Although high-throughput genomic and imaging technologies have revolutionized clinical diagnosis and research, they have also created a significant challenge for the discovery of causal associations among variables. The resulting datasets are large and multi-modal, in that the measured variables are of different types: continuous (omics, fMRI measurements), binary (SNPs, sex), numerical (age, drug dosage), categorical (family history of disease, tissue of metastasis), and ordinal (tumor stage, smoking).

To address these challenges, Drs. Takis Benos (Pitt) and Clark Glymour (CMU) received a $1.3 million award from the National Library of Medicine (R01LM012087) to develop efficient partially causal graphical models using such large, multi-modal datasets for use in patient classification, biomarker selection, drug effect prediction, or mechanistic study of network perturbations in disease.

Their team is developing new methodologies based on mixed variable partially causal graphical models, which will be tested with both synthetic and real datasets, including clinical information, time-series diagnostic image data, and genomic, genetic, and epigenetic data. They will use two representative case studies to drive their development work: the discovery of pathogenesis and predictive features of metastatic melanoma and of predictive features of autistic spectrum disorder subgroups.

Their integrative analysis of multiple data types will be designed to identify complex associations and causal relations between clinical and other disease-relevant features and help decipher disease mechanisms.

They will also develop a fully documented software package, MGM-Learn (Mixed Graphical Model Learning), for MatLab and R that can be seamlessly incorporated in other algorithms and applications beyond their case studies as well as a non-expert, user-friendly graphical interface.