TransferGWAS of T1-weighted brain MRI data from UK Biobank

Alexander Rakowski; Remo Monti; Christoph Lippert

doi:10.1371/journal.pgen.1011332

. 2024 Dec 13;20(12):e1011332. doi: 10.1371/journal.pgen.1011332

TransferGWAS of T1-weighted brain MRI data from UK Biobank

Alexander Rakowski ¹, Remo Monti ^1,², Christoph Lippert ^1,^3,^*

Editor: Xiang Zhou⁴

PMCID: PMC11687923 PMID: 39671448

Abstract

Genome-wide association studies (GWAS) traditionally analyze single traits, e.g., disease diagnoses or biomarkers. Nowadays, large-scale cohorts such as UK Biobank (UKB) collect imaging data with sample sizes large enough to perform genetic association testing. Typical approaches to GWAS on high-dimensional modalities extract predefined features from the data, e.g., volumes of regions of interest. This limits the scope of such studies to predefined traits and can ignore novel patterns present in the data. TransferGWAS employs deep neural networks (DNNs) to extract low-dimensional representations of imaging data for GWAS, eliminating the need for predefined biomarkers. Here, we apply transferGWAS on brain MRI data from UKB. We encoded 36, 311 T1-weighted brain magnetic resonance imaging (MRI) scans using DNN models trained on MRI scans from the Alzheimer’s Disease Neuroimaging Initiative, and on natural images from the ImageNet dataset, and performed a multivariate GWAS on the resulting features. We identified 289 independent loci, associated among others with bone density, brain, or cardiovascular traits, and 11 regions having no previously reported associations. We fitted polygenic scores (PGS) of the deep features, which improved predictions of bone mineral density and several other traits in a multi-PGS setting, and computed genetic correlations with selected phenotypes, which pointed to novel links between diffusion MRI traits and type 2 diabetes. Overall, our findings provided evidence that features learned with DNN models can uncover additional heritable variability in the human brain beyond the predefined measures, and link them to a range of non-brain phenotypes.

Author summary

Genome-wide association studies are a popular framework for identifying regions in the genome influencing a trait of interest. At the same time, the growing sample sizes of medical imaging datasets allow for their incorporation into such studies. However, due to high dimensionalities of imaging modalities, association testing cannot be performed directly on the raw data. Instead, one would extract a set of measurements from the images, typically using predefined algorithms, which has several drawbacks—it requires specialized software, which might not be available for new or less popular modalities, and can ignore features in the data, if they have not yet been defined. An alternative approach is to extract the features using pretrained deep neural network models, which are well suited for complex high-dimensional data and have the potential to uncover patterns not easily discoverable by manual human analysis. Here, we extracted deep feature representations of brain MRI scans from UK Biobank, and performed a genome-wide association study on them. Besides identifying genetic regions with previously reported associations with brain phenotypes, we found novel regions, as well as ones related to several different traits, such as bone mineral density or cardiovascular traits.

1 Introduction

Genome-wide association studies (GWAS) of imaging modalities pose a challenge due to the high dimensionality of the data—for example, structural brain MRI data scans are typically comprised of over 7 million voxels. The naive approach of testing for associations with all elements in an image is infeasible due to both the computational cost and the loss of statistical power when adjusting the p-value thresholds to account for multiple testing against all voxels.

One solution is to extract predefined image-derived phenotypes (IDPs) [1–3], which aggregate spatial information from imaging data into single variables, e.g., volumes or intensities of particular brain regions of interest (ROIs), reducing the number of phenotypes to test against. While being interpretable, such analyses require the availability of automated tools for IDP extraction for the modality of interest and are limited to traits defined a priori, potentially preventing novel genetically-driven phenotypes from being discovered.

Another approach considers all voxels for genetic association testing. However, instead of performing N_variants × N_voxels independent univariate tests, all genetic variants and all phenotypes are modelled jointly in order to increase statistical power and decrease the number of computations. vGWAS proposed by [4] is a voxel-wise method which retains only the genetic variant with the lowest p-value per-voxel and estimates the effective number of tests using linkage disequilibrium (LD) information to relax the significance thresholds. The Multivariate Omnibus Statistical Test (MOSTest) [5, 6] additionally accounts for polygenicity, and is able to detect variants with smaller effect sizes. While improving statistical power, both of these approaches require performing computations for each variant-voxel pair. [7] apply a projection into a lower-dimensional space on both the variants and voxels, reducing the computational burden, and allowing to employ permutation testing instead of parametric testing as a means of controlling the familywise error rate.

A recent line of work employed deep learning (DL) to extract imaging features using pretrained DNN models to perform GWAS on, which has been demonstrated to be successful in a range of imaging modalities, including retinal fundus images [8, 9], cardiovascular magnetic resonance (CMR) images [10, 11], or brain MRI scans [12]. This can be seen as a compromise between the IDP and voxel-wise approaches, reducing the dimensionality of the imaging data, while not being limited to a priori defined IDPs.

Here, we employed the transferGWAS method of [8], which consists of 1) training DNN feature extractors on auxilliary datasets 2) extracting features from imaging data in the “target” GWAS dataest using the trained DNN models 3) further reducing the feature dimensionality using principal component analysis (PCA) 4) conducting GWAS on the resulting feature principal components (PCs). We used transferGWAS to perform an imaging GWAS on N = 36, 311 T1-weighted brain MRI scans from UKB. We encoded the brain scans from UKB using DNN models trained on the ImageNet [13], and the Alzheimer’s Disease Neuroimaging Initiative (ADNI) datasets [14], with the former extracting “general” image features, and the latter focusing on brain MRI and dementia-specific ones. We note that a similar study has been performed on the UKB brain MRI data by [12] using the ENDO approach [9], which reserves a subset of the target dataset for training of the DNN model, therefore reducing the sample size available for GWAS. By employing the transferGWAS method we were able to utilize the full sample size for genetic testing, due to the DNN models being trained on auxilliary datasets.

Our GWAS of 20 DNN features identified 289 associated genetic loci. 21% of them were not previously reported in brain studies, and corresponded to traits such as bone mineral density (BMD), body size, or blood cell counts, indicating connections between these phenotypes and brain structure, while 11 had no previously reported phenotypic associations. A number of the discovered regions was not detected in either the IDP [2] or ENDO [12] brain MRI studies. We further conducted downstream analyses using the discovered genetic variants, demonstrating their utility in creating more predictive PGS, and pointing to novel genetic correlations between type 2 diabetes (T2D) and diffusion magnetic resonance imaging (dMRI) traits (see Fig 1 for an overview of our workflow).

Fig 1 — **(a)** A general overview of the study: (1)—we trained 2 DNN models on external datasets of natural images, and of brain MRI scans (2)—encoded brain MRI data from the target datasets and performed GWAS on the DNN-derived phenotypes (3) performed a series of downstream analyses using the learned DNN features and discovered genetic variants. **(b)** Description of each step involved in the complete workflow of our study.

2 Results

2.1 Extraction of deep features from the T1-weighted UK Biobank brain MRI scans

To obtain the low-dimensional feature representations of the brain MRI scans, we encoded each of the 36, 311 MRI samples using two pretrained DNN models (we will refer to them throughout the text as the ImageNet and the ADNI model). To reduce their dimensionality, we applied PCA on the extracted features, retaining the first 10 PCs for each model, resulting in 20 PCs in total. The choice of 10 PCs was motivated by [8], who found that a larger number of PCs did not lead to substantial improvements in term of discovered variants, and by the fact that the 10 PCs explained 81% and 92% of variance in the ADNI and ImageNet models respectively.

2.2 Interpretation of the DNN features

In order to interpret the signal carried by the DNN PCs, we performed a phenome-wide association study (PheWAS) against each PC and 7, 744 UKB phenotypes (S2 Table). We found 2, 408 and 2, 622 significantly associated phenotypes for the ImageNet and ADNI PCs respectively, having p-values below the Bonferroni-corrected threshold of ≈ 6.5 ⋅ 10⁻⁷. Fig 2 shows the percentage of significantly associated traits per category. The top 35 categories with the highest ratio of significant hits contained 17 brain-related categories, with the other ones being bone density, body composition, or blood-related categories. In almost all cases the ADNI PCs were associated with a higher number of distinct phenotypes than the ImageNet PCs.

Fig 2 — For each phenotype category from UK Biobank (UKB) we plot the number of significant associations per model divided by the total number of traits in that category—in case of multiple PCs being associated with a phenotype, we only count them once. Shown are the top 35 phenotype categories with the highest ratio of significant associations.

Furthermore, we analyzed how the features correlate with different brain ROIs, using brain segmentation masks obtained with the Synthseg software [15]. For each brain ROI, we computed the fraction of voxels correlated with a given PC (Fig 3), the ratio of explained variance per ROI (Fig I in S1 Text), and per-voxel heatmaps (Figs E, F, G, H in S1 Text). Most PCs were correlated with voxels of most ROI, although at different ratios. The ImageNet PC were correlated on average with a higher number of voxels than the ADNI ones. In particular, the first two ImageNet PCs were correlated with 99% of all ROI voxels, and explained 30% and 7% of the total variance. These two PCs, as well as several other ImageNet PCs seemed to be encoding the overall scan brightness (Fig F in S1 Text). While the ADNI PCs were explaining a smaller amount the total variance in voxel intensity (8% of total variance compared to 42% of the ImageNet model), they seemed to be focused on particular brain structures—e.g., PC 5 corresponding to the lateral ventricles—rather than the overall scan intensity (Figs G, E in S1 Text).

Fig 3 — Values are computed as the total number of significantly associated voxels in each ROI divided by the total number of voxels in that ROI.

2.3 GWAS results

At the Bonferroni-corrected significance threshold of 2.5 ⋅ 10⁻⁹, we found 4, 665 peak associations for the ImageNet and 5, 291 for the ADNI pretrained models, resulting in 4, 382 and 4, 360 distinct variants for ImageNet and ADNI. The clumping procedure then identified 194 and 165 independent regions for the ImageNet and the ADNI models respectively, with 70 regions being shared between the two models (see Fig A in S1 Text for a per- chromosome comparison). This amounted to 7, 458 distinct variants and 289 distinct regions across all 20 features of both DNN models. Figs 4 and 5 show the Manhattan plots for both models, aggregated over each of the 10 PCs per model. We estimated the heritability of each PC using linkage disequilibrium score regression (LDSC) (Section 4.5), and found all PCs to be significantly heritable (p-values below 10⁻⁸), with the ADNI-pretrained PCs having a mean h² = 0.19, and the ImageNet PCs having a mean h² = 0.13 (Fig 6). The summary statistics for all PCs are made publicly available as a figshare resource under https://doi.org/10.6084/m9.figshare.25933717.v1.

Fig 4 — The horizontal lines mark the initial significance threshold of 5 ⋅ 10⁻⁸ (dashed line) and Bonferroni-corrected threshold of 2.5 ⋅ 10⁻⁹ (solid line). We plot gene names for leading variants in each locus. For visualization purposes we truncate p-values below 10⁻⁴⁰ and plot only the minimal p-values across the 10 features.

Fig 5 — The horizontal lines mark the initial significance threshold of 5 ⋅ 10⁻⁸ (dashed line) and Bonferroni-corrected threshold of 2.5 ⋅ 10⁻⁹ (solid line). We plot gene names for leading variants in each locus. For visualization purposes we truncate p-values below 10⁻⁴⁰ and plot only the minimal p-values across the 10 features.

Fig 6 — Black lines indicate the standard error of the estimates.

2.3.1 GWAS catalog associations

For each independent locus, we queried associations reported in previous GWA studies from the NHGRI-EBI GWAS Catalog [16] (Fig 7). The dominating phenotype categories included BMD-related traits and a range of brain traits, such as cortical thickness, diffusion, or volumes of brain ROI. We note that the ADNI-pretrained features tagged more regions corresponding to brain-related traits, whereas the ImageNet model tagged more regions related to “general” body structure, such as BMD, height, or body mass index (BMI). Overall, out of the 289 independent loci, 62 did not have brain-related associations reported in the catalog.

Among neuropsychiatric disorders with the highest number of distinct regions, 47 were associated with schizophrenia, 37 with neuroticism, 36 with attention deficit hyperactivity disorder, 35 with bipolar disorder, 33 with depression, 32 with Alzheimer’s disease, 30 with autism, 22 with anorexia nervosa and 21 with anxiety.

3 out of the 10 first traits were not directly brain-related: heel bone mineral density (HBMD) (144 regions), total BMD (125 regions), and height (113 regions). The associations between BMD and the brain have been investigated in the context of neurological disorders [17–19], as well as in samples of healthy subjects [20]. [17, 18] reported a correlation between BMD and an early onset of Alzheimer’s disease (AD), as well as with several brain volumes. HBMD is postulated to be a causal factor for multiple sclerosis (MS) through an increased risk of fractures [19]. [20] showed that osteoporosis increases the pace of parenchymal atrophy and ventricular enlargement during aging of healthy individuals.

Another prevalent category were blood-related traits, such as cell counts: white (67), red (32), monocyte (45), neutrophil (41), eosinophil (40), lymphocyte, (25) reticulocyte (23), blood pressure (95) or hypertension (45), or hemoglobin (68). Blood pressure and hypertension are known factors influencing brain morphology, as well as cognitive performance or dementia [21–24], while aenemia is a causal factor for cognitive decline and AD [25, 26].

2.3.2 Loci without previously reported GWAS catalog associations

In total, we found existing phenotypic associations in the GWAS Catalog for 275 regions in studies conducted on the white British population, and 278 regions among all populations, whereas 11 loci had no previously reported associations. 8 of these loci were located inside 6 distinct gene regions: CPED1, WNT16, TSPAN12, RP11–161H23.9, WNT1, and C16orf95. CPED1, WNT16, and TSPAN12 have recently been identified as a region of BMD genes [27]. WNT1 is protein-coding gene with variants associated with BMD or cognitive function [28, 29]. C16orf95 is a protein-coding gene with associations with a range of brain measurements and BMD [30, 31]. RP11–161H23.9 is a differential expression long non-coding RNA gene whose expression was associated with glioblastoma survival [32]. The remaining 3 loci not residing within genes were located in regions identified as enhancers in the Enhancer Atlas [33]. In particular, the loci around 6:156060806 and 7:121093894 had markers for enhancer activity in human dendritic, fetal brain, and neural stem cells.

As a further means of interpreting the novel regions, we computed per-voxels correlations for each leading single-nucleotide polymorphism (SNP) (Fig 8), and the fractions of volume of each brain region correlated with each lead variant (Fig 9). All SNPs were correlated with the left cerebral cortex and cerebrospinal fluid. Most notable was the variant rs111469125 (16:87268090) located inside the C16orf95 gene, being correlated with 19 out of 22 brain regions, in particular with several ventricle structures: the 3rd and 4th ventricles (3% of total voxels), and the left and right lateral ventricles (1% and 1.5% of total voxels). It was also correlated with 3% of the voxels of cerebrospinal fluid, and was the only new variant correlated with the left cerebellum white matter, the left thalamus, and the right caudate. On the other hand, rs546521618 (6:156060806) might be a potential false positive, being the only variant in its loci, and having a minor allele frequency (MAF) below 1.3%

Fig 8 — Plotted are values of the t-statistics of the correlation coefficients between lead variants of each region and each single voxel in the MRI scans. We plot values below the Bonferroni-corrected significance threshold accounting for the total number of voxels tested.

Fig 9 — The values are computed as the total number of voxels in a given brain region significantly correlated with a lead variant, divided by the total number of voxels in that brain region. White cells indicate no voxels being significantly correlated for a given brain region-genetic region pair.

2.3.4 Comparison with previous studies

We performed another GWAS on the UKB sample, splitting it into discovery and replication cohorts (23,604 and 12,709 samples), replicating 1,631 hits over 1,510 unique variants, which amounted to 70 replicated loci. We compared our results with two GWA studies on UKB brain MRI data—the first one using 3,144 brain imaging-derived phenotypes [2] and the second study using 256 DL-based features [12], which yielded 692 and 43 replicated loci respectively. Out of our 70 replicated loci, 9 were not present in the 692 of [2], and 28 were not present in the 43 loci of [12].

2.4 TransferGWAS polygenic scores

Here, we evaluated the potential of variants discovered in our study for downstream prediction of phenotypes using 20 PGS fitted for each of the 20 DNN PCs with the summary statistics from our GWAS. In order to compute the features of the DNN models, imaging data need to be present, which constitutes less than a tenth of all UKB samples. On the other hand, genotyping data were available for all participants. This allowed us to calculate the PGS for all remaining N = 424, 609 white British participants not included in the GWAS sample. The corresponding methods are described in Section 4.4, while the weights of the fitted scores are made publicly available as a figshare resource under https://doi.org/10.6084/m9.figshare.25933663.v1.

2.4.1 PGS PheWAS

To gain insights into which traits the PGS might be predictive of, we performed a PheWAS on the 20 PGS and the 7, 744 UKB phenotypes (S1 Table). Note that while the “raw” DNN PCs can encode both genetic and environmental signals, the PGS should capture only the former, and thus we expected the associations between the phenotypes to differ from the PheWAS performed on the PCs. The total number of significant PC-phenotype associations and the effect sizes were higher for the original PCs than for the PGS: 28, 767 vs. 13, 199 significant associations in total, 2, 928 vs. 1, 537 distinct associated traits, with mean effect sizes of $\bar{β} = 0.08$ vs. $\bar{β} = 0.04$ . We identified 3 potentially interesting groups of associations (Fig C in S1 Text):

traits related to BMD or bone fractures
height, weight, BMI
cardiovascular traits, and blood biomarkers

which we decided to investigate further in a prediction setting.

2.4.2 Predictive performance compared to trait-specific PGS

We tested the utility of our developed PGS by evaluating whether they can improve predictions of phenotypes from UKB over PGS designed specifically for particular traits in a multi-PGS setting [34]. We chose a set of 9 phenotypes based on the PheWAS results and computed their corresponding scores using PGS available in the PGS Catalog [35]. For each phenotype, we then fitted and evaluated two linear models: one fitted using only the trait-specific PGS, and one additionally using our transferGWAS PGS. While there were statistically significant improvements in predictions for 6 out of 9 traits, they yielded arguably small performance increases (∼1.5% of relative improvement), with the exception of predicting HBMD using a (general) BMD PGS, where the relative improvement was over 24% (Table 1). We decided to further investigate the HBMD results. Since the improvement could have been stemming from a lower signal in the dataset of the external PGS, compared to the UKB, we conducted a further comparison within the UKB. We performed a GWAS on HBMD using 19, 909 samples from our GWAS data which had HBMD measurements available, and 16, 404 randomly drawn from the remaining samples of the white British participants, to match the sample size and population of our GWAS, and fitted a new PGS on the resulting summary statistics. We then evaluated the newly-created HBMD PGS with and without our transferGWAS PGS on the remaining UKB data and observed the same relative improvement of 1% in performance (p < 0.001). This indicates that transferGWAS has the potential to identify additional variants for related traits while using the sample size. We hypothesize that this might be due to certain pleiotropic variants having a larger effect on the DNN PCs than on HBMD, and thus being able to be detected with our DL GWAS and not with the HBMD-dedicated GWAS.

Table 1. Comparison of predictive performance of Multi-PGS models on the white British population of UK Biobank (UKB) using only trait-specific polygenic scores (PGS) (2nd column) and including our TransferGWAS PGS (3rd column) for a set of selected phenotypes from UKB, measured with the R² coefficient of determination.

Significant differences are marked with (*). Heel bone mineral density (1) and (2) correspond to results of using PGS for heel bone mineral density, or (general) bone mineral density respectively. Statistical significance was estimated using pairwise permutation tests with 1, 000 permutations.

Trait	PGS Catalog	Our PGS + PGS Catalog	Δ
Height	0.760	0.760	0.00008*
BMI	0.286	0.286	0.000
Heel bone mineral density (1)	0.274	0.277	0.002*
Heel bone mineral density (2)	0.062	0.077	0.015*
Red blood cell count	0.389	0.389	0.000
White blood cell count	0.097	0.099	0.002*
Systolic blood pressure	0.283	0.283	0.0002*
Diastolic blood pressure	0.175	0.176	0.001*
Ventricular rate	-0.001	-0.005	-0.004

Open in a new tab

2.5 Genetic correlations

The results of the PheWAS conducted on the learned PCs led us to a set of traits that we decided to investigate further. In order to analyze the genetic components of the PCs, we computed genetic correlation coefficients between 102 selected traits and each of the 20 PCs (see Section 4.5 for details). 39 traits were significantly genetically correlated, surpassing the Bonferroni-corrected threshold of ≈ 2.5 ⋅ 10⁻⁵. We grouped the traits into 3 groups:

(volumes of) brain ROI (e.g., ventricles, brain stem, cerebrospinal fluid (CSF))
dMRI traits (e.g., fractional anisotropy (FA), orientation dispersion index (ODI))
“general” traits: Height, T2D, BMI, HBMD

Additionally, we tested for correlations with AD, educational attainment, and unipolar depression, finding no significant correlations when corrected for multiple testing (p > 0.001). The significantly associated traits are shown in Fig 10, where we observed several “clusters” of PC-trait associations.

Fig 10 — Cell colors represent the magnitudes and the signs of the estimated genetic correlation coefficients between each PC and phenotype combination.

Several PCs were genetically correlated with volumes of multiple brain ROI. The first two PCs of ImageNet (IMGNET0, IMGNET1) seemed to capture the overall body size, as they were negatively genetically correlated with height and white matter volume, and positively with ventricular ROI and CSF.

PCs ADNI2, ADNI3, and IMGNET4 were genetically associated with volumes of several brain ROI, e.g., cerebral white matter, putamen, or thalamus. ADNI2 and ADNI3 were also genetically associated with volumes of CSF and the lateral ventricle. Interestingly, ADNI2 had a positive genetic correlation both for CSF and the lateral ventricle, as well as for gray and white matter structures, whereas one might expect the ventricular volumes (and thus CSF) to grow with the shrinkage of brain structures.

PCs genetically associated with HBMD seemed to capture different aspects of brain anatomy. IMGNET2 had a negative genetic correlation with HBMD, BMI, and cerebral white matter, but also with multiple ventricular volumes. On the other hand, ADNI8 and IMGNET4 also had negative genetic correlations with HBMD, but positive ones with cerebral white matter.

ADNI0 and ADNI4 were genetically associated with a range of Diffusion MRI traits, as well as with several ventricular ROI. Furthermore, ADNI4 was genetically correlated with HBMD and BMI, and was the only PC genetically associated with T2D, which we further discuss below.

2.5.1 ADNI4 and T2D

BMI was shown to increase the risk of developing T2D [36, 37], as well as being genetically correlated to T2D [36]. The signs of genetic correlations between ADNI4, and BMI and T2D were also matching. ADNI4 was also positively genetically associated with HBMD. T2D patients have been shown to have a higher bone density [38, 39]. Evidence also exists for shared heritability between BMD and T2D, albeit relatively small [40, 41]. As with BMI, the sign of the genetic correlation between ADNI4 and HBMD was positive. Regarding the brain ROI, ADNI4 was positively genetically correlated with volumes of the lateral, 3rd, and 4th ventricles, as well as with the CSF. Ventricular enlargement and increase in CSF are associated with several neurodegenerative diseases, such as AD, MS, or schizophrenia [42, 43]. Several studies showed an association between T2D and volumes of white matter structures (whole brain volume, frontal lobe), gray matter (overall trend in all structures), as well as CSF and ventricular volumes [44]. Furthermore, ADNI4 was genetically correlated with 35 different dMRI traits:

Mean diffusivity (MD) traits. 4 MD traits were positively genetically correlated with ADNI4: fornix, superior serebellar peduncle (both sides), and the superior fronto-occipital fasciculus (left). Positive associations between T2D and MD have been found in observational studies [45, 46].

Fractional anisotropy (FA) traits. FA traits have been found to be negatively correlated with T2D in literature [44–46]. We found 4 traits to be negatively genetically correlated with ADNI4, however the posterior limb of left internal capsule was positively genetically correlated with the PC. The direction of this correlation seemed to be in opposition to the associations found in observational studies [44]. On the other hand, it is postulated to be causal with the same sign for fasting insulin [47], an increase of which is an indicator of T2D. We identified two regions containing shared variants located at Chr2:27766284 and Chr14:91881387. The first region contains missense and intron variants for GCKR gene (ENSG00000133962), a glucokinase regulator, with no previously reported associations for brain phenotypes, missense and intron variants for C2orf16 (ENSG00000221843) and intron variants for ZNF512 (ENSG00000243943) both protein coding genes with association with neurodegenerative diseases, T2D, and blood measurements. The second region contains intron variants for the CCDC88C (ENSG00000015133), a protein coding gene, with associations with glucose metabolism, brain measurements, and neurodegenerative diseases, and CCDC88C-DT (ENSG00000258798), an RNA gene that is a divergent transcript for CCDC88C, with associations with brain measurements and hypertension. The above may be another indicator of a non-trivial relation between FA of limb of internal capsule and T2D, with potentially different shared heritability and environmental effects.

Orientation dispersion index (ODI) traits. 7 ODI traits were positively genetically correlated with ADNI4, while 3 traits were correlated negatively. ODI of white matter tracts was reported to be positively correlated with duration of T2D and with levels of HbA1c, a marker for T2D, while ODI of internal capsule was reported to have a negative correlation [48], which is consistent with 9 out of 10 of our findings. We found a negative genetic correlation for the posterior right corona radiata, which had shared variants in regions located at chr8:119486034 and 11:27465591. The first region has intron variants for SAMD12 (ENSG00000177570), a protein coding gene with associations with brain measurements, MS, bone density, and blood measurements. The second region has intron variants for LGR4 (ENSG00000205213), a protein coding gene with associations with brain measurements, bone density, and body mass traits.

Mode of anisotropy (MO) traits. 5 MO traits were genetically positively correlated with ADNI4, and 6 negatively. Fasting insulin, a marker for T2D was reported to be negatively associated with anterior corona radiata [47]. We found positive genetic correlations between both sides of the posterior and superior corona radiata and ADNI4, with shared variants with T2D located in the region chr2:27766284 for the superior, and in chr8:119486034 for the posterior. The first region contains missense and intron variants for the GCKR and C2orf16 genes, and an intron variant for ZNF512 (see the FA regions), and the second region has intron variants for SAMD12 (see the ODI regions above). The correlations between the other 7 traits are reported in S3 Table.

3 Discussion

Using the transferGWAS approach we performed a GWAS on 20 DNN feature representations of 36, 311 T1-weighted brain MRI scans from the UKB, identifying 289 loci, 11 of them without any previously reported associations, and 72 without any associations for brain-related traits. Similar to the findings of the initial transferGWAS study of retinal fundus images of [8], the features of an ImageNet-trained model were associated with a higher number of loci related to “general” body structure traits, such as BMD or BMI, whereas features from a model trained directly on brain MRI data identified more loci corresponding to brain measurements and neurodegenerative diseases. Overall, features of both DNN models were associated directly, through PheWAS, or genetically, through GWAS-identified loci, with a large number of BMD traits. For example, the ImageNet and ADNI-derived features were significantly associated with over 50% and 70% of phenotypes under the UKB category 125 “Bone size, mineral, and density by DXA”, and with over 120 and 40 distinct loci associated “Total body bone mineral density” in the NHGRI-EBI GWAS Catalog. Detecting these genetic regions in features derived from brain MRI data seems to confirm the connections between BMD and brain measurements, as well as with neurodegenerative diseases previously reported in the literature (as discussed in Section 2.3.1), which we further investigated with an analysis of genetic correlations (Section 2.5), highlighting particular brain ROI genetically associated with BMD. Furthermore, the genetic correlations identified by our study shed more light on the relations between dMRI measurements and T2D, BMI, as well as cardiovascular traits, also reported in several studies (Section 2.5.1). Finally, we demonstrated a practical application of our findings by constructing PGS of our DNN-derived phenotypes, which improved predictions of existing PGS of BMD, white cell blood count, or diastolic blood pressure. In a further analysis, we fitted a PGS directly to HBMD measurements on a UKB sample of the same size as our GWAS and observed the same improvement in performance when augmented with our DNN PGS, indicating that the transferGWAS approach can identify additional variants for a trait of interest, being complementary to conducting a trait-dedicated GWAS.

We demonstrated how transferGWAS can be applied to discover new variants and in turn, lead to better phenotype predictions. However, a drawback of using features of pretrained DNN models as traits of interest is their reduced interpretability compared to predefined phenotypes. While we analyzed both the DNN-derived traits and the discovered loci with a range of techniques (PheWAS, querying the GWAS Catalog, statistical parametric mappings (SPMs)), we highlight the need for further developing apossibly automated pipeline for interpretability of the DNN features, to foster their utility for consecutive research and clinical applications.

4 Materials and methods

4.1 Training of the neural network models

The first model used for feature extraction was trained on 4,480 T1-weighted scans from the ADNI dataset [49]. The network architecture was a 3D convolutional variational autoencoder (VAE) [50], trained in a multi-task manner. The model consisted of 3 sub-networks: an encoder, a decoder, and a prediction head. The 128-dimensional outputs of the encoder network constituted the latent representations of the input data. The first task was the standard VAE objective, i.e., reconstructing the input scans from the latent representations, while regularizing the representations to match a standard normal prior distribution with a Kullback-Leibler divergence (KLD) loss term. The second task was to predict the clinical dementia rating (CDR) from the latent representations. The aim of the VAE objective was to learn general structural features describing an MRI scan, while the prediction task should promote neurodegenerative features associated with the presence of dementia. Additionally, we input the age and sex of each participant into the decoder and prediction networks, forcing the model to learn latent representations invariant to age and sex, and thus potentially increasing the statistical power of the GWAS. We trained the model for 500 epochs with the Adam optimizer [51], with a mini-batch size of 128. The weights of the reconstruction, KLD, and the predictions loss terms were 1, 10⁻⁴ and 10⁻² respectively. For data preprocessing, we skull-stripped each scan using using the HD-BET tool [52], performed a non-linear registration to the MNI152 template with a 1mm³ resolution using the FLIRT and FNIRT commands from the FSL software [53], and finally downsampled the scans to a size of 96×96×96 voxels each.

Following [8], we also employed a 2D ResNet50 [54] model trained on ImageNet, a non-medical dataset of natural images [13]. We used a readily available trained model from the PyTorch library [55]. We selected the 2048-dimensional output of the penultimate layer as the latent features used for the GWAS. Since the model was trained on 2D data, we could not directly extract features from the 3D MRI scans. Instead, for each scan, we computed the features over each single slice across the axial axis and averaged the results into a single vector.

4.2 GWAS

We selected a sample of N = 36, 311 UKB participants who “self-identified as white British and have very similar genetic ancestry based on a principal components analysis of the genotypes” (UKB field 22006). We performed the association testing within the linear mixed model (LMM) framework using the BOLT-LMM software [56]. We adjusted for confounding using age, sex, the identifiers of the genotyping array and UKB assessment center, and the first 10 genetic principal components. We filtered the SNPs with the following criteria: MAF≥0.1%, Hardy-Weinberg Equilibrium with a significance level of 0.001, pairwise LD-pruning with R² = 0.8, and maximum missingness of 10% per SNP and participant, which resulted in 577, 570 directly genotyped SNPs. Including imputed genotype data resulted in 16,472,121 variants in total, on which we performed the GWAS. We clumped the variants into independent loci using the PLINK software [57], with a physical distance threshold of 250kb and a significance threshold of 10⁻⁹ for the index SNPs. We queried the NHGRI-EBI GWAS Catalog [16] using the LDtrait web application [58], with an R² cutoff of 10⁻¹ and a 250kb window.

4.3 PheWAS

We performed the PheWAS on the PCs of both DNN models using the PHESANT software [59], with a P-value threshold of ≈ 6.5 ⋅ 10⁻⁷ from the Bonferroni correction to account for 20 PCs and 7, 744 different phenotypes from UKB, adjusting for age and sex. We note that PheWAS automatically determines the choice of the appropriate regression model (e.g., linear regression for continuous variables or logistic regression for categorical ones) based on the phenotype metadata.

4.4 Polygenic scores

We fitted the DNN PGS and the custom HBMD PGS using the PRScs method [60], with the prspipe software [61, 62]. For the predictive performance comparison, we queried the PGS Catalog [35] API for a list of PGS developed for each of the 9 phenotypes, ignoring scores that used the UKB for development, to avoid data leakage. We then computed scores for the N = 451, 450 participants who were not in our GWAS sample using the PGS Catalog Calculator [63]. For each phenotype, we fitted a baseline linear model using all corresponding trait-specific PGS and covariates (age, sex, UKB assessment center, UKB genotyping batch, all UKB genetic PCs) and another linear model which additionally included our 20 DNN PGS. We used 60% of the data for model fitting and evaluated it on the remaining 40%. We computed P-values for differences between achieved R² scores of the two linear models using permutation tests with 1, 000 permutations, randomly selecting predictions from either model for each test sample in each permutation.

4.5 Genetic correlations

To compute the genetic correlation scores between the PCs and selected traits, we used the LDSC method [64, 65]. We used the provided LD scores precomputed on 1000 Genomes data [66] over HapMap3 [67] SNPs, and used the default values for other parameters of the LDSC. In order to find regions potentially contributing to the genetic correlations between ADNI4, T2D, and dMRI traits (Section 2.5.1), we selected SNPs with a P-value below 0.0001 for which the magnitude of the product of the z-scores between both ADNI4 and T2D, and ADNI4 and a dMRI trait exceeded a threshold of 15. For the dMRI traits, we selected pairs where the sign of the product of the z-scores matched the sign of the genetic correlation with ADNI4. We consider a region a set of variants within 250, 000 base pairs from a “central” variant.

Supporting information

S1 Text. Supporting Material main text.

Fig A: Distribution of the p-values from the GWAS from ADNI and ImageNet model training. Fig B: Venn diagrams of the number of discovered loci on each chromosome. Fig C: Results of the phenome-wide association study between the polygenic scores fitted on the features of the deep learning models (rows) and phenotypes from the UK Biobank dataset. Fig D: Results of the PheWAS between the polygenic scores fitted on the features of the DL models (rows) and phenotypes from the UK Biobank dataset. Fig E: Difference images for the brain MRI scans sorted according to their corresponding values of the principal components (PCs) of the ADNI trained model. Fig F: Difference images for the brain MRI scans sorted according to their corresponding values of the PCs of the ImageNet trained model. Fig G: Statistical parametric mappings (SPMs) for the first 10 principal components of the model trained on the ADNI dataset. Fig H: Statistical parametric mappings (SPMs) for the first 10 principal components of the model trained on the ImageNet dataset. Fig I: Ratio of variance explained in each brain region of interest (ROI) by each principal component of the neural network models features. Table A: Comparison of predictive performance of Multi-PGS models on the full UKB sample. Table B: Pearson correlation coefficients of each of the 10 PCs of the ADNI-trained variational autoencoder (VAE) model and age and sex covariates on the UKB data.

(PDF)

pgen.1011332.s001.pdf^{(3MB, pdf)}

S1 Table. Results of PheWAS on the Polygenic Scores.

(XZ)

pgen.1011332.s002.tar.xz^{(6.2MB, xz)}

S2 Table. Results of PheWAS on principal components of the trained model features.

(TSV)

pgen.1011332.s003.tsv^{(4.4MB, tsv)}

S3 Table. Additional results of the genetic correlations analyses.

(XLSX)

pgen.1011332.s004.xlsx^{(265.7KB, xlsx)}

S4 Table. Description of columns of S1 and S2 Tables.

(XLSX)

pgen.1011332.s005.xlsx^{(5.5KB, xlsx)}

Acknowledgments

This research has been conducted using the UK Biobank Resource under Application Numbers 77717 and 78537.

Data Availability

The PheWAS summary statistics for the phenotypes and polygenic scores, as well as the genetic correlations are available directly as supplementary material. The GWAS summary statistics as well as the weights of the fitted polygenic scores are available at figshare: https://doi.org/10.6084/m9.figshare.25933717.v1 https://doi.org/10.6084/m9.figshare.25933663.v1.

Funding Statement

This work was supported by the European Commission (Grant agreement ID: 101016775 to CL), the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) (via the research unit KI-FOR 5363 – 459422098 to CL), and the HPI Research School on Data Science and Engineering to AR.AR and RM received a salary from the European Commission (Grant agreement ID: 101016775). AR received a scholarship from the HPI Research School on Data Science and Engineering. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

References

1. Elliott LT, Sharp K, Alfaro-Almagro F, Shi S, Miller KL, Douaud G, et al. Genome-wide association studies of brain imaging phenotypes in UK Biobank. Nature. 2018;562(7726):210–216. doi: 10.1038/s41586-018-0571-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
2. Smith SM, Douaud G, Chen W, Hanayik T, Alfaro-Almagro F, Sharp K, et al. An expanded set of genome-wide association studies of brain imaging phenotypes in UK Biobank. Nature Neuroscience. 2021;24(5):737–745. doi: 10.1038/s41593-021-00826-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
3. Pirruccello JP, Bick A, Wang M, Chaffin M, Friedman S, Yao J, et al. Analysis of cardiac magnetic resonance imaging in 36,000 individuals yields genetic insights into dilated cardiomyopathy. Nature Communications. 2020;11(1):2254. doi: 10.1038/s41467-020-15823-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
4. Stein JL, Hua X, Lee S, Ho AJ, Leow AD, Toga AW, et al. Voxelwise genome-wide association study (vGWAS). Neuroimage. 2010;53(3):1160–1174. doi: 10.1016/j.neuroimage.2010.02.032 [DOI] [PMC free article] [PubMed] [Google Scholar]
5. van der Meer D, Frei O, Kaufmann T, Shadrin AA, Devor A, Smeland OB, et al. Understanding the genetic determinants of the brain with MOSTest. Nature Communications. 2020;11(1):3512. doi: 10.1038/s41467-020-17368-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
6. Shadrin AA, Kaufmann T, van der Meer D, Palmer CE, Makowski C, Loughnan R, et al. Vertex-wise multivariate genome-wide association study identifies 780 unique genetic loci associated with cortical morphology. Neuroimage. 2021;244:118603. doi: 10.1016/j.neuroimage.2021.118603 [DOI] [PMC free article] [PubMed] [Google Scholar]
7. Ganjgahi H, Winkler AM, Glahn DC, Blangero J, Donohue B, Kochunov P, et al. Fast and powerful genome wide association of dense genetic data with high dimensional imaging phenotypes. Nature Communications. 2018;9(1):3254. doi: 10.1038/s41467-018-05444-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
8. Kirchler M, Konigorski S, Norden M, Meltendorf C, Kloft M, Schurmann C, et al. transferGWAS: GWAS of images using deep transfer learning. Bioinformatics. 2022;38(14):3621–3628. doi: 10.1093/bioinformatics/btac369 [DOI] [PubMed] [Google Scholar]
9. Xie Z, Zhang T, Kim S, Lu J, Zhang W, Lin CH, et al. iGWAS: image-based genome-wide association of self-supervised deep phenotyping of human medical images. medRxiv. 2022; p. 2022–05. [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Bonazzola R, Ravikumar N, Attar R, Ferrante E, Syeda-Mahmood T, Frangi AF. Image-derived phenotype extraction for genetic discovery via unsupervised deep learning in CMR images. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part V 24. Springer; 2021. p. 699–708.
11.Bonazzola R, Ferrante E, Ravikumar N, Xia Y, Keavney B, Plein S, et al. Unsupervised ensemble-based phenotyping helps enhance the discoverability of genes related to heart morphology. arXiv preprint arXiv:230102916. 2023;. [DOI] [PMC free article] [PubMed]
12. Patel K, Xie Z, Yuan H, Islam SMS, Xie Y, He W, et al. Unsupervised deep representation learning enables phenotype discovery for genetic association studies of brain imaging. Communications Biology. 2024;7(1):414. doi: 10.1038/s42003-024-06096-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
13. Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision. 2015;115(3):211–252. doi: 10.1007/s11263-015-0816-y [DOI] [Google Scholar]
14. Mueller SG, Weiner MW, Thal LJ, Petersen RC, Jack C, Jagust W, et al. The Alzheimer’s disease neuroimaging initiative. Neuroimaging Clinics. 2005;15(4):869–877. doi: 10.1016/j.nic.2005.09.008 [DOI] [PMC free article] [PubMed] [Google Scholar]
15. Billot B, Greve DN, Puonti O, Thielscher A, Van Leemput K, Fischl B, et al. SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis. 2023;86:102789. doi: 10.1016/j.media.2023.102789 [DOI] [PMC free article] [PubMed] [Google Scholar]
16. Buniello A, MacArthur JAL, Cerezo M, Harris LW, Hayhurst JD, Malangone C, et al. The NHGRI-EBI GWAS Catalog of published genome-wide association studies, targeted arrays and summary statistics 2019. Nucleic Acids Research. 2019;47(Database-Issue):D1005–D1012. doi: 10.1093/nar/gky1120 [DOI] [PMC free article] [PubMed] [Google Scholar]
17. Loskutova N, Honea RA, Vidoni ED, Brooks WM, Burns JM. Bone density and brain atrophy in early Alzheimer’s disease. Journal of Alzheimer’s Disease. 2009;18(4):777–785. doi: 10.3233/JAD-2009-1185 [DOI] [PMC free article] [PubMed] [Google Scholar]
18. Loskutova N, Honea RA, Brooks WM, Burns JM. Reduced limbic and hypothalamic volumes correlate with bone density in early Alzheimer’s disease. Journal of Alzheimer’s Disease. 2010;20(1):313–322. doi: 10.3233/JAD-2010-1364 [DOI] [PMC free article] [PubMed] [Google Scholar]
19. Yao Y, Gao F, Wu Y, Zhang X, Xu J, Du H, et al. Mendelian randomization analysis of the causal association of bone mineral density and fracture with multiple sclerosis. Frontiers in Neurology. 2022; p. 2049. doi: 10.3389/fneur.2022.993150 [DOI] [PMC free article] [PubMed] [Google Scholar]
20. Bae IS, Kim JM, Cheong JH, Ryu JI, Han MH. Association between bone mineral density and brain parenchymal atrophy and ventricular enlargement in healthy individuals. Aging (albany NY). 2019;11(19):8217. doi: 10.18632/aging.102316 [DOI] [PMC free article] [PubMed] [Google Scholar]
21. Swan GE, DeCarli C, Miller B, Reed T, Wolf P, Jack L, et al. Association of midlife blood pressure to late-life cognitive decline and brain morphology. Neurology. 1998;51(4):986–993. doi: 10.1212/WNL.51.4.986 [DOI] [PubMed] [Google Scholar]
22. Dufouil C, de Kersaint-Gilly A, Besancon V, Levy C, Auffray E, Brunnereau L, et al. Longitudinal study of blood pressure and white matter hyperintensities: the EVA MRI Cohort. Neurology. 2001;56(7):921–926. doi: 10.1212/WNL.56.7.921 [DOI] [PubMed] [Google Scholar]
23. Firbank MJ, Wiseman RM, Burton EJ, Saxby BK, O’Brien JT, Ford GA. Brain atrophy and white matter hyperintensity change in older adults and relationship to blood pressure: brain atrophy, WMH change and blood pressure. Journal of Neurology. 2007;254:713–721. doi: 10.1007/s00415-006-0238-4 [DOI] [PubMed] [Google Scholar]
24. Qiu C, Winblad B, Fratiglioni L. The age-dependent relation of blood pressure to cognitive function and dementia. The Lancet Neurology. 2005;4(8):487–499. doi: 10.1016/S1474-4422(05)70141-1 [DOI] [PubMed] [Google Scholar]
25. Peters R, Burch L, Warner J, Beckett N, Poulter R, Bulpitt C. Haemoglobin, anaemia, dementia and cognitive decline in the elderly, a systematic review. BMC Geriatrics. 2008;8:18. doi: 10.1186/1471-2318-8-18 [DOI] [PMC free article] [PubMed] [Google Scholar]
26. Winchester LM, Powell J, Lovestone S, Nevado-Holgado AJ. Red blood cell indices and anaemia as causative factors for cognitive function deficits and for Alzheimer’s disease. Genome Medicine. 2018;10(1):1–12. doi: 10.1186/s13073-018-0556-z [DOI] [PMC free article] [PubMed] [Google Scholar]
27. Gómez AE, Addish S, Alvarado K, Boatemaa P, Onyali AC, Ramirez EG, et al. Multiple mechanisms explain genetic effects at the CPED1-WNT16 bone mineral density locus. Current Osteoporosis Reports. 2023;21(2):173–183. doi: 10.1007/s11914-023-00783-w [DOI] [PMC free article] [PubMed] [Google Scholar]
28. Morris JA, Kemp JP, Youlten SE, Laurent L, Logan JG, Chai RC, et al. An atlas of genetic influences on osteoporosis in humans and mice. Nature Genetics. 2019;51(2):258–266. doi: 10.1038/s41588-018-0302-x [DOI] [PMC free article] [PubMed] [Google Scholar]
29. Lee JJ, Wedow R, Okbay A, Kong E, Maghzian O, Zacher M, et al. Gene discovery and polygenic prediction from a genome-wide association study of educational attainment in 1.1 million individuals. Nature Genetics. 2018;50(8):1112–1121. doi: 10.1038/s41588-018-0147-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
30. Van Der Meer D, Kaufmann T, Shadrin AA, Makowski C, Frei O, Roelfs D, et al. The genetic architecture of human cortical folding. Science Advances. 2021;7(51):eabj9446. doi: 10.1126/sciadv.abj9446 [DOI] [PMC free article] [PubMed] [Google Scholar]
31. Greenbaum J, Su KJ, Zhang X, Liu Y, Liu A, Zhao LJ, et al. A multiethnic whole genome sequencing study to identify novel loci for bone mineral density. Human Molecular Genetics. 2022;31(7):1067–1081. doi: 10.1093/hmg/ddab305 [DOI] [PMC free article] [PubMed] [Google Scholar]
32. Sun H, Liu X, Song W, Feng Y. Systemic analysis of lncRNA and miRNA expression profiles with associated ceRNA network in glioblastoma. Oncotarget. 2018;5. [Google Scholar]
33. Gao T, He B, Liu S, Zhu H, Tan K, Qian J. EnhancerAtlas: a resource for enhancer annotation and analysis in 105 human cell/tissue types. Bioinformatics. 2016;32(23):3543–3551. doi: 10.1093/bioinformatics/btw495 [DOI] [PMC free article] [PubMed] [Google Scholar]
34. Krapohl E, Patel H, Newhouse S, Curtis CJ, von Stumm S, Dale PS, et al. Multi-polygenic score approach to trait prediction. Molecular Psychiatry. 2018;23(5):1368–1374. doi: 10.1038/mp.2017.163 [DOI] [PMC free article] [PubMed] [Google Scholar]
35. Lambert SA, Gil L, Jupp S, Ritchie SC, Xu Y, Buniello A, et al. The Polygenic Score Catalog as an open database for reproducibility and systematic evaluation. Nature Genetics. 2021;53(4):420–425. doi: 10.1038/s41588-021-00783-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
36. Carlsson S, Ahlbom A, Lichtenstein P, Andersson T. Shared genetic influence of BMI, physical activity and type 2 diabetes: a twin study. Diabetologia. 2013;56:1031–1035. doi: 10.1007/s00125-013-2859-3 [DOI] [PubMed] [Google Scholar]
37. Gray N, Picone G, Sloan F, Yashkin A. The relationship between BMI and onset of diabetes mellitus and its complications. Southern Medical Journal. 2015;108(1):29. [DOI] [PMC free article] [PubMed] [Google Scholar]
38. Picke AK, Campbell G, Napoli N, Hofbauer LC, Rauner M. Update on the impact of type 2 diabetes mellitus on bone metabolism and material properties. Endocrine Connections. 2019;8(3):R55–R70. doi: 10.1530/EC-18-0456 [DOI] [PMC free article] [PubMed] [Google Scholar]
39. Liu B, Liu J, Pan J, Zhao C, Wang Z, Zhang Q. The association of diabetes status and bone mineral density among US adults: evidence from NHANES 2005–2018. BMC Endocrine Disorders. 2023;23(1):1–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
40. Reppe S, Wang Y, Thompson WK, McEvoy LK, Schork AJ, Zuber V, et al. Genetic sharing with cardiovascular disease risk factors and diabetes reveals novel bone mineral density loci. PloS one. 2015;10(12):e0144531. doi: 10.1371/journal.pone.0144531 [DOI] [PMC free article] [PubMed] [Google Scholar]
41. Yu YH, Steffensen B, Ridker PM, Buring JE, Chasman DI. Candidate loci shared among periodontal disease, diabetes and bone density. Frontiers in Endocrinology. 2023;13:1016373. doi: 10.3389/fendo.2022.1016373 [DOI] [PMC free article] [PubMed] [Google Scholar]
42. Conley A. In: Ventricular Enlargement. New York, NY: Springer New York; 2011. p. 2600–2601. Available from: 10.1007/978-0-387-79948-3_375. [DOI] [Google Scholar]
43. Orellana C, Ferreira D, Muehlboeck J, Mecocci P, Vellas B, Tsolaki M, et al. Measuring global brain atrophy with the brain volume/cerebrospinal fluid index: normative values, cut-offs and clinical associations. Neurodegenerative Diseases. 2016;16(1-2):77–86. doi: 10.1159/000442443 [DOI] [PubMed] [Google Scholar]
44. Moghaddam HS, Sherbaf FG, Aarabi MH. Brain microstructural abnormalities in type 2 diabetes mellitus: A systematic review of diffusion tensor imaging studies. Frontiers in Neuroendocrinology. 2019;55:100782. doi: 10.1016/j.yfrne.2019.100782 [DOI] [PubMed] [Google Scholar]
45. Hsu JL, Chen YL, Leu JG, Jaw FS, Lee CH, Tsai YF, et al. Microstructural white matter abnormalities in type 2 diabetes mellitus: a diffusion tensor imaging study. Neuroimage. 2012;59(2):1098–1105. doi: 10.1016/j.neuroimage.2011.09.041 [DOI] [PubMed] [Google Scholar]
46. Reijmer YD, Brundel M, De Bresser J, Kappelle LJ, Leemans A, Biessels GJ, et al. Microstructural white matter abnormalities and cognitive functioning in type 2 diabetes: a diffusion tensor imaging study. Diabetes care. 2013;36(1):137–144. doi: 10.2337/dc12-0493 [DOI] [PMC free article] [PubMed] [Google Scholar]
47. Dong SS, Yu K, Tang SH, Guo J, Guo Y, Yang TL. Evaluation of the causal associations between brain imaging-derived phenotypes and type 2 diabetes: a bidirectional Mendelian randomization study. medRxiv. 2022; p. 2022–12. [Google Scholar]
48. Alotaibi A, Podlasek A, AlTokhis A, Tench CR, Mohammadi-Nejad AR, Sotiropoulos SN, et al. White matter microstructural alteration in type 2 diabetes: a combined UK Biobank study of diffusion tensor imaging and neurite orientation dispersion and density imaging. medRxiv. 2022; p. 2022–07. [Google Scholar]
49. Petersen RC, Aisen PS, Beckett LA, Donohue MC, Gamst AC, Harvey DJ, et al. Alzheimer’s Disease Neuroimaging Initiative (ADNI): clinical characterization. Neurology. 2010;74(3):201–209. doi: 10.1212/WNL.0b013e3181cb3e25 [DOI] [PMC free article] [PubMed] [Google Scholar]
50.Kingma DP, Welling M. Auto-Encoding Variational Bayes; 2013. Available from: https://arxiv.org/abs/1312.6114.
51.Kingma DP, Ba J. Adam: A Method for Stochastic Optimization; 2014. Available from: https://arxiv.org/abs/1412.6980.
52. Isensee F, Schell M, Pflueger I, Brugnara G, Bonekamp D, Neuberger U, et al. Automated brain extraction of multisequence MRI using artificial neural networks. Human Brain Mapping. 2019;40(17):4952–4964. doi: 10.1002/hbm.24750 [DOI] [PMC free article] [PubMed] [Google Scholar]
53. Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM. Fsl. Neuroimage. 2012;62(2):782–790. doi: 10.1016/j.neuroimage.2011.09.015 [DOI] [PubMed] [Google Scholar]
54.He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; 2016. p. 770–778.
55. Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, et al. PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems. 2019;32. [Google Scholar]
56. Loh PR, Tucker G, Bulik-Sullivan BK, Vilhjalmsson BJ, Finucane HK, Salem RM, et al. Efficient Bayesian mixed-model analysis increases association power in large cohorts. Nature Genetics. 2015;47(3):284–290. doi: 10.1038/ng.3190 [DOI] [PMC free article] [PubMed] [Google Scholar]
57. Purcell S, Neale B, Todd-Brown K, Thomas L, Ferreira MA, Bender D, et al. PLINK: a tool set for whole-genome association and population-based linkage analyses. The American Journal of Human Genetics. 2007;81(3):559–575. doi: 10.1086/519795 [DOI] [PMC free article] [PubMed] [Google Scholar]
58. Machiela MJ, Chanock SJ. LDlink: a web-based application for exploring population-specific haplotype structure and linking correlated alleles of possible functional variants. Bioinformatics. 2015;31(21):3555–3557. doi: 10.1093/bioinformatics/btv402 [DOI] [PMC free article] [PubMed] [Google Scholar]
59. Millard LA, Davies NM, Gaunt TR, Davey Smith G, Tilling K. Software Application Profile: PHESANT: a tool for performing automated phenome scans in UK Biobank. International Journal of Epidemiology. 2017;47(1):29–35. doi: 10.1093/ije/dyx204 [DOI] [PMC free article] [PubMed] [Google Scholar]
60. Ge T, Chen CY, Ni Y, Feng YCA, Smoller JW. Polygenic prediction via Bayesian regression and continuous shrinkage priors. Nature communications. 2019;10(1):1776. doi: 10.1038/s41467-019-09718-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
61. Pain O, Glanville KP, Hagenaars SP, Selzam S, Fürtjes AE, Gaspar HA, et al. Evaluation of polygenic prediction methodology within a reference-standardized framework. PLOS Genetics. 2021;17(5):e1009021. doi: 10.1371/journal.pgen.1009021 [DOI] [PMC free article] [PubMed] [Google Scholar]
62. Monti R, Eick L, Hudjashov G, Läll K, Kanoni S, Wolford BN, et al. Evaluation of polygenic scoring methods in five biobanks reveals greater variability between biobanks than between methods and highlights benefits of ensemble learning. medRxiv. 2023; p. 2023–11. [Google Scholar]
63.Team PC. PGS Catalog Calculator; in preparation. https://github.com/PGScatalog/pgsc_calc.
64. Bulik-Sullivan BK, Loh PR, Finucane HK, Ripke S, Yang J, of the Psychiatric Genomics Consortium SWG, et al. LD Score regression distinguishes confounding from polygenicity in genome-wide association studies. Nature Genetics. 2015;47(3):291–295. doi: 10.1038/ng.3211 [DOI] [PMC free article] [PubMed] [Google Scholar]
65. Bulik-Sullivan B, Finucane HK, Anttila V, Gusev A, Day FR, Loh PR, et al. An atlas of genetic correlations across human diseases and traits. Nature Genetics. 2015;47(11):1236–1241. doi: 10.1038/ng.3406 [DOI] [PMC free article] [PubMed] [Google Scholar]
66. 1000 Genomes Project Consortium and others. A global reference for human genetic variation. Nature. 2015;526(7571):68. doi: 10.1038/nature15393 [DOI] [PMC free article] [PubMed] [Google Scholar]
67. International HapMap 3 Consortium and others. Integrating common and rare genetic variation in diverse human populations. Nature. 2010;467(7311):52. doi: 10.1038/nature09298 [DOI] [PMC free article] [PubMed] [Google Scholar]

PLoS Genet. doi: 10.1371/journal.pgen.1011332.r001

Decision Letter 0

Hua Tang, Xiang Zhou

29 Jul 2024

Dear Dr Rakowski,

Thank you very much for submitting your Methods entitled 'TransferGWAS of T1-weighted Brain MRI Data from the UK Biobank' to PLOS Genetics.

The manuscript was fully evaluated at the editorial level and by independent peer reviewers. The reviewers appreciated the attention to an important problem, but raised some substantial concerns about the current manuscript. Based on the reviews, we will not be able to accept this version of the manuscript, but we would be willing to review a much-revised version. We cannot, of course, promise publication at that time.

Should you decide to revise the manuscript for further consideration here, your revisions should address the specific points made by each reviewer. We will also require a detailed list of your responses to the review comments and a description of the changes you have made in the manuscript.

If you decide to revise the manuscript for further consideration at PLOS Genetics, please aim to resubmit within the next 60 days, unless it will take extra time to address the concerns of the reviewers, in which case we would appreciate an expected resubmission date by email to plosgenetics@plos.org.

If present, accompanying reviewer attachments are included with this email; please notify the journal office if any appear to be missing. They will also be available for download from the link below. You can use this link to log into the system when you are ready to submit a revised version, having first consulted our Submission Checklist.

To enhance the reproducibility of your results, we recommend that you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols

Please be aware that our data availability policy requires that all numerical data underlying graphs or summary statistics are included with the submission, and you will need to provide this upon resubmission if not already present. In addition, we do not permit the inclusion of phrases such as "data not shown" or "unpublished results" in manuscripts. All points should be backed up by data provided with the submission.

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at figures@plos.org.

PLOS has incorporated Similarity Check, powered by iThenticate, into its journal-wide submission system in order to screen submitted content for originality before publication. Each PLOS journal undertakes screening on a proportion of submitted articles. You will be contacted if needed following the screening process.

To resubmit, log into your Editorial Manager account and select the option 'Revise Submission' in the 'Submissions Needing Revision' folder.

We are sorry that we cannot be more positive about your manuscript at this stage. Please do not hesitate to contact us if you have any concerns or questions.

Yours sincerely,

Xiang Zhou, Ph.D.

Academic Editor

PLOS Genetics

Hua Tang

Section Editor

PLOS Genetics

Reviewer's Responses to Questions

Comments to the Authors:

Please note here if the review is uploaded as an attachment.

Reviewer #1: Rakowski and colleagues used a recently developed approach, i.e., TransferGWAS, to identify more genetic loci associated with neuroimaging-derived phenotypes by extracting low-dimensional representations of imaging data for GWAS. Even though it could be a very sensitive approach to identify more genetic variants that have not been discovered in previous algorithms, there is a subtle improvement of polygenic risk scores in the prediction of phenotypes. The statistical approaches used in the current study sound very robust, but it’s still necessary to rewrite part of the results and clarify how this approach works. I have some comments, which could improve the quality of this study.

1. This study lacks too many details to clarify how the authors investigate the genetics of neuroimaging phenotypes, even though Figure 1 shows the workflow of this study. For example, in the Introduction, the authors mentioned: “We encoded the brain scans using 63 models pre-trained on the ImageNet [48] and Alzheimer’s Disease Neuroimaging Initiative (ADNI) 64 datasets”. However, it’s unclear why they selected two datasets to train the data. Another example is that “As opposed to the ENDO approach…..”. It’s unclear how this ENDO approach works. What’s the difference between ENDO and TransferGWAS. Hopefully, the authors could rewrite the Introduction and introduce more background for this study.

2. In the Introduction, the authors mentioned that “Instead of using manually defined traits, a recent line of work employed deep learning (DL) to derive imaging features using pre-trained deep neural network (DNN)”. Are there any references? What does “a recent line of work” refer to?

3. In Section 3.1, the authors mentioned that “In order to interpret the signal carried by the DNN features, we extracted the first 10 principal components (PCs) of both DNN models”. It’s very confusing. What types of phenotypes did the authors use to extract the top 10 components? When did they extract the top 10, instead of the top 20 components? Since the authors put more Method details in the next section, they should introduce more background about how to confirm the top 10 components.

4. In Section 3.2, the authors tested the heritability of each PC and performed GWAS across all components, which is good. Based on Figure 4, I guess some of the PCs might not have significant heritability. So, it does not make sense to scan genetic variants associated with a trait that is not heritable.

5. I guess the authors performed GWAS for each PC. However, Figure 3 only showed two Manhattan plots for traits derived from ImageNet pretraining and ADNI pretraining. Where are the rest 18 plots? Could the authors explain it?

6. In Section 3.2.3, the authors performed another GWAS using discovery and replication cohorts (23,604 and 12,709 samples). Where do these cohorts come from? Did the authors divide the overall sample into two parts? The authors are missing many details to clarify the results. Hopefully, the authors could add more details to the Results, which could help the readers better understand this study.

7. In Section 3.3, the authors calculate the PGS for all remaining N = 451,450 participants not included in the GWAS sample. Because there are not many details in this section, I am not sure whether this cohort (N=451,450) is a European-ancestry sample. Since GWAS is generated from individuals with European ancestry, the PRS should also be generated in European ancestry samples.

8. Figure 6, I think it would be better if the authors showed these gene-brain associations in a 3D manner. The current 2D visualization could not clearly show these interesting findings, at least I could not see any significant brain regions associated with genetic variants.

9. In the Section 5.2, SNPs were filtered with MAF≥ 0.1%. Since the study aims to investigate common genetic variants associated with brain phenotypes, common genetic variants are typically defined as MAF>1%. I would suggest the authors revise this MAF threshold to avoid false positive signals.

10. In Section 3.3.1, the author performed PheWAS to screen for phenotypes associated with PRS. How did the authors perform phenotypic associations? What type of models did the authors use for testing associations? Given that UKB includes different types of variables (e.g., binary and continuous), do they use a single model for all phenotypes? The authors should add more details to clarify that.

Reviewer #2: In "TransferGWAS of T1-weighted Brain MRI Data from the UK Biobank," the authors provide a deep neural network (DNN) based approach to learn a low dimensional representation of the brain, and then they perform genome-wide association study on this low dimensional representation. Their method TransferGWAS is based on training the DNN (to find the low dimensional representation) on a dataset other than UK Biobank, and then applying the low dimensional representation to UK Biobank, finding 289 associated genetic variants, of which they claim 14 to be novel. They also find a novel link between brain and type 2 diabetes.

This work is relevant to the genetics and neuroscience community. Reliance on preprocessing techniques such as SPM and freesurfer to define volumes of regions of interest is a shortcoming of traditional brain imaging genetics work. Defining brain phenotypes in an unsupervised fashion can improve power, which can lead to new associations, which may improve the clinical relevance of brain imaging GWAS, and further our understanding of the brain. Also, when unsupervised learning is performed on the same dataset that the GWAS is done on, there may be some overfitting leading to false positives. Thus, the idea of TransferGWAS wherein the training stage is done in part on another dataset is interesting.

Major comments:

1. The abstract and introduction of this work however do seem to state quite strongly the extent to which deep neural networks, and transfer learning are novel, and the extent to which voxelwise gwas is not done. The authors cite the ENDO approach of Patel et al. 2022. a) Aside from doing pretraining on a different dataset than the study dataset, in what was is TransferGWAS different from Patel et al. 2022? c) Methodologically, is the application of TransferGWAS in this paper the same as that in Kirchler et al. 2022? c) With regards to voxelwise, there are some pieces from early/mid 2010s such as from Thomas Nichols’ group: “Fast and powerful genome wide association of dense genetic data with high dimensional imaging phenotypes” and Stein et al. “Voxelwise genome-wide association study”. Are there really no updates to voxelwise GWAS since these works? And how does TransferGWAS compare to voxelwise?

2. The authors mention that they train the model for 500 epochs with the Adam optimizer. I understand that this was after pretraining on either ImageNet or ADNI (i.e., these are 500 epochs on UK Biobank. Pretraining on ImageNet has been found to improve accuracy of a resulting classifier, however that improvement can be lost if the length of training on the dataset-specific (target-dataset, in this case UK Biobank) is increased in terms of epochs. The question is this: Is the pre-training improving the nature of the local maximum that the backpropagation on the DNN eventually settles on? Or is it just getting us to the same local maximum (or an equivalent local maximum) faster? Can the authors please investigate this by increasing the number of epochs and plotting a measure of accuracy (or number of positives/amount of overlap of hits/or another unbiased measure of performance) per additional 100 epochs, for a few thousand more epochs? This would be greatly helpful to the community to show if TransferGWAS (or in general, any pretraining method) is truly bringing something that cannot be found when we do not pretrain, but train for longer. What are the compute costs? If 500 epochs takes a day, and if with two weeks of compute we can perform better than pre-training, then there is not much argument for pre-training (two weeks is not a long time).

3. It appears that the downsampling performed by TransferGWAS is quite extensive, downsampling to a side length of 96. Do the authors then introduce a brain mask within the 96x96x96 volume, to select only voxels that intersect the brain? This may considerably reduce the number of voxels that are then fed into the neural net, which could allow less extensive downsampling. I cannot find a mention of this in the paper, regrets if I missed it.

4. Can the authors comment on the clinical or scientific interpretation of the 14 claimed novel variants, and the type 2 diabetes link? I understand that this is not so much a methods paper, as TransferGWAS was already described in Kirchler et al. 2022. What scientific hypotheses regarding the mapping between genotypes, brain phenotypes and disease are supported or refuted by this work?

5. While early work on brain imaging in UK Biobank often focused on the white British subset, the sample size is now large enough that meta-analysis or replication analysis with ancestries other than white British can now be done. How generalizable to ancestries other than white British are the results? An answer to this question is essential, in order to prevent bias from sample inclusion criteria.

6. Please refer to inclusion criteria based on ancestry in a consistent way. On line 332 UKB field 22006 was described “… self-identified as ‘White British’ … ” and on line 9 this set is referred to as “British” (and not specifically white) and on line 179 as “white-British”. I would recommend referring to this population always as “white British”.

Minor comments:

1. Please increase the font size of Figure 5, possibly after rotating it 90 degrees clockwise. Increase font size of Figure 6, after tiling in two columns. Greatly increase font size of Figure 3, possibly tile in one column.

2. Please refer to UK Biobank as “from UK Biobank” rather than “from the UK Biobank”. While “the UK Biobank” is prevalent usage, as UK Biobank is a proper noun, omitting a definite article is preferred.

3. Please make the writing tighter in the introduction and author summary, and avoid cliches such as “At the same time, ” “The growing size, ”

4. In the fourth sentence, the em dash should be longer (its length is currently that of a hyphen, not an em dash).

5. The references need formatting attention. For example, “early alzheimer’s disease” -> “early Alzheimer’s disease” and “Nature genetics” -> “Nature Genetics” and “The relationship between bmi and onset …” -> “The relationship between BMI and onset …” and “PLoS genetics” -> “PLOS Genetics”

6. The tables in the supplementary material could do with captions indicating units, and describing in more detail what the columns mean. I struggled to read the supplementary tables.

Reviewer #3: The authors employ a previously published approach, TransferGWAS (by Kirchler et al.), originally applied to retinal fundus photographs, to a new modality: T1-weighted Brain MRI data from UK Biobank. They find novel genetic findings, and show how these can help in building polygenic scores (PGS) with higher predictive performance for some phenotypes.

Regarding interpretability of the embeddings:

- I believe the "Interpretation of DNN Features" Section should be improved and enlarged. While the presence of many significant correlations with different traits is shown in an aggregated way in Fig. 2 and the individual correlations are included as Supplementary Material, the question of what's the impact of the different embeddings in brain morphology is central and should be addressed. I would suggest distilling the conclusions from the correlation study into a couple of additional paragraphs, so that the reader can better understand what each of the 20 PCs is modelling. An additional figure (e.g. another panel in Fig. 2) showing the correlation of some brain structural traits with the PCs could perhaps help to this end. While some correlations seem to be described in Section 3.4 "Genetic correlations", it's not clear if these are genetic correlations (which the section title suggests) or "direct" correlations (as the text suggests). In any case, I believe Section 3.1 would be a better place to include this information.

- It would also be reassuring to determine whether these insights are confirmed by directly examining the impact of the DNN-derived PC embeddings in the image space. For example, in the case of the VAE network, traversals can be performed on the PCs, e.g. from (mean - 3 sd) to (mean + 3 sd), and examine changes in the output of the decoder. This could help localize the changes to some specific region of the brain. The PCA inverse transformation could be pre-appended to the decoder network as a fully connected layer, so that your DL code can be readily used to create these figures.

In the case of the ImageNet, since you have no decoder you could try to build attribution maps by employing explainable AI techniques. For example, if using Pytorch the library Captum might be handy to this end. The average operation can be added as an additional Pytorch module, and Captum could be used on a 3D model built in this way.

- In Fig. 2, please clarify if you are counting a single association whenever you have more than one significant PC-trait association with a given trait. I understand this is the case since you refer to these values as fractions, however the caption seems to suggest otherwise (i.e. "we plot the number of significant associations per model" this reads as though there could be double-counting). Also, could more fine-grained results be presented, showing whether

Regarding the models:

- Is there an expected benefit in training the VAE in a separate dataset instead of directly on the UK Biobank population? Is this done because the ADNI dataset contains a greater number individuals with brain pathologies which can be used to guide the training? Also related to this, I am curious as to whether the authors have assessed the heritability of features that come from a VAE trained simply with the reconstruction task. The conclusion that a benefit is found in using multi-task learning with a prediction head would be really interesting to report.

- I like the idea of inputting demographic data to the decoder to "explain them away" from the embeddings. The correlation of the embeddings (or the PCs) with age and sex can be reported to ensure that this procedure is having the expected effect.

- I wonder if the data matrix for PCA in each of the two experiments, consists of hidden features that are properly scaled. In principle, the neural network training could converge to a set of hidden features that span different ranges. For instance, hidden feature h1 could span the range [-3, 3] and h2 could span [0.1, 0.2], where the amplitude of these ranges is not necessarily meaningful and which would affect the results of PCA. In particular, some hidden feature with a large range can be overshadowing another, more important one, but with a smaller range. It would be reassuring to know this is taken into account somehow, either by scaling the hidden features to a common range or by justifying why this is not actually desirable.

- The choice of 10 as the number of PCs to be considered should also be justified, e.g. in terms of the proportion of variance explained or the absence of GWAS hits.

- It would be interesting to examine how PCs from the two different models relate to each other. In particular, are they linearly identifiable, i.e. can a linear transformation be applied to one set of PCs to obtain the other? Canonical correlation analysis could be helpful to this end. If the answer is no, i.e. if some PCs can't be accurately reconstructed as linear combinations of the other set of PCs, this would provide evidence that the two networks actually extract orthogonal information, further justifying the use of both.

Some additional minor comments:

- In the abstract, I suggest adding a closing statement regarding the possible implications of the findings.

- For Table 1: is a statistical test being used to declare the improvement significant? If so, what's the test? If not, I would discourage the use of the word "significant" since it normally implies a statistical test being employed, at least in my experience.

- Line 337: Usually, a threshold in INFO imputation score is applied (typically 0.3). Is this filtering step being performed here? It looks to be the case based on the number of SNPs, but it would be good to explicitly mention it.

- For Fig. 3, I suggest adding gene names. Given the large number of loci and the central role of these figures, also consider using landscape orientation and splitting the Manhattan plots into two different figures. It would also be informative to add some indication for the GWAS hits of whether the loci are shared between the two models, or whether effect sizes are significantly higher in one case or the other.

- Line 246: and -> an.

**********

Have all data underlying the figures and results presented in the manuscript been provided?

Large-scale datasets should be made available via a public repository as described in the PLOS Genetics data availability policy, and numerical data that underlies graphs or summary statistics should be provided in spreadsheet form as supporting information.

Reviewer #1: Yes

Reviewer #2: No: The summary statistics for the GWAS appear to be not available (only summary statistics available for PHEWAS).

Reviewer #3: None

**********

PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

PLoS Genet. 2024 Dec 13;20(12):e1011332. doi: 10.1371/journal.pgen.1011332.r002

Author response to Decision Letter 0

13 Sep 2024

Attachment

Submitted filename: Brain_MRI_Transfer_GWAS_Response_to_Reviewers.pdf

pgen.1011332.s006.pdf^{(117.5KB, pdf)}

PLoS Genet. doi: 10.1371/journal.pgen.1011332.r003

Decision Letter 1

Hua Tang, Xiang Zhou

7 Nov 2024

Dear Dr Rakowski,

We are pleased to inform you that your manuscript entitled "TransferGWAS of T1-weighted Brain MRI Data from UK Biobank" has been editorially accepted for publication in PLOS Genetics. Congratulations!

Before your submission can be formally accepted and sent to production you will need to complete our formatting changes, which you will receive in a follow up email. Please be aware that it may take several days for you to receive this email; during this time no action is required by you. Please note: the accept date on your published article will reflect the date of this provisional acceptance, but your manuscript will not be scheduled for publication until the required changes have been made.

Once your paper is formally accepted, an uncorrected proof of your manuscript will be published online ahead of the final version, unless you’ve already opted out via the online submission form. If, for any reason, you do not want an earlier version of your manuscript published online or are unsure if you have already indicated as such, please let the journal staff know immediately at plosgenetics@plos.org.

In the meantime, please log into Editorial Manager at https://www.editorialmanager.com/pgenetics/, click the "Update My Information" link at the top of the page, and update your user information to ensure an efficient production and billing process. Note that PLOS requires an ORCID iD for all corresponding authors. Therefore, please ensure that you have an ORCID iD and that it is validated in Editorial Manager. To do this, go to ‘Update my Information’ (in the upper left-hand corner of the main menu), and click on the Fetch/Validate link next to the ORCID field. This will take you to the ORCID site and allow you to create a new iD or authenticate a pre-existing iD in Editorial Manager.

If you have a press-related query, or would like to know about making your underlying data available (as you will be aware, this is required for publication), please see the end of this email. If your institution or institutions have a press office, please notify them about your upcoming article at this point, to enable them to help maximise its impact. Inform journal staff as soon as possible if you are preparing a press release for your article and need a publication date.

Thank you again for supporting open-access publishing; we are looking forward to publishing your work in PLOS Genetics!

Yours sincerely,

Xiang Zhou, Ph.D.

Academic Editor

PLOS Genetics

Hua Tang

Section Editor

PLOS Genetics

Aimée Dudley

Editor-in-Chief

PLOS Genetics

Anne Goriely

Editor-in-Chief

PLOS Genetics

www.plosgenetics.org

Twitter: @PLOSGenetics

----------------------------------------------------

Comments from the reviewers (if applicable):

Reviewer's Responses to Questions

Comments to the Authors:

Please note here if the review is uploaded as an attachment.

Reviewer #1: The authors addressed all of my concerns. I have no comments on this study. I think the current version of the study has improved a lot.

Reviewer #2: The manuscript under review is a revision to "TransferGWAS of T1-weighted Brain MRI Data from UK Biobank." The paper presents a deep learning technique to uncover brain imaging phenotypes. The phenotypes are learned in an unsupervised way, potentially improving on hand-crafted pipelines used to make phenotypes from MRI scans. The manuscript focusses on the transferability of the learned phenotypes. In particular, in the main experiment, the phenotypes are learned on one dataset (Alzheimer’s Disease Neuroimaging Initiative), and then GWAS is performed on another dataset (UKB). This design is crutial to ensure that the phenotypes learned are useful, and not overfitting to the training dataset. This work has the potential to improve brain imaging genetics by reducing the reliance on hand-crafted features.

The authors have made extensive revisions to the initial submission, and improved the manuscript. The revisions include 1) improvements to the clarity, 2) investigation of whether their previous main results were confounded by ancestry, 3) improvements in Figures and clarity of p-values. The author's revisions are all valid, and I think that they cover all of the reviewer's questions. I have no further comments on this manuscript.

Reviewer #3: I thank the authors for properly addressing my comments and those of my fellow reviewers from the previous round of revision.

I believe the manuscript is now in good shape for publication.

I would suggest adding the p-values in Table 1 (or otherwise in the Supplementary Material), since the differences seem too small to be declared significant. This addition would give more reassurance to the readers.

**********

Have all data underlying the figures and results presented in the manuscript been provided?

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

----------------------------------------------------

Data Deposition

If you have submitted a Research Article or Front Matter that has associated data that are not suitable for deposition in a subject-specific public repository (such as GenBank or ArrayExpress), one way to make that data available is to deposit it in the Dryad Digital Repository. As you may recall, we ask all authors to agree to make data available; this is one way to achieve that. A full list of recommended repositories can be found on our website.

The following link will take you to the Dryad record for your article, so you won't have to re‐enter its bibliographic information, and can upload your files directly:

http://datadryad.org/submit?journalID=pgenetics&manu=PGENETICS-D-24-00628R1

More information about depositing data in Dryad is available at http://www.datadryad.org/depositing. If you experience any difficulties in submitting your data, please contact help@datadryad.org for support.

Additionally, please be aware that our data availability policy requires that all numerical data underlying display items are included with the submission, and you will need to provide this before we can formally accept your manuscript, if not already present.

----------------------------------------------------

Press Queries

If you or your institution will be preparing press materials for this manuscript, or if you need to know your paper's publication date for media purposes, please inform the journal staff as soon as possible so that your submission can be scheduled accordingly. Your manuscript will remain under a strict press embargo until the publication date and time. This means an early version of your manuscript will not be published ahead of your final version. PLOS Genetics may also choose to issue a press release for your article. If there's anything the journal should know or you'd like more information, please get in touch via plosgenetics@plos.org.

PLoS Genet. doi: 10.1371/journal.pgen.1011332.r004

Acceptance letter

Hua Tang, Xiang Zhou

5 Dec 2024

PGENETICS-D-24-00628R1

TransferGWAS of T1-weighted Brain MRI Data from UK Biobank

Dear Dr Rakowski,

We are pleased to inform you that your manuscript entitled "TransferGWAS of T1-weighted Brain MRI Data from UK Biobank" has been formally accepted for publication in PLOS Genetics! Your manuscript is now with our production department and you will be notified of the publication date in due course.

The corresponding author will soon be receiving a typeset proof for review, to ensure errors have not been introduced during production. Please review the PDF proof of your manuscript carefully, as this is the last chance to correct any errors. Please note that major changes, or those which affect the scientific understanding of the work, will likely cause delays to the publication date of your manuscript.

Soon after your final files are uploaded, unless you have opted out or your manuscript is a front-matter piece, the early version of your manuscript will be published online. The date of the early version will be your article's publication date. The final article will be published to the same URL, and all versions of the paper will be accessible to readers.

Thank you again for supporting PLOS Genetics and open-access publishing. We are looking forward to publishing your work!

With kind regards,

Anita Estes

PLOS Genetics

On behalf of:

The PLOS Genetics Team

Carlyle House, Carlyle Road, Cambridge CB4 3DN | United Kingdom

plosgenetics@plos.org | +44 (0) 1223-442823

plosgenetics.org | Twitter: @PLOSGenetics

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

S1 Text. Supporting Material main text.

(PDF)

pgen.1011332.s001.pdf^{(3MB, pdf)}

S1 Table. Results of PheWAS on the Polygenic Scores.

(XZ)

pgen.1011332.s002.tar.xz^{(6.2MB, xz)}

S2 Table. Results of PheWAS on principal components of the trained model features.

(TSV)

pgen.1011332.s003.tsv^{(4.4MB, tsv)}

S3 Table. Additional results of the genetic correlations analyses.

(XLSX)

pgen.1011332.s004.xlsx^{(265.7KB, xlsx)}

S4 Table. Description of columns of S1 and S2 Tables.

(XLSX)

pgen.1011332.s005.xlsx^{(5.5KB, xlsx)}

Attachment

Submitted filename: Brain_MRI_Transfer_GWAS_Response_to_Reviewers.pdf

pgen.1011332.s006.pdf^{(117.5KB, pdf)}

Data Availability Statement

[pgen.1011332.ref001] 1. Elliott LT, Sharp K, Alfaro-Almagro F, Shi S, Miller KL, Douaud G, et al. Genome-wide association studies of brain imaging phenotypes in UK Biobank. Nature. 2018;562(7726):210–216. doi: 10.1038/s41586-018-0571-7 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref002] 2. Smith SM, Douaud G, Chen W, Hanayik T, Alfaro-Almagro F, Sharp K, et al. An expanded set of genome-wide association studies of brain imaging phenotypes in UK Biobank. Nature Neuroscience. 2021;24(5):737–745. doi: 10.1038/s41593-021-00826-4 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref003] 3. Pirruccello JP, Bick A, Wang M, Chaffin M, Friedman S, Yao J, et al. Analysis of cardiac magnetic resonance imaging in 36,000 individuals yields genetic insights into dilated cardiomyopathy. Nature Communications. 2020;11(1):2254. doi: 10.1038/s41467-020-15823-7 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref004] 4. Stein JL, Hua X, Lee S, Ho AJ, Leow AD, Toga AW, et al. Voxelwise genome-wide association study (vGWAS). Neuroimage. 2010;53(3):1160–1174. doi: 10.1016/j.neuroimage.2010.02.032 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref005] 5. van der Meer D, Frei O, Kaufmann T, Shadrin AA, Devor A, Smeland OB, et al. Understanding the genetic determinants of the brain with MOSTest. Nature Communications. 2020;11(1):3512. doi: 10.1038/s41467-020-17368-1 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref006] 6. Shadrin AA, Kaufmann T, van der Meer D, Palmer CE, Makowski C, Loughnan R, et al. Vertex-wise multivariate genome-wide association study identifies 780 unique genetic loci associated with cortical morphology. Neuroimage. 2021;244:118603. doi: 10.1016/j.neuroimage.2021.118603 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref007] 7. Ganjgahi H, Winkler AM, Glahn DC, Blangero J, Donohue B, Kochunov P, et al. Fast and powerful genome wide association of dense genetic data with high dimensional imaging phenotypes. Nature Communications. 2018;9(1):3254. doi: 10.1038/s41467-018-05444-6 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref008] 8. Kirchler M, Konigorski S, Norden M, Meltendorf C, Kloft M, Schurmann C, et al. transferGWAS: GWAS of images using deep transfer learning. Bioinformatics. 2022;38(14):3621–3628. doi: 10.1093/bioinformatics/btac369 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref009] 9. Xie Z, Zhang T, Kim S, Lu J, Zhang W, Lin CH, et al. iGWAS: image-based genome-wide association of self-supervised deep phenotyping of human medical images. medRxiv. 2022; p. 2022–05. [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref010] 10.Bonazzola R, Ravikumar N, Attar R, Ferrante E, Syeda-Mahmood T, Frangi AF. Image-derived phenotype extraction for genetic discovery via unsupervised deep learning in CMR images. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part V 24. Springer; 2021. p. 699–708.

[pgen.1011332.ref011] 11.Bonazzola R, Ferrante E, Ravikumar N, Xia Y, Keavney B, Plein S, et al. Unsupervised ensemble-based phenotyping helps enhance the discoverability of genes related to heart morphology. arXiv preprint arXiv:230102916. 2023;. [DOI] [PMC free article] [PubMed]

[pgen.1011332.ref012] 12. Patel K, Xie Z, Yuan H, Islam SMS, Xie Y, He W, et al. Unsupervised deep representation learning enables phenotype discovery for genetic association studies of brain imaging. Communications Biology. 2024;7(1):414. doi: 10.1038/s42003-024-06096-7 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref013] 13. Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision. 2015;115(3):211–252. doi: 10.1007/s11263-015-0816-y [DOI] [Google Scholar]

[pgen.1011332.ref014] 14. Mueller SG, Weiner MW, Thal LJ, Petersen RC, Jack C, Jagust W, et al. The Alzheimer’s disease neuroimaging initiative. Neuroimaging Clinics. 2005;15(4):869–877. doi: 10.1016/j.nic.2005.09.008 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref015] 15. Billot B, Greve DN, Puonti O, Thielscher A, Van Leemput K, Fischl B, et al. SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis. 2023;86:102789. doi: 10.1016/j.media.2023.102789 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref016] 16. Buniello A, MacArthur JAL, Cerezo M, Harris LW, Hayhurst JD, Malangone C, et al. The NHGRI-EBI GWAS Catalog of published genome-wide association studies, targeted arrays and summary statistics 2019. Nucleic Acids Research. 2019;47(Database-Issue):D1005–D1012. doi: 10.1093/nar/gky1120 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref017] 17. Loskutova N, Honea RA, Vidoni ED, Brooks WM, Burns JM. Bone density and brain atrophy in early Alzheimer’s disease. Journal of Alzheimer’s Disease. 2009;18(4):777–785. doi: 10.3233/JAD-2009-1185 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref018] 18. Loskutova N, Honea RA, Brooks WM, Burns JM. Reduced limbic and hypothalamic volumes correlate with bone density in early Alzheimer’s disease. Journal of Alzheimer’s Disease. 2010;20(1):313–322. doi: 10.3233/JAD-2010-1364 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref019] 19. Yao Y, Gao F, Wu Y, Zhang X, Xu J, Du H, et al. Mendelian randomization analysis of the causal association of bone mineral density and fracture with multiple sclerosis. Frontiers in Neurology. 2022; p. 2049. doi: 10.3389/fneur.2022.993150 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref020] 20. Bae IS, Kim JM, Cheong JH, Ryu JI, Han MH. Association between bone mineral density and brain parenchymal atrophy and ventricular enlargement in healthy individuals. Aging (albany NY). 2019;11(19):8217. doi: 10.18632/aging.102316 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref021] 21. Swan GE, DeCarli C, Miller B, Reed T, Wolf P, Jack L, et al. Association of midlife blood pressure to late-life cognitive decline and brain morphology. Neurology. 1998;51(4):986–993. doi: 10.1212/WNL.51.4.986 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref022] 22. Dufouil C, de Kersaint-Gilly A, Besancon V, Levy C, Auffray E, Brunnereau L, et al. Longitudinal study of blood pressure and white matter hyperintensities: the EVA MRI Cohort. Neurology. 2001;56(7):921–926. doi: 10.1212/WNL.56.7.921 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref023] 23. Firbank MJ, Wiseman RM, Burton EJ, Saxby BK, O’Brien JT, Ford GA. Brain atrophy and white matter hyperintensity change in older adults and relationship to blood pressure: brain atrophy, WMH change and blood pressure. Journal of Neurology. 2007;254:713–721. doi: 10.1007/s00415-006-0238-4 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref024] 24. Qiu C, Winblad B, Fratiglioni L. The age-dependent relation of blood pressure to cognitive function and dementia. The Lancet Neurology. 2005;4(8):487–499. doi: 10.1016/S1474-4422(05)70141-1 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref025] 25. Peters R, Burch L, Warner J, Beckett N, Poulter R, Bulpitt C. Haemoglobin, anaemia, dementia and cognitive decline in the elderly, a systematic review. BMC Geriatrics. 2008;8:18. doi: 10.1186/1471-2318-8-18 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref026] 26. Winchester LM, Powell J, Lovestone S, Nevado-Holgado AJ. Red blood cell indices and anaemia as causative factors for cognitive function deficits and for Alzheimer’s disease. Genome Medicine. 2018;10(1):1–12. doi: 10.1186/s13073-018-0556-z [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref027] 27. Gómez AE, Addish S, Alvarado K, Boatemaa P, Onyali AC, Ramirez EG, et al. Multiple mechanisms explain genetic effects at the CPED1-WNT16 bone mineral density locus. Current Osteoporosis Reports. 2023;21(2):173–183. doi: 10.1007/s11914-023-00783-w [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref028] 28. Morris JA, Kemp JP, Youlten SE, Laurent L, Logan JG, Chai RC, et al. An atlas of genetic influences on osteoporosis in humans and mice. Nature Genetics. 2019;51(2):258–266. doi: 10.1038/s41588-018-0302-x [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref029] 29. Lee JJ, Wedow R, Okbay A, Kong E, Maghzian O, Zacher M, et al. Gene discovery and polygenic prediction from a genome-wide association study of educational attainment in 1.1 million individuals. Nature Genetics. 2018;50(8):1112–1121. doi: 10.1038/s41588-018-0147-3 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref030] 30. Van Der Meer D, Kaufmann T, Shadrin AA, Makowski C, Frei O, Roelfs D, et al. The genetic architecture of human cortical folding. Science Advances. 2021;7(51):eabj9446. doi: 10.1126/sciadv.abj9446 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref031] 31. Greenbaum J, Su KJ, Zhang X, Liu Y, Liu A, Zhao LJ, et al. A multiethnic whole genome sequencing study to identify novel loci for bone mineral density. Human Molecular Genetics. 2022;31(7):1067–1081. doi: 10.1093/hmg/ddab305 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref032] 32. Sun H, Liu X, Song W, Feng Y. Systemic analysis of lncRNA and miRNA expression profiles with associated ceRNA network in glioblastoma. Oncotarget. 2018;5. [Google Scholar]

[pgen.1011332.ref033] 33. Gao T, He B, Liu S, Zhu H, Tan K, Qian J. EnhancerAtlas: a resource for enhancer annotation and analysis in 105 human cell/tissue types. Bioinformatics. 2016;32(23):3543–3551. doi: 10.1093/bioinformatics/btw495 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref034] 34. Krapohl E, Patel H, Newhouse S, Curtis CJ, von Stumm S, Dale PS, et al. Multi-polygenic score approach to trait prediction. Molecular Psychiatry. 2018;23(5):1368–1374. doi: 10.1038/mp.2017.163 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref035] 35. Lambert SA, Gil L, Jupp S, Ritchie SC, Xu Y, Buniello A, et al. The Polygenic Score Catalog as an open database for reproducibility and systematic evaluation. Nature Genetics. 2021;53(4):420–425. doi: 10.1038/s41588-021-00783-5 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref036] 36. Carlsson S, Ahlbom A, Lichtenstein P, Andersson T. Shared genetic influence of BMI, physical activity and type 2 diabetes: a twin study. Diabetologia. 2013;56:1031–1035. doi: 10.1007/s00125-013-2859-3 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref037] 37. Gray N, Picone G, Sloan F, Yashkin A. The relationship between BMI and onset of diabetes mellitus and its complications. Southern Medical Journal. 2015;108(1):29. [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref038] 38. Picke AK, Campbell G, Napoli N, Hofbauer LC, Rauner M. Update on the impact of type 2 diabetes mellitus on bone metabolism and material properties. Endocrine Connections. 2019;8(3):R55–R70. doi: 10.1530/EC-18-0456 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref039] 39. Liu B, Liu J, Pan J, Zhao C, Wang Z, Zhang Q. The association of diabetes status and bone mineral density among US adults: evidence from NHANES 2005–2018. BMC Endocrine Disorders. 2023;23(1):1–9. [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref040] 40. Reppe S, Wang Y, Thompson WK, McEvoy LK, Schork AJ, Zuber V, et al. Genetic sharing with cardiovascular disease risk factors and diabetes reveals novel bone mineral density loci. PloS one. 2015;10(12):e0144531. doi: 10.1371/journal.pone.0144531 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref041] 41. Yu YH, Steffensen B, Ridker PM, Buring JE, Chasman DI. Candidate loci shared among periodontal disease, diabetes and bone density. Frontiers in Endocrinology. 2023;13:1016373. doi: 10.3389/fendo.2022.1016373 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref042] 42. Conley A. In: Ventricular Enlargement. New York, NY: Springer New York; 2011. p. 2600–2601. Available from: 10.1007/978-0-387-79948-3_375. [DOI] [Google Scholar]

[pgen.1011332.ref043] 43. Orellana C, Ferreira D, Muehlboeck J, Mecocci P, Vellas B, Tsolaki M, et al. Measuring global brain atrophy with the brain volume/cerebrospinal fluid index: normative values, cut-offs and clinical associations. Neurodegenerative Diseases. 2016;16(1-2):77–86. doi: 10.1159/000442443 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref044] 44. Moghaddam HS, Sherbaf FG, Aarabi MH. Brain microstructural abnormalities in type 2 diabetes mellitus: A systematic review of diffusion tensor imaging studies. Frontiers in Neuroendocrinology. 2019;55:100782. doi: 10.1016/j.yfrne.2019.100782 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref045] 45. Hsu JL, Chen YL, Leu JG, Jaw FS, Lee CH, Tsai YF, et al. Microstructural white matter abnormalities in type 2 diabetes mellitus: a diffusion tensor imaging study. Neuroimage. 2012;59(2):1098–1105. doi: 10.1016/j.neuroimage.2011.09.041 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref046] 46. Reijmer YD, Brundel M, De Bresser J, Kappelle LJ, Leemans A, Biessels GJ, et al. Microstructural white matter abnormalities and cognitive functioning in type 2 diabetes: a diffusion tensor imaging study. Diabetes care. 2013;36(1):137–144. doi: 10.2337/dc12-0493 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref047] 47. Dong SS, Yu K, Tang SH, Guo J, Guo Y, Yang TL. Evaluation of the causal associations between brain imaging-derived phenotypes and type 2 diabetes: a bidirectional Mendelian randomization study. medRxiv. 2022; p. 2022–12. [Google Scholar]

[pgen.1011332.ref048] 48. Alotaibi A, Podlasek A, AlTokhis A, Tench CR, Mohammadi-Nejad AR, Sotiropoulos SN, et al. White matter microstructural alteration in type 2 diabetes: a combined UK Biobank study of diffusion tensor imaging and neurite orientation dispersion and density imaging. medRxiv. 2022; p. 2022–07. [Google Scholar]

[pgen.1011332.ref049] 49. Petersen RC, Aisen PS, Beckett LA, Donohue MC, Gamst AC, Harvey DJ, et al. Alzheimer’s Disease Neuroimaging Initiative (ADNI): clinical characterization. Neurology. 2010;74(3):201–209. doi: 10.1212/WNL.0b013e3181cb3e25 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref050] 50.Kingma DP, Welling M. Auto-Encoding Variational Bayes; 2013. Available from: https://arxiv.org/abs/1312.6114.

[pgen.1011332.ref051] 51.Kingma DP, Ba J. Adam: A Method for Stochastic Optimization; 2014. Available from: https://arxiv.org/abs/1412.6980.

[pgen.1011332.ref052] 52. Isensee F, Schell M, Pflueger I, Brugnara G, Bonekamp D, Neuberger U, et al. Automated brain extraction of multisequence MRI using artificial neural networks. Human Brain Mapping. 2019;40(17):4952–4964. doi: 10.1002/hbm.24750 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref053] 53. Jenkinson M, Beckmann CF, Behrens TE, Woolrich MW, Smith SM. Fsl. Neuroimage. 2012;62(2):782–790. doi: 10.1016/j.neuroimage.2011.09.015 [DOI] [PubMed] [Google Scholar]

[pgen.1011332.ref054] 54.He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; 2016. p. 770–778.

[pgen.1011332.ref055] 55. Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, et al. PyTorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems. 2019;32. [Google Scholar]

[pgen.1011332.ref056] 56. Loh PR, Tucker G, Bulik-Sullivan BK, Vilhjalmsson BJ, Finucane HK, Salem RM, et al. Efficient Bayesian mixed-model analysis increases association power in large cohorts. Nature Genetics. 2015;47(3):284–290. doi: 10.1038/ng.3190 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref057] 57. Purcell S, Neale B, Todd-Brown K, Thomas L, Ferreira MA, Bender D, et al. PLINK: a tool set for whole-genome association and population-based linkage analyses. The American Journal of Human Genetics. 2007;81(3):559–575. doi: 10.1086/519795 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref058] 58. Machiela MJ, Chanock SJ. LDlink: a web-based application for exploring population-specific haplotype structure and linking correlated alleles of possible functional variants. Bioinformatics. 2015;31(21):3555–3557. doi: 10.1093/bioinformatics/btv402 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref059] 59. Millard LA, Davies NM, Gaunt TR, Davey Smith G, Tilling K. Software Application Profile: PHESANT: a tool for performing automated phenome scans in UK Biobank. International Journal of Epidemiology. 2017;47(1):29–35. doi: 10.1093/ije/dyx204 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref060] 60. Ge T, Chen CY, Ni Y, Feng YCA, Smoller JW. Polygenic prediction via Bayesian regression and continuous shrinkage priors. Nature communications. 2019;10(1):1776. doi: 10.1038/s41467-019-09718-5 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref061] 61. Pain O, Glanville KP, Hagenaars SP, Selzam S, Fürtjes AE, Gaspar HA, et al. Evaluation of polygenic prediction methodology within a reference-standardized framework. PLOS Genetics. 2021;17(5):e1009021. doi: 10.1371/journal.pgen.1009021 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref062] 62. Monti R, Eick L, Hudjashov G, Läll K, Kanoni S, Wolford BN, et al. Evaluation of polygenic scoring methods in five biobanks reveals greater variability between biobanks than between methods and highlights benefits of ensemble learning. medRxiv. 2023; p. 2023–11. [Google Scholar]

[pgen.1011332.ref063] 63.Team PC. PGS Catalog Calculator; in preparation. https://github.com/PGScatalog/pgsc_calc.

[pgen.1011332.ref064] 64. Bulik-Sullivan BK, Loh PR, Finucane HK, Ripke S, Yang J, of the Psychiatric Genomics Consortium SWG, et al. LD Score regression distinguishes confounding from polygenicity in genome-wide association studies. Nature Genetics. 2015;47(3):291–295. doi: 10.1038/ng.3211 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref065] 65. Bulik-Sullivan B, Finucane HK, Anttila V, Gusev A, Day FR, Loh PR, et al. An atlas of genetic correlations across human diseases and traits. Nature Genetics. 2015;47(11):1236–1241. doi: 10.1038/ng.3406 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref066] 66. 1000 Genomes Project Consortium and others. A global reference for human genetic variation. Nature. 2015;526(7571):68. doi: 10.1038/nature15393 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pgen.1011332.ref067] 67. International HapMap 3 Consortium and others. Integrating common and rare genetic variation in diverse human populations. Nature. 2010;467(7311):52. doi: 10.1038/nature09298 [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

TransferGWAS of T1-weighted brain MRI data from UK Biobank

Alexander Rakowski

Remo Monti

Christoph Lippert

Roles

Abstract

Author summary

1 Introduction

Fig 1. Overview of our study and workflow.

2 Results

2.1 Extraction of deep features from the T1-weighted UK Biobank brain MRI scans

2.2 Interpretation of the DNN features

Fig 2. Results of the PheWAS performed on the principal components (PCs) of the ImageNet (blue) and ADNI (yellow) pretrained models.

Fig 3. Fraction of significantly correlated voxels in each brain region of interest (ROI) for each principal component of the neural network models features.

2.3 GWAS results

Fig 4. Manhattan plot of GWAS (n = 36, 311 individuals, 16, 472, 121 SNPs) performed on features of the ImageNet-pretrained model aggregated over all 10 features of the model.

Fig 5. Manhattan plot of GWAS (n = 36, 311 individuals, 16, 472, 121 SNPs) performed on features of the ADNI-pretrained model aggregated over all 10 features of the model.

Fig 6. h2 heritability estimates of principal components (PCs) of the ImageNet (blue) and ADNI (yellow) pretrained models, obtained using linkage disequilibrium score regression (LDSC).

2.3.1 GWAS catalog associations

Fig 7. Number of independent loci per trait with associations reported in previous studies included in the NHGRI-EBI GWAS Catalog [16].

2.3.2 Loci without previously reported GWAS catalog associations

Fig 8. Brain MRI voxels corresponding to genetic regions with no previously reported GWAS associations.

Fig 9. Fractions of volume of brain regions correlated with genetic regions with no previously reported GWAS associations.

2.3.4 Comparison with previous studies

2.4 TransferGWAS polygenic scores

2.4.1 PGS PheWAS

2.4.2 Predictive performance compared to trait-specific PGS

2.5 Genetic correlations

Fig 10. Genetic correlation coefficients between the 20 deep neural network (DNN) principal components (PCs) (rows) and 23 significantly associated phenotypes (columns), out of 27 candidate traits from the UK Biobank (UKB).

2.5.1 ADNI4 and T2D

3 Discussion

4 Materials and methods

4.1 Training of the neural network models

4.2 GWAS

4.3 PheWAS

4.4 Polygenic scores

4.5 Genetic correlations

Supporting information

Acknowledgments

Data Availability

Funding Statement

References

Decision Letter 0

Hua Tang

Xiang Zhou

Roles

Author response to Decision Letter 0

Decision Letter 1

Hua Tang

Xiang Zhou

Roles

Acceptance letter

Hua Tang

Xiang Zhou

Roles

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases

Fig 6. h² heritability estimates of principal components (PCs) of the ImageNet (blue) and ADNI (yellow) pretrained models, obtained using linkage disequilibrium score regression (LDSC).