Research advances in microbial source tracking models for pollution source identification in aquatic environments

Microbial Source Tracking (MST) integrates high-throughput sequencing with machine learning to enable accurate identification and quantitative apportionment of pollution sources. To address the lack of systematic model characterization, this review compared seven representative MST models-Naive Bayes, Random Forest, SourceTracker (ST), Fast Expectation-Maximization for Microbial Source Tracking (FEAST), Relative Aitchison Difference (RAD), Ontology-aware Neural Network for Microbial Source Tracking (ONN4MST), and Microbial Source Tracking with Environment Selection (STENSL)-across principles, accuracy, applicability, and computational efficiency. Our analysis reveals a clear evolutionary trajectory from conventional statistical methods (NB/RF/ST) to high-efficiency algorithms (FEAST/RAD), and more recently to deep learning and sparsity-based frameworks (ONN4MST/STENSL). While ST demonstrates robust performance across diverse aquatic systems, with applications in 14 countries, FEAST offers superior sensitivity (detection limit: 0.01%) and computational speed (30-300 × faster than ST). Despite these advances, all models share common limitations, including sensitivity to source library representativeness, sample size sufficiency, and performance variability under seasonal hydrological dynamics. Beyond model comparison, this review systematically diagnoses these key factors affecting tracking accuracy in real-world scenarios and proposes an integrated technical framework to address them. The framework encompasses four interconnected components-sampling optimization, source library quality assessment via Leave-One-Out Cross-Validation (LOOCV) and Identification Accuracy (IA)-sample size curves, parameter tuning with uncertainty quantification, and multi-evidence cross-validation integrating molecular, chemical, and hydrological spatial data. By explicitly linking problem diagnosis with solution design, this framework provides a practical pathway to enhance MST reliability for watershed pollution control and ecological management. Future research should prioritize standardized dynamic source libraries and hybrid modeling algorithms to broaden MST practical application.