A novel classification framework for genome-wide association study of whole brain MRI images using deep learning

Abstract

Genome-wide association studies (GWASs) have been widely applied in the neuroimaging field to discover genetic variants associated with brain-related traits. So far, almost all GWASs conducted in neuroimaging genetics are performed on univariate quantitative features summarized from brain images. On the other hand, powerful deep learning technologies have dramatically improved our ability to classify images. In this study, we proposed and implemented a novel machine learning strategy for systematically identifying genetic variants that lead to detectable nuances on Magnetic Resonance Images (MRI). For a specific single nucleotide polymorphism (SNP), if MRI images labeled by genotypes of this SNP can be reliably distinguished using machine learning, we then hypothesized that this SNP is likely to be associated with brain anatomy or function which is manifested in MRI brain images. We applied this strategy to a catalog of MRI image and genotype data collected by the Alzheimer's Disease Neuroimaging Initiative (ADNI) consortium. From the results, we identified novel variants that show strong association to brain phenotypes.

Fig 1

Illustration of the proposed classification-based GWAS framework (A) Schematic representation of the proposed classification-based GWAS framework for conducting association study on Magnetic Resonance (MR) images, where full-frame brain MR images from different directions—axial, coronal, and sagittal, serve as input and SNP genotypes treated as labels. Classification was carried out by convolutional neural network (CNN) models. (B) Detailed illustration of the classification-based framework for detecting the associations between brain MR images and SNPs. In the training stage, true labels are assumed known. Yellow label indicates class 1, blue label indicates class 2. In the testing stage, true labels are only used at the end to evaluate performance. Image with yellow frame indicates the image is predicted to be in class 1, image with blue frame indicates the image is predicted to be in class 2. Green check mark indicates a correct prediction, red “x mark” indicates an incorrect prediction.

Fig 2

Simulation study and data splitting. (A) Schematic illustration of the simulation study, where classification and testing are performed on the same dataset. Data are generated from two distinct univariate normal distributions with the same variance but different means. (B) Comparing p-values of two-sided independent t-test with classification performance metrics Matthew’s Correlation Coefficient (MCC) in the simulation study. (C) Illustration of evaluation process of the classification task in which all images are randomly divided into training, validation and testing sets by subjects in the ratio of 7:1:2.

Fig 3

Results from sex classification using 2D full-frame MR images of the brain (A) Sample 2D full-frame brain MR images from three different planes—axial, coronal, and sagittal. (B) Side-by-side boxplots showing MCC values from sex classification in axial, coronal, and sagittal planes before and after randomly permuting the sex labels. (C) Summary statistics and p-values from the fine tuning step of sex classification using 2D full-frame MR images of the brain. The male and female icons used in this figure were obtained from a free online source (https://www.iconfinder.com/icons/7787571/female_woman_avatar_man_icon, https://www.iconfinder.com/icons/9034992/male_icon), and can be used under MIT license or Creative Commons (Attribution 3.0 Unported, https://creativecommons.org/licenses/by/3.0/).

Fig 4. Main results from classification-based GWAS applied to Alzheimer’s Disease Neuroimaging Initiative (ADNI) MR image data.

(A) Histograms of all MCC values derived from the GWASs conducted in each of the axial, coronal, and sagittal planes. (B) Illustration of the difference in the side-by-side boxplots between top-ranked SNPs and randomly selected low-ranked SNPs obtained from the GWAS conducted on three different planes—axial, coronal, and sagittal. The boxplot shows MCC values for classifying genotypes of the SNP, before and after randomly permuting the genotype labels. (C) Venn diagram displaying the number of overlapping SNPs among the three lists of top 500 SNPs ranked by MCC in the in the axial, coronal, and sagittal planes.

Fig 5. Exploring biological properties of genes close to the top SNPs.

(A) A word cloud showing the major domains of phenotypes show up in PheWAS analyses of the 500 top-ranked SNPs identified from three different planes (from top to bottom): axial, coronal and sagittal. (B) The heatmap showing the enrichment levels (measured by -log transformed p-values) of GO terms and pathways from five lists of top-ranked genes. The first three lists correspond to top genes from our studies of classifying full-frame brain MR images from three different planes—axial, coronal, and sagittal. The fourth set was sourced from the GWAS catalog, the fifth one was acquired from the Oxford Brain Imaging Genetics (BIG) Server. Top 10 most enriched pathways identified from each of the five gene lists were used to generate this heatmap. A grey color indicates that the p-value is not significant (p > 0.05). (C) A heatmap comparing the enrichment (measured by -log transformed p-values) of the expression levels of the same five lists of top-ranked genes from a panel of 16 tissues using the TissueEnrich tool. (D) Venn diagram showing overlaps among top SNPs from three different sources: proposed classification-based GWAS, Brain Measurement (GWAS Catalog, 1762 SNPs) and the Oxford BIG (4323 SNPs). (E) Gene Ontology (GO) term and pathway enrichment analyses results from the GWASs conducted in each of the axial, coronal, and sagittal planes.

Fig 6. Saliency maps for the sex classification task and for the rs11845184 genotype classification task in the axial plane.

The leftmost column shows the saliency map for male individuals, the middle column for female individuals, and the rightmost column for the difference between the two. The saliency maps in the first column depict the regions that are most important for the model to identify a male individual, while the second column illustrates the regions that are most important for the model to identify a female individual. The third column highlights the regions that are the most different between the two sexes, providing insight into the specific brain regions that are most differentiable.