NSF Award
2430516
With the recent advances in multimodal data acquisition technologies in healthcare, diverse and large volumes of biological omics, imaging and physiological data are collected at an exponentially increasing rate. The availability of this vast amount of data has allowed us to expand our understanding of physiological and pathological processes, enabling the development of novel healthcare solutions. However, this abundance of data also presents major challenges in analysis due to its complexity and volume. Complex networks (graphs) such as gene regulatory networks (GRNs) and functional connectivity networks (FCNs) of the brain have emerged as valuable tools for describing the interactions between biological targets, providing effective analytical frameworks to characterize the inherent complexities of biological data. Aspiring to address these challenges, this project introduces a comprehensive disentangled graph representation learning framework tailored to address the complexities of graph structured data commonly encountered in biomedical applications. Disentangled graph representation learning will benefit a gamut of research areas, including multimodal machine learning, social analytics, biology, health informatics, and infrastructure systems. In particular, it will have an impact in life science and medicine, where large volumes of unpaired data from various domains are common. Disentangling shared and distinct representations across domains can uncover associations between modalities and disease characteristics, refine patient stratification, and ultimately improve treatments by tailoring therapies to each subgroup more effectively. In addition to research, the project will impact the education and training of the next generation engineers and data scientists at all levels at both MSU and EPFL. <br/><br/>The goal of this project is to use principles of disentangled learning to incorporate insights from various domains and modalities and to extract shared and unique representations across different views, such as patients or data modalities. The proposed research is centered around four intertwined thrusts that broadly aim at: (T1) Structured Disentangled Multi-View Graph Inference; (T2) Multi-view Disentangled Graph Representation Learning; (T3) Multimodal Disentangled Learning and (T4) Application to biomedical data. The first thrust will develop a structured disentangled graph learning framework to infer both common and individual graph structures across multi-view datasets from observed data. The second thrust will move from dealing with unknown data domain to handling data across multiple known domains, represented as different layers in the same graph. Various interactions within biological data will be integrated using a multiplex graph representation learning framework. The third thrust will expand upon the framework developed in Thrust 2 by incorporating heterogeneous data from various modalities, including clinical, imaging, and genomic data. Finally, Thrust 4 will apply these methodological advancements to data from two domains: brain connectomics and cancer biology. Both disease and individual level representations will be learned for ultimately paving the way for precision medicine.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
