Multi-level statistical classification of substance use disorder

ABSTRACT
This application represents our ongoing commitment to developing an innovative and interdisciplinary research
program on the classification of substance use disorders (SUDs). This research is achieved through
quantitative analysis of multidimensional data that combine clinical symptoms and diagnoses, imaging
markers, and genotypes. The team has a PI with expertise in computational science and the development and
implementation of innovative statistical algorithms to understand multidimensional data; a PI with extensive
experience in systems, imaging and addiction neuroscience; and a co-I who has expertise in the genetics of
SUDs. Our previous R01 project employed a sample of ~12,000 individuals aggregated from multiple genetic
studies of alcohol and drug dependence to generate SUD subtypes based on clinical symptoms. Because
clinical manifestations are distal endpoints in the biological pathway, the genetic effects identified are often
weak and inconsistent, and consequently difficult to detect even in large samples. As championed by the NIMH
Research Domain Criteria (RDoC) research, the etiologies of psychiatric disorders, including SUDs, can be
fruitfully characterized by dimensional neural features. This project thus extends our ongoing work to include
imaging neural features in the classification of SUDs. Specifically, we will utilize a large database from the UK
Biobank Project that provides both genetic and multi-modality magnetic resonance imaging (MRI) data.
Building on our work with the US Human Connectome Project, we aim in the current project to integrate
clinical, imaging, and genotype data to investigate the neurobiological substrates of SUD diagnostic labels, and
to derive SUD subtypes that are optimized for gene finding. Methodologically, we replace the classic statistical
analysis that is confirmatory and biased to an a priori hypothesis by an approach that emphasizes pattern
discoveries from big data. Our specific aims are to: (I): identify neuroimaging features that represent robust
markers of addiction and differentiate SUD subtypes that can be confirmed by multi-modality evidence; (II)
employ a novel brain connectivity model, on the basis of graph convolutional neural networks, to identify neural
markers that precisely characterize the differences in structural changes and functional circuits related to
SUDs; and (III) derive an innovative machine learning model to identify highly heritable neurobiological
subtypes of SUDs that facilitate investigation of the genetic basis of addiction. We will focus on alcohol and
nicotine use disorders to demonstrate the conceptual and methodological approaches. We believe that, by
providing a productive conceptual and methodological platform to integrate imaging and genetic data to
understand the etiologies of SUDs, this research is highly responsive to the RFA “Leveraging Big Data Science
to Elucidate the Neural Mechanisms of Addiction and SUD.” The machine learning tools developed for this
project will provide an innovative and reliable foundation to enhance the aggregation and analysis of
multidimensional data, and to meet the diagnostic and predictive challenges in mental health research.

Grant Number: 5R01DA051922-04
NIH Institute/Center: NIH
Principal Investigator: Jinbo Bi