The Krackencoder: a multi-modal connectome fusion tool for brain-behavior mapping - The human brain is a highly complex system of interconnected neurons, from which arise thought, emotion, and behavior. Neuroimaging via MRI can capture structural connectivity (SC) and functional connectivity (FC) net- works to allow insight into how behavior arises from the brain, or so-called brain-behavior mapping. Our group and others have shown that measures of the brain’s SC and FC networks and their inter-relationship can eluci- date brain-behavior relationships in health and disease/recovery. However, there is an urgent, unmet need for multi-modal connectome models that preserve inter-individual differences across the lifespan in health and dis- ease. Without such models, we cannot perform accurate brain-behavior mapping that allows probing of circuitry underlying behavior. Our long-term goal is to create computational tools for integrating metrics of and predicting behavior from brain structure and function, so we can probe circuitry underlying behavior. This project’s over- arching objective is to create and validate tools that integrate and estimate multi-modal connectome data across many individuals, with and without diagnoses, in a behaviorally relevant way. Our central hypothesis is our Kra- kencoder, which fuses and maps between multi-modal connectomes, will be more accurate and better preserve inter-individual differences compared to existing techniques, allowing more robust brain-behavior modeling. We hypothesize the Krakencoder can be used with our Network Modification (NeMo) Tool to estimate FC and SC from clinical MRI in lesioned individuals, to quantify their impact on the connectomes and probe circuitry under- lying behavior. Our hypothesis is supported by data showing i) the Krakencoder accurately maps between multi- modal connectomes in lifespan data and in lesion cohorts and ii) the NeMo Tool’s estimated SC and the Kra- kencoder’s latent space can more accurately perform brain-behavior mapping compared to other connectome metrics. Our rationale is that having an accurate model that preserves individual differences and integrates multi- modal connectomes will enable more robust brain-behavior models for use in mapping the brain’s circuitry un- derlying behavior. We will test our hypotheses via three specific aims: 1) expand the Krakencoder’s ability to accurately map between connectome flavors while preserving inter-individual differences in data from across the lifespan, 2) integrate the Network Modification Tool and the Krakencoder to estimate multi-modal connectomes and 3) probe multi-modal connectome circuitry associated with demographics and behavior. We will use brain MRI, cognitive, and behavioral data from 11 sources (total N>8000), including from healthy young adults, devel- oping and aging populations (HCP lifespan and ABCD), as well as individuals with lesions. The approach is innovative in that our work focuses on preserving inter-individual differences when combining and mapping be- tween multi-modal connectome flavors; this will increase SNR and result in more accurate brain-behavior mod- els. The proposed research is significant in that understanding of how brain anatomy and physiology relate and give rise to behavior could pave the way for novel, personalized interventions to support brain health.
