Xuan Hou, Yuxin Liu, Bohan Li, Jiahui Jiang, Bing Feng, Ting Lei, Ran Yin, Hongqiang Ren
Cyanobacterial harmful algal blooms threaten freshwater security because biomass proxies do not reliably indicate cyanotoxin concentrations. This critical review examines how genetic, physiological, ecological and hydrodynamic processes generate biomass-toxicity decoupling and evaluates approaches for predicting direct cyanotoxin endpoints. The review distinguishes bloom/biomass forecasting, toxin occurrence or threshold-risk prediction, and quantitative prediction of total, phase-specific and congener-specific toxin concentrations. It synthesizes evidence from mechanistic, statistical, molecular, remote-sensing and machine-learning studies according to endpoint definition, predictor availability, forecast horizon and validation design. Reported short-horizon forecasts have achieved threshold accuracies of approximately 85-94%. Leave-one-location-out testing of models based on passive-sampling data from 10 lakes and 12 monitoring locations yielded AUCs of 0.54-0.96, indicating substantial variation in cross-location transferability. Molecular measurements improve toxin specificity but provide upstream signals rather than direct toxin concentrations, and their useful lead time is not universal. Biomass and remote-sensing models are therefore treated as enabling evidence unless linked to independently validated toxin relationships. Explainable artificial intelligence (XAI) supports model auditing but does not establish causality, while direct cyanotoxin-specific physics-informed machine learning (PIML) evidence remains limited. The review proposes a site-specific, uncertainty-aware workflow linking ecological observations, toxin forecasting, direct chemical verification, plant-specific phase-aware treatment evaluation and operator-supervised feedback. Future progress depends on harmonized toxin-specific data, latency-aware multimodal integration, validated transfer under domain shift and toxicity-resolved prediction.