PROBABILISTIC NETWORK ALIGNMENT APPLIED TO BRAIN CONNECTOMES

The network alignment problem, which seeks to find correspondence mappings between two or more graphs, is a widely studied open problem. Recent advances in connectomics have created a demand for tools to compare these datasets, since identifying mappings between neurons across brains is a challenging and non-trivial task. Current approaches to network alignment mainly focus on pairwise alignment, which, often fails to capture the complexity of multiple networks. In our work, we address this limitation with a probabilistic method based on Bayesian inference. We assume the existence of an underlying blueprint, L, for the connectome graph of a given species such that the brain of that species is a noisy copy of the underlying blueprint L with differences in the node labels. In this probabilistic framework, the problem of finding the best alignment reduces to maximizing a proposed posterior probability, derived from Baye's theorem. Our method has been effectively applied to both weighted and unweighted graphs, achieving improvements over the current state-of-the-art for the alignment of four C.elegans connectomes with 224 neurons,and the two hemispheres of the larval D.melanogaster with 1,235 neurons (https://www.nature.com/articles/s41467-025-59077-7). Additionally, the interpretability of the Bayesian framework allows us for adapting the model to various scenarios, such as neurons with group annotations or networks with varying numbers of neurons. These advancements are particularly promising for real-world data with larger brains where some neurons may be missing from brain to brain due to experimental limitations or biological variation across individuals.