Genetics of deep-learning-derived neuroimaging endophenotypes for Alzheimer's Disease

点击翻译按钮获取中文摘要

Project Summary/Abstract Alzheimer's disease (AD) is a severe neurodegenerative disease affecting aging Americans and generates dra- matic costs and pains on patients, caregivers, and the society. This administrative supplement proposes to develop new deep learning methods that integrate neuroimaging data with genetic information, thereby generating enriched gene-related data representations. There are at least three challenges to accomplish this research objective. Firstly, commonly used image encoders employ deep neural networks to extract high-level features for downstream tasks, but gene-related information can be contained in features at different levels and resolutions in deep neural networks. Secondly, using deep learning to process genetic data has been largely unexplored in literature. It is hard to design an effective encoder to tackle high-dimensional and discrete genetic data based on deep learning methods. Thirdly, there lacks a principled contrastive learning framework to learn from both imaging and genetic data for GWAS pur- poses. In this administrative supplement, we propose a novel trans-modality contrastive learning framework (TM-CL) to address these limitations and then faithfully accomplish our research goal. Our TM-CL contains novel and spe- cially designed imaging and genetic encoders to process brain MRI data and high-dimensional genetic data as well as a novel contrastive learning scheme to learn enriched gene-related information to benefit downstream tasks such as GWAS. Specifically, TM-CL contains a uniquely designed MRI encoder to integrate features at different scales and resolutions. In addition, our MRI encoder contains an attention-based multi-scale global transformation to extract global information from MRI data. Overall, gene-related information contained in MRI data can be largely captured in the representations. We also design a transformer-based genetic encoder for computing genetic representations. As genetic data is high-dimensional and discrete, it is hard to design deep learning based encoders to process such data. Our genetic encoder is proposed based on swin-transformer, where window attention and shifted window attention are designed to perform attention within splitting windows. By doing this, our genetic encoder can aggregate sufficient gene-related information, while largely reducing the computing cost. More importantly, when performing attention in our genetic encoder, the computed attention scores can capture three types of dependencies among different SNPs, As a result, the complicated genetic dependencies of the input genome data can be effectively captured. Finally, we propose a novel trans-modality contrastive learning scheme for integrating imaging data and genetics. Based on the proposed MRI and genetic encoders, we perform mutual information maximization between the MRI representation and genetic representation as the learning objective. Our contrastive framework is able to generate more informa- tive and discriminate gene-related representations, boosting the performance of downstream tasks such as GWAS. By faithfully accomplishing these research goals, we expect to facilitate the discovery of new genes relevant to AD, thereby shedding light on the causes and cures of AD.