Parsing the Functional Impact of Noncoding Genetic Variants in the Brain Epigenome

Abstract

The heritability of common psychiatric disorders has motivated global efforts to identify risk-associated genetic variants and elucidate molecular pathways connecting DNA sequence to disease-associated brain dysfunction. The overrepresentation of risk variants among gene regulatory loci instead of protein-coding loci, however, poses a unique challenge in discerning which among the many thousands of variants identified contribute functionally to disease etiology. Defined broadly, psychiatric epigenomics seeks to understand the effects of disease-associated genetic variation on functional readouts of chromatin in an effort to prioritize variants in terms of their impact on gene expression in the brain. Here, we provide an overview of epigenomic mapping in the human brain and highlight findings of particular relevance to psychiatric genetics. Computational methods, including convolutional neuronal networks, and other machine learning approaches hold great promise for elucidating the functional impact of both common and rare genetic variants, thereby refining the epigenomic architecture of psychiatric disorders and enabling integrative analyses of regulatory noncoding variants in the context of large population-level genome and phenome databases.

Keywords: Artificial intelligence; Chromatin; Convolutional neuronal networks; Epigenome; Prefrontal cortex; PsychENCODE.

Figure 1:. Epigenomic Exploration of the Human Brain in Context of Psychiatric Disease

(left) highly schematic illustration of (top to bottom) GWAS locus with risk-associated polymorphism located in non-coding, putative regulatory region, and epigenomic determinants for gene expression of function, scaling from single base DNA modifications to kilo- and megabase spanning higher order chromatin, as indicated. (right, top) highly schematic illustration of population-scale genotype-phenotype resources including, for example, Psychiatric Genomics Consortium (PGC) and the UK Biobank collecting information on a larger number of subjects in 10⁵ – 10⁷ range. (right bottom) Postmortem brain studies as representative examples for the field. Number of assays in the postmortem brains are shown and are many orders of magnitude smaller as compared to population-scale resources. ROS/MAP = Religious Order Study and Memory and Aging Project. PFC = prefrontal cortex; ACC = anterior cingulate cortex; TC = temporal cortex; CBC = cerebellar cortex. Epigenomic and transcriptomic exploration of the prefrontal cortex. Note that number of brains profiled is many orders of magnitude smaller as compared to population-scale resources. Number of brains shown for specific cohorts as representive examples for the field:(Hannon et al. (23); Jaffe et al. (24); ROS/MAP, Religious Orders Study and Rush Memory and Aging Project, Ng et al. 2017(101); PsychENCODE, Wang et al. (4). Figure was created with BioRender.com.

Figure 2:. Approach of Deep Learning Techniques to Generate Data-Prediction Models

(A - Neural network training workflow) Before the neural network begins training, the sum total of data used to train and evaluate the network is divided in three parts through two successive holdouts. In the test holdout, a random selection of the input data—roughly 10–20% depending on the size of the data set—is set aside as test data, which will be used in the final evaluation of the trained model. The remaining data is referred to as the training data and is itself subsequently divided in two parts during the validation holdout. The validation data provides a measure of the network’s accuracy after each round of training (or epoch). In the training loop, the network generates predictions from input training data and refines the importance (or weights) assigned to input features based on its error rate (or loss) through a process known as backpropagation. After a set number of cycles, a trained model is produced from the network and its accuracy determined by using previously withheld test data. (B - Prediction-Annotation comparison) During training and testing, the network or model takes as input some number of data points (x_i) previously paired with annotations (y_i) and is tasked with predicting $({\widehat{\mathrm{y}}}_{\mathrm{i}})$ the original annotations for the input data. Both accuracy and loss are calculated from the discrepancy therein. The loss function is a measure of the quantitative distance between predicted and expected data and may take many forms depending on the task the network has been assigned. The function displayed is a measure of mean-squared error and would be implemented for a regression task. Loss is then used to inform the backpropagation algorithm, which adjusts the weights of each neuron in the network beginning with the output layer and working backwards. (C - Applications of Deep Learning in Psychiatric Genomics) Deep learning approaches in psychiatric genomics seek to generate predictive models that accurately predict function-based annotations from input datasets that lack those annotations. After adequate training, deep learning-generated models can predict disease-relevant functional annotations without additional resource-intensive experiments being performed. A = PMID 30013180; B = 30247488; C = https://www.biorxiv.org/content/10.1101/103614v3; D = 26873929