Machine Learning and Model Inference for Biological Physicists

Standard physics modeling approaches rely on symmetries and conservation laws to determine the relevant degrees of freedom and their dynamical rules. This method is challenged in biological physics where, for example, neither proteins nor genes can be identified through symmetries, but nonetheless play a significant role in an organism’s behavior. Instead, across biology, data-driven methods are now being used to unravel the physical laws underlying complex living systems. These approaches often take the form of inverse problems, aiming to infer underlying stochastic equations of motion or other statistical mechanics descriptions. Examples of systems where these approaches are applied range from chromosome organization and interacting proteins to cell motility, and from postural dynamics in animals, to the movement of interacting swarms of fish, insects or birds. These inverse problems are notoriously hard, posing exciting challenges for theory. Such data-driven theory can now capitalize on the recent surge in high-quality data coming from rapid advances in quantitative experiments across biology. An example is given by the explosion of available genome sequences, allowing data-driven modeling of biological sequence data. Other example is in the broad area of morphogenesis where a complex interplay of genes, protein and biomechanical fields can be elucidated with neural networks thanks to rapid advances in imaging and optogenetics techniques.

Modern statistical inference and machine learning (ML) methods can extract generalizable rules from such large datasets and learn to approximate existing underlying functional relationships. For instance, in the case of protein structure determination, pairwise maximum entropy models brought substantial advances, and deep learning models based on natural language processing methods contributed to recent breakthroughs.