Hybrid mechanistic-machine learning models in biosciences: causality, forecasting, and lab-to-field translation

Biological and environmental systems rarely present a clean choice between first-principles equations and black-box learning. More often, core mechanisms such as disease progression, conservation relations, reaction stoichiometry, or transport direction are known, while key drivers such as behavior, regulation, unresolved forcing, field transport, and measurement processes remain latent. This review examines hybrid mechanistic-machine learning models as a disciplined response to this partial-knowledge regime. We trace their development from early bioprocess hybrids to scientific machine learning, organize current methods into a practical taxonomy of coupling strategies, and synthesize applications in infectious disease forecasting, environmental methane monitoring, hydrology, systems biology, and bioprocess engineering. Across these domains, successful hybrids share a clear division of labor: mechanistic cores retain interpretable states, constraints, and intervention pathways, whereas learning modules are reserved for unresolved closures, latent drivers, observation maps, and site-specific corrections. Particular attention is given to lab-to-field translation, where controlled experiments can identify source mechanisms but field measurements are shaped by transport, aggregation, and sensor context. Stylized examples illustrate how hybrids can separate causal drivers from observed outcomes and connect latent source dynamics to field observations; these examples are conceptual demonstrations rather than empirical benchmarks. The review concludes with design principles, evaluation criteria, and future directions for building hybrid models that are predictive, interpretable, transportable, and useful for scientific decision-making.