BrainGENIE: The Brain Gene Expression and Network Imputation Engine

Ex vivo molecular analysis of the human brain is virtually impossible given major risks and ethical concerns. Transcriptome imputation offers a promising and non-invasive alternative for developing models (albeit imperfect) of brain gene expression in lieu of biopsying brain tissue. Popular tools such as FUSION (Gusev et al., 2016) and PrediXcan (Gamazon et al., 2015) use genotypes at common cis-expression quantitative trait loci (eQTLs) to predict tissue-specific gene expression levels. However, those tools cannot reliably predict expression levels for a majority of genes in the brain. This raises the question of whether an alternative modeling approach should be evaluated to capture greater variance in more genes in the brain that are not yet imputable with existing cis-eQTL imputation tools. To address this problem, we developed a novel transcriptome-imputation method called the Brain Gene Expression and Network Imputation Engine (BrainGENIE) that imputes brain-region-specific gene expression levels from peripheral blood gene expression. BrainGENIE predicted brain-region-specific expression levels for 1,733 – 11,569 genes (cross-validation R2≥0.01, false-discovery rate-adjusted p<0.05), few of which are imputable by PrediXcan. Disease-related transcriptome signals detected by BrainGENIE showed stronger agreement with known transcriptome signatures from postmortem brain when compared with findings from analyses of peripheral blood or S-PrediXcan. BrainGENIE complements and outperforms existing transcriptome-imputation tools, provides biologically meaningful predictions, and opens avenues for studying brain transcriptomes longitudinally. BrainGENIE was developed using R (v.3.6.3, tested in 4.0.2) and is freely available at: https://github.com/hessJ/BrainGENIE.


Introduction
Brain disorders cause much disability worldwide (James et al., 2018). Typically, ex vivo molecular assessment of human disease centers on the primarily affected tissue(s), or the site of pathogenesis, but that is not possible for brain disorders. It is virtually impossible to conduct brain biopsy in a risk-free manner, which makes experimental analysis of ex vivo human brain tissue highly improbable. Consequently, there are a limited set of research questions pertaining to the brain that can be dealt with using human participants. Transcriptome imputation offers a non-invasive alternative to brain biopsy by allowing investigators to infer tissue-specific gene expression without directly assaying gene expression levels. Instead, brain-region-specific predictions about gene expression can be made in living persons by proxy of genetic variants genotyped from peripheral tissues (e.g., blood), capitalizing on the known regulatory relationships between those variants and the expression of transcripts.
FUSION and PrediXcan focus on tissue-specific effects of expression quantitative trait loci (eQTLs) on expression of proximal genes (cis-eQTLs) to impute transcriptome profiles.
These methods have been successful in prioritizing genome-wide association study (GWAS) hits and have helped reveal putative mechanisms underlying complex disorders (Hall et al., 2020;Liao et al., 2019;Cheng et al., 2019;Smit et al., 2019;Hohman et al., 2017). There is a striking disparity between the number of genes imputable in the brain versus non-central nervous system tissues: FUSION imputes an average of 3,158 genes in brain (range = 1,604 -5,855 across 12 brain regions) compared with 5,592 in the non-central nervous system tissues, while PrediXcan imputes an average of 4,337 genes in the brain (range = 2,559 -6,794) compared with 6,262 genes (range = 1,642 -10,012) in non-central nervous system tissues.
The clear majority of genes in the brain transcriptome are not reliably predicted with FUSION or PrediXcan, suggesting that other predictors may be needed to account for greater variance in transcriptome profiles.
Compared with PrediXcan, TIGAR, a Bayesian modeling framework for predicting gene expression from eQTL data by more robustly measuring cis-eQTL signals, increases the number of imputable genes, and improves accuracy of predictions (Nagpal et al., 2019). A Bayesian hierarchical model called EpiXcan also sought to improve weaknesses of PrediXcan by applying epigenetic annotations to optimize the weights assigned to cis-eQTLs and increase predictability of gene expression levels (Zhang et al., 2019). It is likely that this trend of optimizing transcriptome imputation will continue apace with efforts to provide deep annotations of the human genome and amass catalogs of transcriptome profiles for bulk tissue and single cells (Gandal, Zhang, et al., 2018;Fromer et al., 2016;Polioudakis et al., 2019;Feingold et al., 2004).
Tissue-specific and tissue-dependent gene expression profiles help differentiate between brain and peripheral tissues, but compelling evidence shows that brain and blood exhibit comparable transcriptome profiles (Tylee et al., 2013;Hess et al., 2016;McKenzie et al., 2014;Qi et al., 2018). Our group systematically reviewed relevant literature on this topic, and found that gene expression profiles in blood and brain are moderately correlated (Pearson's r of 0.24 to 0.64), with 35% to 80% of genes expressed in both tissues (Tylee et al., 2013). In a later study, we found empirical evidence that ~90% of weighted gene-gene interaction networks identified in the blood from transcriptome profiles were significantly preserved in the prefrontal cortex (Hess et al., 2016). Brain and blood also show significant overlap with respect to eQTLs (McKenzie et al., 2014;Qi et al., 2018), signifying that shared genetic effects (albeit small effect sizes) may explain gene expression comparability between blood and brain. Although previous studies have suffered drawbacks with respect to inherent design limitations (i.e., lack of withinsubject studies done on living humans), findings suggest that blood gene expression is a reasonable (albeit imperfect) proxy for brain gene expression.
Based on this evidence, we sought to capitalize on the transcriptome similarity between brain and blood (and the easy accessibility of blood) to make predictions about gene expression in the brain solely based on observed expression in peripheral blood. Simultaneously, we sought to develop a transcriptome-imputation approach that complements cis-eQTL-based transcriptome imputation methods. We achieve these goals with a novel Brain Gene-Expression and Network-Imputation Engine (BrainGENIE), which imputes brain-region-specific gene-expression profiles based on gene-expression profiles assayed from peripheral blood.
BrainGENIE is implemented in the R statistical environment and is distributed as a freely available software (https://github.com/hessJ/BrainGENIE). We sought to develop a toolset that capitalizes on human blood transcriptomes because these data are widely available through public repositories or can be generated de novo with relative ease and cost-effectiveness.
Moreover, because gene profiles in blood fluctuate over time they hold valuable information about corresponding temporal changes in the brain. In contrast, temporal components of gene expression regulation are not yet predictable with existing cis-eQTL-based toolsets.
There is no other method that uses blood-based gene expression to predict brain gene expression, thus we were unable to directly benchmark performance of BrainGENIE against a equivalent algorithm. We compared BrainGENIE with PrediXcan since the two methods are conceptually similar, which helped us to understand differences between using blood-based gene expression profiles versus eQTLs to impute region-specific gene expression in the brain.
Our findings show that BrainGENIE is capable of imputing brain-region-specific expression levels for an average 6,302 genes (range: 1,733 -11,569 genes), of which 18.7%, are reliably predicted by PrediXcan. We used BrainGENIE to impute brain transcriptomes using external ex vivo blood-based transcriptome data, and found that disease-related differential expression patterns detected in the imputed data better aligned with results found directly in postmortem brain compared with results obtained by PrediXcan or results directly measured in blood.

Data
RNA-sequencing-derived gene-expression data from human postmortem brain tissue and whole blood for 267 adult donors were downloaded from the dbGaP repository from the current release of the GTEx Project (v.8). The inclusion criteria for donors collected by the GTEx Consortium were as follows: a body mass index (BMI) of 18.5 -35, time of death less than 24 hours by tissue collection, no blood transfusions within 48 hours of time of death, no metastatic cancer, no chemotherapy or radiation within 2 years of time of death, and no communicable diseases that would disqualify donors from tissue donation. Transcriptome profiles from twelve brain regions were available (amygdala, anterior cingulate cortex, caudate, cerebellar hemisphere, cerebellum, cortex, frontal cortex, hippocampus, hypothalamus, nucleus accumbens, putamen, substantia nigra). Genome-wide microarray genotyping data were available for 218 donors. Some brain regions were not available from some donors, hence pairs of blood-brain transcriptome profiles ranged from n = 86 -153. Ages of donors ranged from 20 -70 years old with the distribution skewed toward older persons (50% over the age of 61 years, mean age = 57.9 years). Approximately 30% of donors were female. Donors were recorded predominantly as European (n=199, 91%), with 8% recorded as black or African American (n=17), and less than 1% recorded as Asian, Alaska Native, or American Indian (n=2).

Normalization of RNA-sequencing data from GTEx
Total RNA was extracted from whole blood stored in PAXgene Blood RNA (Qiagen®) tubes and snap frozen brain tissue. All RNA samples that were used for RNA-sequencing had a RNA integrity number ≥ 6.0 as measured by Agilent Bioanalyzer. Detailed procedures for tissue collection, library preparation, and sequencing were previously described by the GTEx Consortium (GTEx Consortium et al., 2015). Gene-level read counts released by GTEx were summarized based on the Gencode 26 (GRCh38) transcript model. We used the following steps used by the GTEx Consortium (v6p) to pre-process gene counts for analysis, including: retain genes with > 0.1 read per kilobase per million (RPKM) and ≥5 read counts in at least 10 donors, quantile normalize RPKM to adjust for between-sample variation (limma v.3.24.3) (Ritchie et al., 2015), and inverse-rank normalization.
A methodological difference between BrainGENIE and PrediXcan is that we did not include probability estimates of expression residuals (PEER) factors as covariates in our multiple regression models in order to preserve variation in transcriptome levels for model training. The GTEx Consortium cautioned that the use of PEER factors may be overly aggressive, citing that PEER factors have a large impact on variation in gene expression levels (58 -78%) (Aguet et al., 2017), which may remove useful biological variation. In light of this concern, we used clearly defined technical and biological covariates instead of PEER factors when performing corrections on transcriptome profiles. Multiple regression was used to compute residual expression values for blood and brain tissues by removing additive effects attributed to age, sex, ancestry covariates attributed to the top three genotype-derived principal components, PCR method used in library preparation, RNA-sequencing platform, RNA integrity number (RIN), postmortem interval, and death classification of each donor measured by the Hardy scale. For whole blood, we included three additional covariates (defined as the top 3 PCs from CIBERSORT) to adjust for between-donor variation in circulating leukocyte abundance (Supplementary Figure 1).
Residual expression levels were used for building prediction models of brain gene expression.

Training process for BrainGENIE and evaluating prediction performance
We performed a single 10-fold cross-validation to estimate the predictive performance of BrainGENIE separately for each brain region. Paired blood-brain transcriptome profiles from GTEx donors were randomly assigned to 10 folds. For each training set, a PCA was performed on normalized blood transcriptome profiles and linear regression was trained to predict brainregion-specific expression levels per gene using an arbitrarily chosen number of the 20 PCs, which captured roughly 58% of the variation in blood gene expression, to represent gene expression profiles en masse in peripheral blood. To evaluate robustness of predicted brainregional gene expression levels relative to a different choice of top k PCs, we developed BrainGENIE models using the top 5 (11% variance explained) , 10 (41% variance explained), and 40 PCs (80% variance explained) of peripheral blood gene expression. Our initial work uncovered that prediction accuracies achieved by linear regression were as good as or better than elastic net regression (the model used by PrediXcan) but computationally faster to train.
The trained models were then deployed in the validation set to estimate the predictive performance on unseen data. The metric for prediction performance was the coefficient of determination for observed and predicted expression levels (R 2 ) in the hold-out fold. This process was repeated until each fold was used as the validation set, and prediction performance was averaged over the validation sets. In order to have a reasonable side-by-side comparison between BrainGENIE and PrediXcan, we adopted the same criterion for "reliably predicted" as PrediXcan, i.e., genes that could be predicted with a cross-validation R 2 ≥ 0.01 and with false-discovery rate adjusted p-value (FDRp) < 0.05.

Comparison of prediction accuracies between methods
We compared the average cross-validation R 2 prediction accuracies obtained by BrainGENIE and PrediXcan. We then used binomial exact test to determine if one method significantly out-performed the other in terms of the proportion genes that met the criterion for being "reliably predicted" (i.e., R 2 ≥ 0.01, FDRp < 0.05). The expected probability was set to 50%, which represents the value expected under the null hypothesis that there is no systematic difference between the two methods.

Synaptic gene set enrichment test
The aim of this analysis was to test for a difference between BrainGENIE and PrediXcan in terms of enrichment of synaptic ontologies among genes reliably predicted by these methods.
Synaptic ontologies were chosen to provide another means of comparing methods, and to limit the scope of this analysis to gene sets that are fundamental to the brain, many of which are overrepresented with common genetic variation associated with neuropsychiatric disorders (Koopmans et al., 2019). We obtained gene annotations for synaptic gene ontologies from the SynGO database (release 2018-07-31) (Koopmans et al., 2019). We used a two-proportions 2 test to compare BrainGENIE and PrediXcan in terms of the proportions of presynaptic (no. of genes = 747) and postsynaptic genes (no. of genes = 1,351) reliably predicted by each method.

Brain cell-type marker gene enrichment test
The aim of this analysis was to test for a difference between BrainGENIE and PrediXcan in terms of enrichment of brain cell-type marker genes that were reliably predicted by these methods. We downloaded lists of cell-type marker genes via the R package BRETIGEA (v1.0.0) (McKenzie et al., 2018) for neurons, astrocytes, endothelial cells, oligodendrocytes, microglia, and oligodendrocyte precursor cells identified from postmortem human brain tissue derived from the Allen Brain Atlas. We used a two-proportions 2 test to compare BrainGENIE and PrediXcan in terms of the proportions of marker genes for the six brain cell types reliably predicted by each method.

Proportion of cross-disorder pleotropic gene sets imputable with BrainGENIE
The goal of this analysis was to employ gene set enrichment to determine whether BrainGENIE and PrediXcan differ in terms of their ability to reliably predict gene sets that show significant association with major neuropsychiatric disorders in GWAS. We obtained Gene Ontology (GO) identifiers for 45 gene sets identified by GWAS meta-analysis as having a shared association across eight neuropsychiatric disorders (Lee et al., 2019). GO identifiers were annotated with HGNC gene symbols using the Molecular Signatures Database (v.6.2) (Liberzon et al., 2011). We calculated the proportion of genes per gene set that were reliably predicted using BrainGENIE (20 PC model), and compared this with the proportion of each gene set reliably predicted using PrediXcan using a two-proportions 2 test.
Briefly, array-based expression data were processed using standard procedures best-suited for each study's particular array platform, expression values were log2 transformed and quantile normalized, probes were mapped to HGNC human gene symbols, median expression levels were calculated for groups of probes mapping to a common gene to collapse expression to a single value, normalized gene expression levels were z-standardized (mean=0, s.d.=1) per study to minimize unwanted variation due to platform differences, and standardized expression data were merged across studies based on common gene symbols. Additional details for our normalization procedure are available in the supplement (Supplementary Methods). The combined peripheral blood transcriptome data for each disorder was then supplied to BrainGENIE in order to impute transcriptome profiles for frontal cortex with two modes: 1) using the top 10 blood-based gene expression PCs, as that parameter led to imputation of fewer genes with better overall accuracy, and 2) using the top 40 blood-based gene expression PCs, as that parameter maximized the total number of imputable genes.
We calculated differential gene expression (DGE) estimates of blood-based gene expression levels among affected cases versus unaffected comparison subjects with combinedsamples mega-analyses, adjusting for age, sex, and study site. Similarly, we calculated DGE estimates using predicted gene expression profiles for frontal cortex obtained from BrainGENIE using the same mega-analysis approach. In addition, we inferred DGE signals for SCZ, BD, and ASD by applying the method S-PrediXcan applied to GWAS summary statistics for each disorder (Ripke et al., 2014;Grove et al., 2019;Stahl et al., 2019). Transcriptome-wide DGE effect sizes for each disorder obtained from peripheral blood mega-analyses (t-values), BrainGENIE mega-analyses (t-values), and S-PrediXcan were then compared with DGE effect sizes directly measured from postmortem brain published by the PsychENCODE Consortium (log2 fold-changes) using Pearson's correlation test. Pearson's correlation coefficient was chosen in order to assess the linear monotonic relationship between DGE signals derived from different methods. Table 1 summarizes the cross-validation prediction performance of BrainGENIE per brain region. Table 2 depicts the number of genes that surpassed our criteria for reliably predicted (i.e., average R 2 ≥ 0.01, FDRp < 0.05) per brain region, alongside the number of genes that PrediXcan is capable of imputing, and the overlap between BrainGENIE and PrediXcan. The statistics in Table 1 and Table 2 reflect the prediction performance for the top 20 PCs of blood-based transcriptome-wide gene expression, which yielded a higher average number of imputable genes per brain region relative to 5, 10, or 40 PCs. The prediction performance of BrainGENIE, measured by the average cross-validation R 2 , ranged from 0.08 -0.16 for genes that met the criteria of significance based on predictability during cross-validation (average R 2 ≥0.01, FDRp<0.05). The distributions of cross-validation R 2 values produced by BrainGENIE and PrediXcan for all reliably predicted genes are shown in Supplementary   Figure 2. The shape of the distributions found using BrainGENIE were similar PrediXcan, however, PrediXcan featured heavier right-tails compared to BrainGENIE (Supplementary   Figure 2). The proportion of genes in the brain that met the significance criterion (R 2 ≥0.01, FDRp<0.05) ranged from 9 -57% (mean number of genes = 6,302; range: 1,733 -11,569 genes). The maximum average cross-validation prediction accuracy of BrainGENIE across all brain regions ranged from R 2 = 0.37 -0.67. Brain-region-specific expression levels of three genes could be predicted with at an average cross-validation accuracy of R 2 ≥ 0.5, namely:
Prediction accuracies generated by BrainGENIE were significantly better (binomial FDRp < 0.05) than those produced by PrediXcan for genes that both methods could reliably predict in the amygdala (no. of genes = 709) and substantia nigra (no. of genes = 207,  Supplementary Table 3, BrainGENIE reliably predicted a significantly larger fraction of presynaptic and postsynaptic genes compared with PrediXcan (FDRp<0.05) for eight out of 12 brain regions. PrediXcan predicted a larger fraction of presynaptic and postsynaptic genes in the anterior cingulate cortex, cortex, and substantia nigra compared with BrainGENIE. In the hippocampus, BrainGENIE predicted a larger proportion of postsynaptic genes compared with PrediXcan. However, no statistically significant difference was found between BrainGENIE and PrediXcan in terms of the proportion of presynaptic genes that the methods predicted in the hippocampus. The largest difference seen between these two methods in terms of coverage of pre-and post-synaptic genes was among those genes expressed in frontal cortex, wherein the two methods showed an absolute difference of about 56% (~68% coverage by BrainGENIE versus ~11% coverage by PrediXcan).

Brain cell-type marker gene enrichment test
BrainGENIE reliably predicted more human brain cell-type marker genes than PrediXcan for at least one brain cell type in 10 regions (Supplementary Table 4). The following patterns emerged: markers for astrocytes, neurons, oligodendrocytes, and oligodendrocyte precursor cells (OPCs) were better predicted by BrainGENIE compared with PrediXcan in six brain regions; endothelial marker genes were better predicted by BrainGENIE in nine brain regions; and microglia marker genes were better predicted by BrainGENIE in 10 brain regions. The most significant difference that emerged in favor of BrainGENIE was the proportion of neuronal marker genes that were reliably predicted within the frontal cortex by BrainGENIE (65.2%) compared with PrediXcan (11.8s %) ( 2 = 599.9, p = 1.74x10 -132 , FDRp = 1.25×10 -130 ).

Proportion of cross-disorder gene sets imputable with BrainGENIE versus PrediXcan
Supplementary Figure 3 shows that, compared with PrediXcan, BrainGENIE provided significantly better coverage for 14 of 45 Gene Ontology (GO) gene sets found to have the strongest relationship with cross-disorder genetic risk across eight neuropsychiatric disorders (Lee et al., 2019). After multiple-testing correction, no statistically significant difference was found between the two methods in their coverage for 31 of the 45 neuropsychiatric crossdisorder gene sets. On average, BrainGENIE reliably predicted 1.73-fold (95% confidence interval [CI]: 1.62, 1.84) more genes across the 45 neuropsychiatric cross-disorder gene sets than did PrediXcan.

Comparison of differential gene expression signatures across methods
Pearson's correlation tests revealed that DGE effects for SCZ measured in peripheral blood were significantly congruent with DGE signals directly measured in postmortem brain from the PsychENCODE Consortium's microarray meta-analysis (Pearson's r = 0.053, 95% CI [0.035, 0.072], n genes = 10,832, p = 2.6×10 -08 ), and from the CommonMind Consortium (Pearson's r = 0.11, 95% CI [0.093, 0.133], n genes = 9,434, p = 2.1×10 -28 ), but not the PsychENCODE Consortium's RNA-sequencing analysis (Figure 1). DGE effects for ASD from peripheral blood were inversely correlated with postmortem brain (Figure 1), which may reflect age differences between samples considering that individuals in the peripheral blood datasets were predominantly children whereas those in the postmortem brain studies were predominantly adults. DGE signals found using S-PrediXcan were not significantly correlated with postmortem brain DGE signals for SCZ, BD, or ASD (Figure 1). Conversely, DGE effects estimated from predicted genes' expression profiles in brain using BrainGENIE were significantly correlated with results directly measured in postmortem brain (Figure 1). The strongest correlation that emerged was between DGE signals obtained using BrainGENIE and DGE signals directly measured in postmortem brain for SCZ found by the PsychENCODE Consortium's microarray meta-analysis (Pearson's r = 0.52, 95% CI [0.495, 0.548], n genes = 2,913, p = 1.12×10 -203 ).
The DGE correlations for ASD, BD, and SCZ found between BrainGENIE and postmortem brain replicated with significance using results from an independent PsychENCODE Consortium cohort profiled via RNA-sequencing (Figure 1). Furthermore, the SCZ DGE correlation between BrainGENIE and postmortem brain replicated with significance in a second independent cohort from the CommonMind Consortium (Pearson's r = 0.34, 95% CI [0.33, 0.37], n genes = 8,077, p = 4.9×10 -230 , Figure 1). In general, the DGE agreement between BrainGENIE and postmortem brain was significantly stronger than peripheral blood with postmortem brain for ASD, BD, and SCZ for at least one of the four BrainGENIE models (z-test p-values < 0.05, Supplementary   Table 5). Similarly, DGE signals from BrainGENIE were significantly more concordant with postmortem brain than S-PrediXcan with postmortem brain for ASD, BD, and SCZ (Supplementary Table 5).

This study introduced and benchmarked a novel computational method called
BrainGENIE, which predicts brain-region-specific gene expression profiles based on observed expression profiles in peripheral blood. Over the past decade, there has been rapid growth in the number of blood-based transcriptome studies aimed at identifying biomarkers for neuropsychiatric disorders. This has led to a vast amount of useful data that may hold untapped information about the brain. Much of the raw data from published blood-based transcriptome studies of neuropsychiatric disorders can be readily downloaded from public repositories (i.e.,

Gene Expression Omnibus [GEO], ArrayExpress) or made available to investigators with
controlled access (i.e., dbGaP, NIMHGR, Synapse). It is therefore possible to further mine those stores of transcriptome data with BrainGENIE, thus generating novel mechanistic hypotheses about disease and advancing our understanding of psychopathology.
BrainGENIE exploits PCA as an efficient method of data-reduction that can also reduce the potential of overfitting, i.e., limiting the number of input features steers a model away from "learning" random noise. However, risk for over-fitting is not eliminated by PCA alone. We applied a standard machine learning approach of k-fold cross-validation to estimate the ability of BrainGENIE to generalize its predictions on unseen data. The predictability of BrainGENIE is undoubtedly affected by the number of samples available for training. Brain regions that had large sample sizes (i.e., cortex, frontal cortex) showed better prediction performance relative to brain regions with fewer samples (i.e., amygdala, substantia nigra). Furthermore, there were 48 fewer donors (range: 28 -63), on average, with paired blood-brain transcriptome data from GTEx available for BrainGENIE compared with the number of donors with paired genetic and transcriptome data used for PrediXcan. It is challenging to draw strong conclusions about differences in model performance for individual genes between BrainGENIE and PrediXcan given this disparity of sample size. However, global differences between the methods could not be explained by variation in sample size alone. Even with BrainGENIE limited by smaller sample size for model training, we found that BrainGENIE can impute a substantial fraction of genes that were not imputable using PrediXcan. This suggests that non-genetic components of gene expression ignored by PrediXcan models hold valuable information for transcriptome imputation.
Considering that transcriptomic data from bulk postmortem brain tissue was used to develop BrainGENIE, we are not able to specifically model gene expression for any specific brain cell type. We are instead exploiting cross-tissue overlap at the level of cell mixtures in the brain and blood. It is possible that commonalities seen between brain and blood gene expression could be driven by a possible shared lineage between macrophages and microglia (Chan et al., 2007;Prinz and Priller, 2014). Nevertheless, studies have implicated glial cell dysfunction in neuropsychiatric disorders (Tay et al., 2018), which makes them an interesting target for gene expression modeling. The neuronal component of brain gene expression levels imputed by BrainGENIE is unknown. However, inferences can still be made about how different brain cell types explain components of gene expression through cell deconvolution analysis, which can be accomplished by several algorithms that make use of gene expression data from single cells.
The current version of BrainGENIE can predict expression levels for 1,733 to 11,569 of genes in the human brain (depending on the brain region), which accounts for about 9 to 57% of the brain transcriptome. Iterations of BrainGENIE have made continual but gradual improvements in the number of reliably predicted genes and variance accounted for in brainregion-specific gene expression levels. This suggests that further refinement of our models will continue to improve predictions until they reach their (unknown) maximum per gene and per brain region. As we would expect, not all or even most genes are imputable with BrainGENIE, but the number of new genes that can be imputed with BrainGENIE and not by PrediXcan is considerable. We found that some genes are imputable by both BrainGENIE and PrediXcan, but the amount of overlap between methods in terms of imputable genes was relatively small. This suggests that there is value to integrating BrainGENIE and PrediXcan for a combined and complementary approach to transcriptome imputation. Ideally, the strengths of multiple modeling approaches like those in BrainGENIE, PrediXcan, and others, would be combined into a unified framework (or multiple models simply ran separately and output merged) to deliver a holistic and effectual portrait of the landscape of the human brain transcriptome.
Our results showed that DGE results obtained using BrainGENIE were in better agreement with DGE signals measured directly in postmortem brain compared with measures made in peripheral blood and those imputed by PrediXcan. This advantage of BrainGENIE over peripheral blood and PrediXcan was most striking for SCZ but was still evident for BD and ASD. An interesting finding that emerged for ASD from our DGE concordance analysis was the inverse correlation between DGE measured in peripheral blood and DGE directly measured in postmortem brain. There are a variety of possible explanations for this finding, but no definitive conclusion could be drawn. One consideration is that the participants in the ASD dataset are mostly young children and toddlers (mean age = 5.8 years, 95% CI [4.8, 5.3]), whereas the GTEx dataset used to train BrainGENIE contained only adults (aged 20 -70).

DGE concordance between
Additionally, the PsychENCODE dataset for ASD contained mostly adults (mean age = 29.7 years, 95% CI [26.3, 33.1]). Either way these results suggest that BrainGENIE needs to be further optimized before wide-scale application on subjects out-of-range relative to the age of the GTEx (or other emergent) training samples.
We applied statistical corrections to remove effects of age, sex, and genetic ancestry from the gene expression data so that those factors would not systematically bias our models.
Still, it is possible that characteristics encoded in the GTEx dataset are not fully representative of the entire population. For example, donors in the GTEx Project were predominantly of European ancestry, hence limiting applicability of transcriptome imputation across ancestries.
Amassing large sample sizes that encompass a broader range of characteristics (e.g., environmental exposures, genetic background, and demographics, to name a few) would make it so that BrainGENIE can make use of more biological (useful) variability that may help increase the number of reliably predicted genes and improve variance accounted for. Devoting efforts to increasing sample ascertainment from diverse human populations, coupled with deeper phenotyping, are strategic ways to encourage more effective transcriptome imputation models.
In sum, BrainGENIE offers a novel approach to investigate brain-region-specific geneexpression profiles in living individuals. We demonstrated that gene-expression changes associated with disease and imputed in the brain by BrainGENIE were in better agreement with corresponding gene-expression changes detected in postmortem brain studies, relative to cis-eQTL-based predictions of gene expression by PrediXcan and gene expression changes detected in peripheral blood. The main challenge of transcriptome imputation is identifying a model and set of predictor variables that can efficiently and reliably predict gene expression levels and ensuring that downstream analyses of predicted expression levels can yield biologically meaningful results. PrediXcan and FUSION, respectively can impute an average of 18% and 16% of the brain transcriptome (compared with an average of 30% by BrainGENIE).
They have been successful in identifying novel tissue-specific gene expression dysregulation profiles for complex disorders. BrainGENIE overcomes some (but not all) of the limitations of cis-eQTL-based transcriptome imputation; namely, the strength of BrainGENIE is that it captures components of gene expression regulation that may be missed by cis-eQTL-based methods, which improves predictability of genes that are not imputable by PrediXcan and similar methods. BrainGENIE fills a void in transcriptome imputation by allowing analyses of genes that were not previously imputable or improving the predictability of disease-relevant gene sets for which PrediXcan can only partially impute. Though we posit that BrainGENIE has advantages over conceptually similar methods, our intention was for it to serve as a complement to geneticbased transcriptome imputations methods. In practice, our recommendation would be to integrate BrainGENIE with other methods whenever possible, to boost confidence in genedisease associations, hence expediting a deeper understanding of complex phenotypes. As such, BrainGENIE is a valuable toolset that has the potential to shed light on problems related to psychopathology and serve as a hypothesis-generating tool for mechanistic studies. Potential applications of BrainGENIE are far-reaching and would be best suited (relative to PrediXcan and FUSION) to study gene expression longitudinally, including: across developmental timepoints of the brain, pre-and post-exposure (e.g., environmental risks, traumatic life experiences), and modeling medication effects. BrainGENIE could be used to impute brainregion-specific transcriptomes at any point in a person's lifetime, opening the possibility that we could find causal and longitudinal mechanisms underlying neuropsychiatric disease.

Data and code availability
Data and source code can be accessed from the following GitHub repository: https://github.com/hessJ/BrainGENIE. This software was developed and tested on a personal laptop (2.6GHz 6-Core Intel i7 processor, 16 GB 2400 MHz DDR4). The number of samples in the target dataset will affect runtime of BrainGENIE, but it is generally efficient at handling typically sized datasets (runtime time for one brain region: ~30 seconds for ≤1,000 samples, ~1 minute for ≥10,000 samples). Statistics are shown for genes that were considered reliably predicted based on the following criteria: cross-validation [CV] R 2 ≥ 0.01, CV FDRp ≤ 0.05.