Linking fine-grained detail to function through information-preserving reduced representations of complex systems

As advances in computational power enable a relentless increase in the size and complexity of the systems studied and engineered via in silico approaches, identifying the components that play key roles in their function represents a growing challenge. From uncovering biologically relevant regions in biomolecular systems by mining high-dimensional simulation datasets to determining task-relevant units in artificial neural networks (ANNs), addressing this cross-disciplinary challenge calls for the development of broadly applicable and automated analysis methods. In this talk, I will present the mapping entropy optimization workflow (MEOW), an approach rooted in coarse-graining theory for identifying simplified representations of complex systems that minimize the loss of statistical information relative to their high-resolution descriptions. I will illustrate the application of MEOW to two distinct settings, proteins and ANNs, in both of which the resulting information-preserving reduced representations were found to capture functionally relevant components. In proteins, these include biologically significant regions such as catalytic and binding sites, with the combination of MEOW and transition path theory further revealing how the relevance of individual residues changes along a biomolecular transition. In ANNs, the approach identifies subsets of hidden neurons yielding subnetworks whose predictive performance suggests MEOW as a promising complementary strategy to existing pruning methods. These results highlight the potential of information-preserving coarse-graining as a general, unsupervised framework for relating fine-grained detail to function in complex systems.