Gene expression associated with human brain activations in facial expression recognition

Previous studies identified some genetic loci of emotion, but few focused on human emotion-related gene expression. In this study, the facial expression recognition (FER) task-based high-resolution fMRI data of 203 subjects in the Human Connectome Project (HCP) and expression data of the six healthy human postmortem brain tissues in the Allen Human Brain Atlas (AHBA) were used to conduct a transcriptome-neuroimaging spatial association analysis. Finally, 371 genes were identified to be significantly associated with FER-related brain activations. Enrichment analyses revealed that FER-related genes were mainly expressed in the brain, especially neurons, and might be related to cell junction organization, synaptic functions, and nervous system development regulation, indicating that FER was a complex polygenetic biological process involving multiple pathways. Moreover, these genes exhibited higher enrichment for psychiatric diseases with heavy emotion impairments. This study provided new insight into understanding the FER-related biological mechanisms and might be helpful to explore treatment methods for emotion-related psychiatric disorders.


Introduction
Emotion is important in most aspects of human social behavior and cognitive activities (Niedenthal & Brauer, 2012;Pessoa, 2008). Facial expression recognition (FER) is a crucial feature of emotional expressions and is instrumental for our mental health. Facial expressions are created with the morphological changes of the face and reflect internal emotions, needs, and action tendencies of a person, such as frowning, widening eyes, and tightening lips (Ekman, 1992). The ability of FER helps us infer emotions displayed by others and then respond appropriately to such social and emotional information (Lopes et al., 2005). FER impairments are associated with many psychiatric diseases such as attention deficit and hyperactivity disorder (ADHD) (Shaw et al., 2014), autism spectrum disorder (ASD) (Harms et al., 2010), bipolar disorder (BP) (Lima et al., 2018), major depression disorder (MDD) (Bourke et al., 2010), and schizophrenia (SCZ) (Kring & Elis, 2013). For example, several studies investigated that recognizing facial expressions on others could be difficult in individuals with depression (Dalili et al., 2015;Groves et al., 2018;Kohler et al., 2011). Understanding the neurotic mechanism of FER has been becoming an important research topic in neuroscience.
Although many studies investigated the heritability of emotion, and both candidate genes and GWAS studies revealed genetic loci associated with emotion processing, the biological mechanisms of those emotion-related genes are still unclear. Those studies usually provide limited insight into the transcriptional mechanism of emotion processing. Gene expression assay is a direct measure of a gene's biological functions; assessing the association between emotion processing and gene expression may help understand the neuronal mechanisms of emotion processing. However, few gene expression studies have been performed to explore the neuronal mechanisms associated with emotion processing. Therefore, it is necessary to investigate underlying gene expressions which modulate emotion processing.
Across-individual transcription-neuroimaging association analysis is an excellent method that could be applied to investigate the association between brain functions and gene expression. In this study, the emotion task fMRI data in Human Connectome Project (HCP) was used to detect FER-related brain activation. HCP is a public resource that includes high-resolution task-based fMRI data of healthy adults (Smith et al., 2013) and provides a reliable emotion recognition-related brain activation map (Barch et al., 2013;Hariri et al., 2002b). Gene expression data were extracted from a public dataset, Allen Human Brain Atlas (AHBA) (Hawrylycz et al., 2012). AHBA provides high-resolution coverage mRNA transcriptome in more than 20,000 genes profiled in 3702 samples in different brain regions of six healthy human postmortem brain tissues. The samples could be mapped into Montreal Neurological Institute (MNI) space, allowing researchers to link gene expression with human neuroimaging data. Recent studies using a spatial pattern of gene expression maps provided by AHBA yielded new insights into transcriptional mechanisms of many cognitive activities and psychiatric disorders (Liu et al., 2019;Ritchie et al., 2018;Xie et al., 2020).
In the present study, the across spatial association was performed between the AHBA dataset and group-level FER activation maps of 203 healthy young adults in the HCP. First, a set of genes were identified as FER-related genes. Then, enrichment analyses of these genes about gene ontology, cell types, tissues, and diseases were performed. Moreover, protein-protein interaction (PPI) networks were conducted to gain an in-depth understanding of the molecular mechanisms of emotion recognition. Finally, hub genes were derived through PPI network analysis. The flowchart of the framework showed in Fig. 1.

Subjects
MRI Data in our study were from an open-source database, WU-Minn Human Connectome Project Data-500 Subjects (HCP_500) (Smith et al., 2013). Four hundred and sixtyfour subjects were scanned during performing six cognitive tasks (emotion, language, social, working memory, gambling, and relational tasks) in the HCP-500. After excluding 46 left-handed subjects, 418 subjects from 203 families were retained. Then, one subject was chosen from each family with gender balance. Finally, 203 right-handed healthy young adults (101 males and 102 females; age: 29.11 ± 3.50 years, range: 22-36 years) were included in our study. All subjects were screened for a history of neurodevelopmental, neuropsychiatric disorders, and genetic disorders. Written informed consents were provided by all subjects. The HCP study was ethically approved by the Washington University Institutional Review Board (IRB). The detailed inclusion and exclusion criteria were listed elsewhere . The demographic information is shown in Table 1.

FER task
The detailed HCP emotion task procedure was described in the previous literature (Hariri et al., 2002b). In brief, the subjects were asked to choose the correct face or shape on the bottom of the screen, which matches the figure on the top of the screen. The facial expression figures shown

MRI data acquisition, preprocessing, and statistical analysis
The MRI data were collected using a customized Siemens 3.0 T Connectome Skyra scanner with a standard 32-channel Siemens receive head coil at Washington University in St. Louis. The HCP data preprocessing methods are detailed elsewhere (Barch et al., 2013;Glasser et al., 2013). The parameters in MRI data acquisition and preprocessing steps are shown in Supplementary materials. A general linear model (GLM) was used to identify FER-related brain activation. Face versus shape contrast was used in the first-level analysis. Then, the 203 subjects were randomly divided into 102 subjects in one group and 101 subjects in another group. This step was repeated five times, and ten groups were formed. A group-level activation map for each group was obtained using a voxel-wise one-sample t-test in Statistical Parametric Mapping 12 (SPM12; https:// www. fil. ion. ucl. ac. uk/ spm/). The familywise (FWE) rate method with P < 0.05 and the minimum cluster size of 100 voxels were used to conduct multiple comparison corrections.
We hypothesized that FER-related genes would be distinct from the genes correlated with other cognitive functions to some extent. Therefore, other cognitive tasks in the HCP dataset, including language, social, working memory, gambling, and relational tasks, were also analyzed using the same method as the emotion task to confirm this hypothesis. The statistical contrast used in other tasks: reward versus punish in the gambling task, story versus math in the language task, relational versus match in the relational task, social versus random in the social task, and two back versus 0 back in the working memory task.

Gene expression data processes
In this study, we leveraged gene expression data of six donated human postmortem brain tissues from the AHBA open-source dataset (http:// human. brain-map. org) (Hawrylycz et al., 2012). According to a published pipeline, the gene expression data were processed (Arnatkeviciute et al., 2019). Detailed processing steps are shown in Supplementary materials. In this study, only 1782 samples in six left brain hemispheres were used because only two donors included samples of the right hemisphere. These processing procedures generated a 1782 samples × 10,185 genes transcription matrix.

Identification of the FER-related genes
The FER-related brain activation of each sample was represented by the mean t-score extracted from the group-level emotion activation map with a 4-mm radius sphere centered at this sample's MNI coordinates. Spearman correlation analyses were performed between brain activations and the expression levels of each given gene across 1782 samples. We randomly shuffled sample labels and re-performed the correlation analyses between group-level emotion activation and the gene expression data of 10,185 genes in ten groups, respectively. This analysis was repeated 10,000 times, and the maximum absolute R-value in 10,185 (genes) × 10 (groups) of each permutation was obtained to build an FWE-corrected null distribution. The genes that meet the permutation significance (P < 0.05) in this distribution were defined as FER-related genes in each group. To ensure the results' reliability, only genes that exhibited significant correlation in the same direction (i.e., positive or negative) in the ten groups were identified as final FER-related genes and used for further analyses. The same correlation analyses were conducted between the gene expression and the other cognitive task-related brain activations.

Gene Set Enrichment Analysis (GSEA) of the FER-related genes
Functional enrichment analysis of Gene Ontology (GO) was performed to analyze the FER-related genes' biological functions. The GO enrichment contains biological processes (BP), molecular functions (MF), and cellular components (CC). The enrichment analysis was performed using an open online tool g: Profiler (Raudvere et al., 2019) (https:// biit. cs. ut. ee/ gprofi ler) with a threshold of P < 0.05 (Bonferroni corrected). Moreover, another open online tool: Toppgene (https:// toppg ene. cchmc. org) (Chen et al., 2009), was used to perform GO enrichment analysis for validation.

Tissue-Specific Expression Analysis and Cell Type Enrichment Analysis
Tissue-specific expression analysis was performed using an online tool, TSEA (http:// genet ics. wustl. edu/ jdlab/ tsea/), to detect in which type of tissue the FER-related genes were predominantly expressed. To characterize differential expressions of FER-related genes in major cell types, an online cell type-specific expression analysis (CSEA) tool (http:// genet ics. wustl. edu/ jdlab/ csea-tool-2/) (Dougherty et al., 2010;Xu et al., 2014) was used to identify the cell types in which FER-related genes specially expressed. The multiple comparison corrections method is the Benjamini-Hochberg FDR (BH-FDR) method (P < 0.05). The specificity index (SI) was used to assess the specificity of a given RNA in one sample relative to all other samples analyzed. The smaller specificity index probability (pSI = 0.05, 0.01, 0.001, and 0.0001) represented a gene more likely expressed in a given tissue or cell type relative to other tissues or cell types.

Protein-Protein Interaction network
We constructed the PPI network analysis of the FER-related genes using the STRING (Szklarczyk et al., 2019) database version 11.0 (https:// string-db. org/), with a medium confidence value of 0.4, and other parameters were set to default. Key nodes in the PPI network were explored by cytoHubba (Chin et al., 2014) in Cytoscape software (version 3.8.0) (Shannon et al., 2003). Among the different topological algorithms provided by CytoHubba, Maximal Clique Centrality (MCC) has been reported to be the best option (Chin et al., 2014). Therefore, the MCC method was used to identify the top five key genes in the PPI network, which were considered hub genes with important biological functions.

Enrichment analysis for psychiatric diseases
Enrichment of psychiatric diseases for FER-related genes was conducted to assess the association between the FERrelated genes and the psychiatric conditions. Common genetic variants related to psychiatric diseases, including Alzheimer's disease (AD), attention-deficit disorder (ADHD), autism spectrum disorder (ASD), bipolar disorder (BP), major depression disorder (MDD), and schizophrenia (SCZ), were obtained from the Psychiatric Genomics Con-

Group-level activations in emotion task and other cognitive tasks
FER-related brain activation patterns were similar across ten groups under the face versus shape contrast and consistent with previous work in HCP (Barch et al., 2013). Positive activation areas mainly included the bilateral amygdala, middle occipital gyrus, middle temporal gyrus, superior frontal gyrus; negative activation was observed in the inferior parietal lobule, postcentral gyrus, and cingulate gyrus (FWE, P < 0.05) (Fig. 2). The ten group-level FER-related activation maps are shown in Fig. S1. The overlap regions of the ten group-level FER-related activations are shown in Fig. S2. The activation maps of other cognitive tasks (language, social, working memory, gambling, and relational tasks) are shown in Fig. S3.

Genes associated with FER-related brain activations
With a transcription-neuroimaging association analysis, 714 genes in group one, 706 genes in group two, 795 genes in group three, 584 genes in group four, 691 genes in group five, 752 genes in group six, 958 genes in group seven, 516 genes in group eight, 673 genes in group nine, and 724 genes in group ten were found significantly related to FER-related brain activation (P < 0.05, permutation test). Among these genes, 371 exhibited significant correlation in the same direction (positive, 213 genes; negative, 158 genes) in all the ten groups. The detailed information of FER-related genes in the ten groups was listed in Table S1. The representative scatter plots of the three top genes with significant positive and three top genes with negative correlations in group one were shown in Fig. S4.

FER-related genes were partly distinct from genes associated with other cognitive tasks
Using the same transcription-neuroimaging association analysis with the emotion task, 464 gamblingrelated genes, 2887 language-related genes, 659 working memory-related genes, 749 social-related genes, and 97 relational-related genes were identified, respectively (P < 0.05, permutation test). Among those genes, only 20 gambling-related genes, 167 language-related genes, 141 working memory-related genes, 127 social-related genes, and 12 relational-related genes overlapped with FERrelated genes. The detailed information of genes significantly related to other cognition-related brain activations is listed in Table S2. This analysis revealed that genes associated with FER were partly different from other cognitive functions.

FER-related genes are specifically expressed in brain tissue
Tissue-specific expression analysis was performed. FERrelated genes were significantly enriched in brain tissue (P = 2.32 × 10 −14 , BH-FDR corrected), blood vessel tissue (P = 1.21 × 10 −7 , BH-FDR corrected), and nerve (P = 4.60 × 10 −5 , Fig. 2 The group-level activation maps of group one in the facial expression recognition task. The color bar represents t-values of brain activation BH-FDR corrected) under a pSI threshold of 0.05 (Fig. 4). Brain tissue was the most significant, even under a more stringent pSI threshold of 0.001 (P = 2.38 × 10 −4 , BH-FDR corrected).

Discussion
This study carried out an across-sample transcriptome-neuroimaging spatial association analysis between the FERrelated brain activation and gene expressions in the human brain. A total of 371 genes were identified to associate with FER-related brain activation. Enrichment analyses showed that these FER-related genes are mainly overexpressed in the brain, especially neurons, and closely related to the cell junction, neuron development, and synaptic function. Furthermore, diseases enrichment analysis confirmed that FERrelated genes were associated with psychiatric diseases with emotion impairments. Consistent with previous works using the HCP dataset (Barch et al., 2013;Hariri et al., 2002b), we found that the FER task could robustly activate the emotion-related brain regions such as the amygdala in all ten groups. Moreover, the activation pattern was consistent with many previous studies on the expression recognition tasks (Fusar-Poli et al., 2009;Phan et al., 2002;Vytal & Hamann, 2010). These results demonstrated that emotion task in the HCP could stimulate reliable FER-related brain activation.
Although candidate gene studies and GWAS revealed some genetic loci associated with emotion, the studies usually provided limited insight into the transcriptional mechanism of emotion. An alternate method is the across-sample transcriptome-neuroimaging spatial association analysis which was widely used to investigate transcriptional mechanisms of cognitive activities and psychiatric disorders (Forest et al., 2017;Ritchie et al., 2018;Xie et al., 2020). It has been demonstrated that this method could reliably link neuroimaging maps from living brains with gene expression from postmortem brains (Berto et al., 2018;Forest et al., 2017;Hawrylycz et al., 2015).
To increase the reliability of the results, we randomly divided the 203 subjects into two groups, and this step was repeated five times. Finally, 371 genes that exhibited association with FER-related brain activation in the same direction in all ten groups were defined as the FER-related genes. We further compared the FER-related genes with the other cognitive task-related genes to evaluate whether 371 genes were specific to the FER. Our results indicated that FER-related genes were partly different from the other cognitive task-related genes, and FER shared some genes with working memory (141 genes), language (167 genes), and social cognition (127 genes). These results suggested that FER might have some specific regulatory genes and might also be modulated by some genes related to other cognitive functions. This explanation was supported by the previous studies that showed emotion had shared some similar neuroanatomical and neurophysiological basis with language, social cognition, and working memory. For instance, a study demonstrated a strong positive correlation between emotional competence and language competence in children (Beck et al., 2012). Emotion shared bilateral prefrontal function with working memory in an fMRI study (Mitchell, 2007). Robust correlations were found between social cognition skills and facial processing (Petroni et al., 2011).
Enrichment analyses revealed that FER-related genes were mainly overexpressed in the brain, especially neurons. GO enrichment showed that FER-related genes were mainly associated with anterograde trans-synaptic signaling, cell junction organization, chemical synaptic transmission, and neurodevelopment, including neuron projection development, neuron development, and nervous system development. And a couple of cellular components, including synapse, distal axon, and neuron projection, showed significant association. Those findings highlighted the importance of the genes modulating the synaptic functions, neuronal structures, and development in FER processing. For example, MET in the term nervous system development, which encodes Met Receptor Tyrosine Kinase, was known to affect facial expression perception (Lin et al., 2012) and could increase the risk for ASD (Abrahams et al., 2013;Campbell et al., 2006Campbell et al., , 2007Jackson et al., 2009). SHANK2 (H3-and multiple ankyrin repeats protein 2), a gene involved in ASD (Bavamian et al., 2015), is related to cell junction organization and chemical synaptic transmission. and CDH3. (C) Enrichment of FER-related genes for common psychiatric disorders. The size of each bubble represents the number of FER-related genes overlapped with genes associated with common psychiatric disorders. The x-axis represents -log10 (P, Bonferroni corrected). The y-axis represents odds ratio values. The red line represents P = 0.05 A PPI network was constructed using FER-related genes. The top five most connecting key genes were identified, including CTNNB1, CDH2, KRT19, EPCAM, and CDH3. Those genes may play an important role in emotion recognition. For instance, CTNNB1 in the term neuron development, which encodes catenin beta 1, is associated with fear-memory consolidation, ASD, and anxiety-related behavior (Dong et al., 2016;Maguschak & Ressler, 2008;Wang et al., 2017). A previous study suggested that CDH2, encoding cadherin 2, was associated with ADHD in humans and mice (Halperin et al., 2021) and could maintain migration and postmitotic survival of cortical interneuron precursors (László et al., 2020).
FER deficits have been observed in individuals with ASD (Harms et al., 2010), SCZ (Kohler et al., 2010;Kring & Elis, 2013), BP (Kohler et al., 2011), and MDD (Bourke et al., 2010;Demenescu et al., 2010). For example, Individuals with ASD have a limited repertoire of facial expressions (Harms et al., 2010;Yeung et al., 2020), especially negative emotions (Enticott et al., 2014). In this study, the FERrelated genes showed significant enrichment for the psychiatric conditions with heavy emotion impairments such as ADHD, ASD, SCZ, BP, and MDD. However, AD showed no significance. These results suggested that FER-related genes were associated with psychiatric diseases, which might help understand the mechanism of the FER impairments in these psychiatric disorders.
Several limitations should be mentioned in this study. First, transcriptome data was extracted from the left cerebral hemispheres of six postmortem adult brains in AHBA. The small sample size might bias our findings. Second, gene expression data and FER-related brain activation data were derived from the different subjects. Although gene expression patterns were confirmed to be conserved across individuals (Hawrylycz et al., 2015;Zeng et al., 2012), this influence could not be completely ruled out. To reduce this impact, we randomly divided neuroimaging data into two groups and repeated it five times. Only genes that exhibited consistently significant associations in the same direction in ten groups were defined as FER-related genes. Third, the task-evoked response magnitude across regions in this task had a relatively large individual difference (Miller et al., 2009, which might affect our results to some extent. Finally, causal effects between gene expression and FERrelated brain activation could not be clarified by this transcription-neuroimaging association analysis.

Conclusions
This study identified 371 significant FER-related genes using the transcriptome-neuroimaging spatial association analysis. Enrichment analyses revealed that FER-related genes were mainly enriched in the brain, especially neurons, and might be associated with cell junction organization, synaptic functions, and nervous system development regulation, indicating that FER was a complex polygenetic biological process involving multiple pathways. Moreover, these genes exhibited higher enrichment for psychiatric diseases with heavy emotion impairments. This study provided new insight into understanding the FER-related biological mechanisms and might be helpful to explore treatment methods for emotionrelated psychiatric disorders.