Abstract
Recent advances in human connectome research have led to the development of models that reveal the brain
circuits associated with behaviors or symptoms. The networks that define these circuits yield functional
phenotypes that can be measured in individuals and are unique to each individual. Such work holds
tremendous promise for providing a biological basis for understanding brain function and brain disorders, it
allows us to characterize trajectories of growth, development, and aging, to categorize patients according to
their functional phenotype, ultimately aiding treatment decisions, and predicting outcomes.
Building such connectome based predictive models, involves 3 distinct steps: 1) construction of the
connectivity matrix summarizing the connections across the defined nodes/parcellation; 2) a subsequent
association step linking edge strength to the behavior or clinical symptom of interest; 3) and finally a predictive
model step for validation and to ensure the models generalize and the associations are not spurious. While
many atlases are available, there has been no consensus on which atlas to use to define the nodes in building
the connectome, making the sharing of models and validation across sites difficult. A second, often overlooked
problem, is that the node configuration supporting one behavior may not be the same for a different behavior
due to the functional flexibility in brain organization. Thus, while the parcellation and brain modeling steps
have historically been treated separately, they are not independent and should not be treated as such.
In this work we will develop a joint parcellation/brain-phenotype modeling approach that provides statistically
powerful, analytically robust, and biologically interpretable Bayesian models that are not dependent upon the
choice of the initial atlas. We will validate the models through measures of predictive power, reliability, and
generalizability, and compare to existing state-of-the-art methods. Data for validation will include the healthy
adult data from the human connectome project and a transdiagnostic sample of 450 individuals (after adding
150 subjects in this study) collected at Yale, spanning a range from healthy control subjects to those with
psychiatric illnesses. Normative models for 16 behavioral measures and 6 clinical scores will be developed
and shared with the neuroscience community.
A key aspect of validation and reproducibility in research is the sharing of data and models. The use of
approximately a dozen or so arbitrary atlases in the field inhibits the sharing of models. This work will move
the field forward by improving the methodology of brain-phenotype predictive modeling, identifying the circuits
supporting behavior, without a priori imposition of an arbitrary atlas. The results could advance our
understanding of the brain networks supporting behavior and impact a wide range of psychiatric illnesses.
Facilitating the release of generalized models to the research community will aid in understanding how to use
these methods for assigning treatments and monitoring the response to treatment.