Machine learning for high-dimensional data in genetics and neuroscience

点击翻译按钮获取中文摘要

My recent work was a collaborative creation of a catalogue of the genetic basis of human brain structure and function (L. Elliott et al. Nature. 2018). I also co-created new software (Bycroft et al. Nature. 2018) to conduct efficient genome wide association studies (GWAS) on a new generation of large-scale consortia such as UK Biobank, the China Kadoorie Biobank and AllofUs (each with more than 0.5 million subjects). Current methods uncover associations in such data, but modern machine learning (ML) s required to find patterns in the large number of associations found, and also to extend associations to phenotype predictions. This proposal involves four aims designed to advance ML methods for genetics and bioinformatics in application to large-scale GWAS. My first aim is to develop ML techniques that operate on association results from GWAS. Patterns in the associations of GWAS are exploited by current methods to estimate heritability of phenotypes. My work towards this aim will use techniques from ML disciplines of multi-task learning to further analyse these associations, uncovering latent genetic processes involving groups of variants and groups of phenotypes. My second aim is to use ML techniques such as deep learning and Bayesian nonparametrics to predict phenotypes. There is much recent interest in deep learning with convolutional neural networks, but there is no definitive application of deep learning to GWAS. This is due to the lack of context sensitive information in the formulation of GWAS. I will investigate methods for adding context sensitive information to the input layers of deep neural network. This work could have groundbreaking impact in deep learning. My third aim is to conduct a GWAS on brain images at the level of the volumetric pixel (voxel). Recent work in voxelwise GWAS has not produced statistically significant results. By using data from large-scale consortia, and efficient methods for controlling for spatial noise in brain image signals, I will create fine-scale maps of genetic variation in human brain structure. This work will allow greater understanding of brain variation. My fourth aim is to advance methods related to kinship matrices. The kinship matrix is a fundamental unit encoding genetic similarity between pairs of subjects of a study. It is used to control for socioeconomic and geographical variation within study participants, but the computational cost of computing the kinship matrix precludes its' widespread adoption. I will develop new and efficient methods for computing and approximating the kinship matrix on large-scale consortia, allowing more accurate GWAS results. This research program will support the training of highly qualified personnel, and the software developed in this research program will be released under open source licenses. I anticipate that this research program will form a foundation for ML on the new generation of large-scale consortia, with impact in broad areas of neuroscience.