Knowledge Guided Machine Learning in Biology

The study of biological systems holds great promise for understanding the origin and evolution of life and the interplay of biological processes with environmental effects, which influences policy decisions relating to public health and conservation. While the state-of-the-art for understanding biological systems have conventionally relied on numerical or statistical models for making predictions or performing in silico experimentation, these techniques struggle to capture the nonlinear response of many natural systems. On the other hand, machine learning (ML) methods, that are able to extract highly complex and non-linear patterns and models solely from data, are increasingly being considered as promising alternatives to scientific discovery in biological applications. However, black-box ML methods, that are developed and deployed agnostic to underlying scientific theories, face several barriers in understanding real-world biological systems, primarily due to the absence of ML-ready data in biological applications at the scales possible in commercial applications of ML (e.g., on benchmark problems in computer vision and speech recognition). As a result, there is a growing realization in the scientific community to embrace a deeper integration of scientific knowledge with machine learning frameworks, referred to as the paradigm of Knowledge Guided Machine Learning (KGML). While this emerging paradigm has already started to show successful applications in a number of disciplines including fluid dynamics, particle physics, computational chemistry, and climate science, there is a need for concerted efforts to realize the full potential of “KGML in biology,” by integrating complex forms of biological knowledge (available as process-based models, ontologies, rules, heuristics, etc.), with ML methodologies.