and masked spans (AURC 0.9902). Interpretability analysis identified five localized ITS1 windows with concentrated species-discriminative signals.

conclusionThe pipeline provides a reproducible, leakage-aware framework for reliable ITS1-based Fasciola identification emphasizing decision readiness and robustness over raw accuracy alone. Both deep learning and simple statistical baselines saturate accuracy on the curated benchmark, indicating strong ITS1 separability under leakage controls. Calibrated probabilities, abstention mechanisms, and robustness characterization support practical diagnostic workflows. The results are presented as internal validation on a small public dataset; external validation across geographic regions, time periods, and laboratories is required before field deployment claims can be made.