III:Small: Representing and integrating genomics and pathomics data for phenotypic predictive models

University Of Connecticut, Storrs CT

Investigators

Abstract

This project is dedicated to advancing the development of novel algorithms and computational methods that aim to achieve precise disease prediction through the efficient and effective integration of diverse molecular data types and histology images. Recent advances in sequencing technologies and bioinformatics have facilitated the availability of various types of genomics data and have significantly enhanced our understanding of the relationship between genomes and diseases. Additionally, the digitization of histopathology slides, has transformed pathology by enabling the identification and analysis of complex patterns and structures within tissue samples through computational methods. This project establishes a robust computational foundation by leveraging advanced artificial intelligence (AI) methods, especially deep learning, to integrate multiple types of molecular and pathology data. The integration of these data opens up new avenues for understanding the underlying mechanisms associated with diseases, thereby allowing for personalized prognosis, diagnosis, and treatments. The outcomes of this project contribute significantly to advancing biological sciences and improving human health by providing a means to address complex questions about the pathology and molecular mechanisms of various diseases. Furthermore, it contributes to education and training in the high-demand and multidisciplinary fields of bioinformatics, computational genomics, and medical image analysis. The integrative analysis of multimodal omics data poses challenges owing to its inherent complexity, encompassing issues like noise, nonlinearity, heterogeneity, correlation, high dimensionality, and multimodality. Traditional predictive methods struggle with these challenges, and while deep learning methods hold promise for handling complex data, they face limitations due to the aforementioned challenges and the relatively small sample sizes of omics data. To extend beyond conventional learning models, this project builds a precise data-driven phenotypic predictive model by developing novel methods within advanced deep learning frameworks to integrate multimodal omics data and leverage prior biological knowledge. This project significantly advances knowledge in extracting accurate information from complex omics data by: i) introducing an attention-based graph neural network learning algorithm to represent and integrate multimodal genomics data while incorporating prior biological knowledge; ii) developing a Transformer-based multiple instant learning algorithm to identify discriminative regions of pathology images using unannotated training images, and adapting graph neural network methodology to aggregate information from discriminative regions to represent pathology images; and iii) implementing contrastive learning for aligning features from multimodal omics data and their fusion for phenotype prediction while preserving the unique data distribution of each omic and prioritizing discriminative features and omics. 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.