Student: Abdulah Fawaz
The aim of this project is to develop new methods for intelligent feature selection, that will allow advanced joint modelling of vast imaging and genomics data sets, for the purpose of determining new gene candidates for targeted neuroprotective therapy of vulnerable preterm infants. Imaging genetics is an emerging field that has huge potential to improve understanding of complex neurological conditions, through identifying concrete links between morphological or functional changes in the brain and genetic variants linked to disease. Unfortunately, the immense numbers of imaging and genetic features involved challenge current methods of analysis, limiting their capacity to find statistically significant relationships from the data. This project will therefore use advanced techniques from sparse predictive modelling and Deep Learning to intelligently compress and combine state-of-the-art multi-modality, developmental imaging and genomics data sets so as to propose sensitive genotype-phenotype candidates as targets for future clinical trials.
Preterm births are the commonest cause of loss of disability adjusted life-years in children under five years; about one third have neurocognitive impairments, and the population faces a 7-fold increase in risk for serious psychiatric illness. The causes of these conditions are complex and multifaceted (having both genetic and environmental causes). However, candidate genes for neuroprotective therapy can be identified through first determining how genetic variants, linked to these conditions, impact brain morphology and function.
This style of approach has recently proved highly successful, discovering an unexpected role for the genes DLG4 (PSD95) and PPARg in abnormal preterm brain development (1,2), leading to a planned clinical trial of PPARg agonist drugs as therapy for vulnerable preterm infants. evertheless, it is likely more genes are implicated in the mechanisms of these complex neurological conditions. Unfortunately, current methods of analysis lack the sensitivity to detect them due to the vast numbers of genes and considerable space of multimodality imaging features to be searched.
This project therefore seeks to detect candidate genes and associated imaging phenotypes (linked to autistic spectrum disorder, attention deficit disorder and other challenging neurological conditions) through the development of advanced methods for intelligent compression and sparse predictive modelling of imaging genomics data.
Two approaches for image compression will be explored, including the mapping of an advanced cyto-architecturally accurate atlas of the adult human cerebral cortex (3) (using state-of-the-art machine and/or Deep Learning techniques for propagation of regional labels), as well as the use of convolutional auto-encoders, to learn sensitive latent descriptions of the image domain without requirement for hand engineering of features of prior hypotheses; such methods have previously been used to compelling effect for segmentation tasks (4).
Following identification of candidate image features, data-driven clustering of genomics data will be performed to derive sets of candidate genes (5). Then, the relationship between candidate imaging phenotypes and gene sets will be learnt through advanced sparse predictive modelling (5,6) and/or end-to-end Deep Learning.
The project will take advantage of state-of-the-art collections of multi-modality neuroimaging and genomics data sets, collected within the department as part of the Developing Human Connectome Project (dHCP) and ePrime studies (PI Professor Edwards, 7). These have pushed the boundaries of image acquisition and analysis to acquire large numbers of high spatial and temporal resolution functional, microstructural and morphological Magnetic Resonance Imaging (MRI) of neonatal and fetal brains, collected alongside clinical information, outcome data and Genome-Wide Association Studies (GWAS).
Use of this unique data set in combination with methodological advances proposed in this project will present unique opportunities for the exploration the relative impacts of genetics and the environment on the development of cognition. In this way, new imaging phenotypes and target genes will be identified for focus of future clinical trials for vulnerable preterm infants.
1. Krishnan ML, Van Steenwinckel J, Schang AL, Yan J, Arnadottir J, Le Charpentier T, Csaba Z, Dournaud P, Cipriani S, Auvynet C, Titomanlio L, Pansiot J, Ball G, Boardman JP, Walley AJ, Saxena A, Mirza G, Fleiss B, Edwards AD, Petretto E, Gressens P. Integrative genomics of microglia implicates DLG4 (PSD95) in the white matter development of preterm infants. Nat Commun. 2017 Sep 5;8(1):428. doi: 10.1038/s41467-017-00422-w.
2. Krishnan ML, Wang Z, Aljabar P, Ball G, Mirza G, Saxena A, Counsell SJ, Hajnal JV, Montana G, Edwards AD. Machine learning shows association between genetic variability in Peroxisome Proliferator Activated Receptor Gamma and cerebral connectivity in preterm infants. Proc Nat Acad Sci (USA) 2017 ;114(52):13744-13749. doi: 10.1073/pnas.1704907114.
3. Glasser MF, Coalson TS, Robinson EC, et al Nature. 2016 Aug 11;536(7615):171-178. doi: 10.1038/nature18933.
4 Oktay, Ozan, et al. “Anatomically constrained neural networks (ACNN): application to cardiac image enhancement and segmentation.” IEEE transactions on medical imaging (2017).
5 Liu J, Calhoun VD. A review of multivariate analyses in imaging genetics. Frontiers in neuroinformatics. 2014 Mar 26;8:29.
6 Wang H et al. Alzheimer’s Disease Neuroimaging Initiative. Identifying quantitative trait loci via group-sparse multitask regression and feature selection: an imaging genetics study of the ADNI cohort. Bioinformatics. 2011 Dec 6;28(2):229-37.
7 Makropoulos A, Robinson EC, et al. The developing human connectome project: a minimal processing pipeline for neonatal cortical surface reconstruction. Neuroimage (in press)