Tissue-dependent day-night variation in gene expression

We first used MetaCycle [16], a suite designed to analyze rhythmic data and that has been previously used to investigate circadian patterns (e.g., Ruberto and colleagues [17] and Mishra and colleagues [18]). While MetaCycle uses three methods to identify oscillating genes, only Lomb-Scargle is suited for GTEx data, due to the uneven distribution of the time of death. Using this method, from the 18,018 protein-coding genes expressed in at least one tissue (median transcripts per million (TPM) ≥ 1), we identified 187 (1%) that were circadian in at least one tissue (non-adjusted P ≤ 0.05; S1 Table). This number is much smaller than that previously reported in baboon [2], where 82% of the protein-coding genes had circadian patterns in at least one of the 64 studied tissues. Lack of power to detect genes with circadian gene expression patterns can be partially attributed to the characteristics of the GTEx resource. As reported, the GTEx metadata only includes the time of the day and the season of death, and there is no information about the location of death. Therefore, we focused only on the individuals in which the time of death could be unequivocally classified as either day [8:00 to 17:00) (351 individuals) or night [21:00 to 5:00) (315 individuals) and excluded those in which death occurred outside these intervals (uncertain data points, referred to hereafter as the twilight; 222 individuals) (Fig 1A). We then identified genes that were differentially expressed between day and night.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Day-night and seasonal classification of the GTEx samples.

(A) Samples from GTEx were classified as either day or night depending on the time of death of the donor, respectively, between 08:00 and 17:00 (351 individuals) and between 21:00 and 05:00 (315 individuals). (B) Samples from GTEx were assigned to the different seasons (spring, summer, fall, and winter), according to the reported season of death of the donor.

https://doi.org/10.1371/journal.pbio.3001986.g001

For this purpose, we carried out a differential gene expression analysis using voom/limma [19] controlling for the effects of season of death, sex, body mass index (BMI), age, and postmortem interval. We performed an initial comparison across tissues with high sensitivity, applying loose cutoffs of a non-adjusted P ≤ 0.05 and an absolute log₂ fold-change ≥ 0.1. From the 18,022 protein-coding genes expressed in at least one tissue (median TPM ≥ 1), we found that 12,530 (70%) were differentially expressed between day and night in at least one of the 46 tested tissues (day-night genes, Supp Dataset 1 - https://doi.org/10.6084/m9.figshare.21906252.v1), which is a number in line with the results found in baboon [2], with a Simpson index of similarity of 0.7 between our day-night genes and the baboon circadian genes (see Methods and S2 Table). Using a more stringent cutoff, with false discovery rate (FDR) ≤ 0.1, we identified 4,928 (27%) of genes as differentially expressed between day and night (Supp Dataset 1 - https://doi.org/10.6084/m9.figshare.21906252.v1).

Per tissue, 5.5% of expressed protein-coding genes were day-night on average, using the most sensitive settings (non-adjusted P ≤ 0.05 and an absolute log₂ fold-change ≥ 0.1). The tissue with the largest number of day-night genes was lung, with 2,418 genes (17.2% of genes expressed in this tissue; Fig 2A). Other tissues with a proportionally large number of day-night genes were the heart left ventricle (2,202 genes, 19.2%) and whole blood (1,900 genes, 19%). On the other hand, the tissue with the fewest differentially expressed day-night genes was salivary gland, with only 85 genes (0.63% of genes expressed in this tissue; Fig 2A). Other tissues with proportionally low number of day-night genes were colon transverse (92, 0.67%) and testis (105, 0.66%). Brain regions also showed a relatively small number of day-night changes (ranging from 0.86% to 7.8%; Fig 2A). Caudate was the brain region with the highest number of day-night genes (1,026 7.8%), followed by the cerebellum (766, 5.7%), which was reported to have a sleep stage–dependent activity [20]. Using FDR ≤ 0.1 as a cutoff for the differential expression analysis showed fewer differentially expressed genes for all tissues, but the tissues with the highest relative numbers were maintained (S1 Fig). Moreover, randomizations by swapping the times of death of individuals within each tissue (see Methods) showed that these relative patterns were not due to differences in sample size and/or the structure of samples within each tissue (S2 Fig), and the top tissues displayed a strong enrichment in the number of differentially expressed genes between day and night compared to the randomized samples. Finally, some tissues showed a bias in the number of genes overexpressed during the day (diurnal genes) versus the night (nocturnal genes) (Fig 2A and S3 Table). Stomach was the tissue with the strongest diurnal preference, while non-sun-exposed skin was the tissue with the strongest nocturnal preference. In contrast, sun-exposed skin showed diurnal preference (0.7 log₂ day/night genes ratio), suggesting that UV plays a role in the activation of gene expression.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Distribution of the day-night genes in the GTEx tissues.

(A) Number of genes found as day-night, i.e., genes differentially expressed between day and night (barplots; bottom x-axis), and log₂ ratio between the number of genes up-regulated during the day vs. the night (red dots; top x-axis) for each tissue. Tissues are sorted by the total number of day-night genes. (B) Yellow dots represent the number of tissues in which a given gene is classified as day-night (y-axis) vs. the number of genes that are classified as day-night in that number of tissues (x-axis, log₁₀ scale). Gene names for core clock genes are shown next to the dot with the corresponding number of tissues in which they are identified as day-night. Black diamonds show the number of high-confidence day-night genes per number of tissues (see S5 Fig for further details). (C) Expression of ARNTL in the thyroid in GTEx samples at the time of death of the GTEx donors (in hours). The colors of the dots represent the classification of the individuals according to their time of death: during the day (yellow), during the night (blue), or during twilight (grey). The samples classified as twilight were discarded for the day-night analysis. The “circadian” curve was created using the geom_smooth function from ggplot2 in R with the “loess” method. The data underlying this figure can be found in S1 Data.

https://doi.org/10.1371/journal.pbio.3001986.g002

Importantly, while our approach increases the power to detect significant changes in gene expression between day and night, it does not necessarily lead to the identification of all genes with circadian patterns of gene expression, since the circadian peaks may occur at the twilight time zones that we are ignoring. Consistently, most of the gene-tissue pairs detected by MetaCycle as circadian were also identified as day-night by our analyses (210/339, 62%; S1 Table), with most exceptions peaking at or around the twilight (S4 Table). To further understand the contribution of day-night variation to global circadian variation in gene expression, we implemented a sliding window approach in which we identified all genes that were differentially expressed between pairs of 6-hour window intervals separated by 6 hours (see Methods). Aggregating the results across all tissues, and despite the differences in the time interval definitions, the largest number of differentially expressed genes between two specific time windows was obtained for 7/8 AM to 1/2 PM versus 7/8 PM to 1/2 AM, highly overlapping with our day/night (S3A Fig). This suggests that day-night variation is, as expected, a major contributor to circadian-like variation in gene expression. Moreover, there was a strong and significant correspondence between the number of genes that were differentially expressed between at least one pair of time windows (to which we collectively refer as "circadian-like") and day-night genes across tissues (S3B Fig; Spearman rho = 0.56, P = 0).

Recurrent day-night variation in gene expression across tissues

We next computed the number of tissues in which a given gene showed a day-night pattern (Fig 2B). We found that, on average, day-night genes exhibited a day-night pattern in 2.6 tissues out of 39 tissues in which they had detectable expression (median TPM ≥ 1), indicating a rather tissue-specific response to day-night cycles. Overall, we found that genes identified as day-night in more than one tissue significantly tended to be consistently up-regulated in either day or night, and this consistency largely increased with the number of tissues in which genes were detected as day-night (S5 Table; see Methods).

We focused on a set of 16 genes that form the molecular core clock and are the main regulators of the circadian rhythm [21–23] (S6 Table). With the exception of RORC and RORB, the core clock genes were expressed in almost all tissues, and, with the exception of RORB, they showed differential day-night expression in multiple tissues (Fig 2B). This is in agreement with the hypothesis that one molecular clock program is present in all tissues but the circadian processes triggered downstream are highly tissue-specific [24]. Core clock genes usually have effect sizes (difference between day and night expressions) that are larger than those of non-core clock day-night genes (S4 Fig). ARNTL, a positive regulator of the core clock, was the gene with a day-night pattern in the largest number of tissues (33 tissues; Fig 2B). Its expression at the time of death, aggregated across the GTEx donors, nicely captures the known circadian behavior of this gene (Fig 2C), strongly indicating that, in spite of all the caveats associated with the data collection and available metadata, the GTEx data can be effectively used to investigate gene expression patterns during the day-night cycle. A recent study described ARNTL as the main regulator of the intertissue timekeeping function in mouse [25], and our results suggest that it may play a similar role in humans. Similarly, NPAS2 was the second gene with a day-night pattern in the largest number of tissues (31 tissues; Fig 2B and 2D) and has also been reported as being circadian in many mouse [26] and baboon [2] tissues.

Among the core clock genes, the thyroid had the highest number of day-night cases (12 out of 16 genes; Fig 3A). In contrast, we did not detect any of them as day-night in stomach, testis, and vagina, consistent with their overall low number of day-night changes (Fig 2A) and with previous studies in testis in rodents [27]. Clock genes showed a largely consistent pattern of day-night expression across tissues (i.e., they were either consistently up- or down-regulated at the same time of the day; Fig 3A). One exception was NR1D2, whose expression was consistently higher diurnally in brain subregions but nocturnally in all other tissues (Figs 3B and S5). For the core clock genes that showed day-night differences in both human and baboon in any of the available 20 homologous tissues (S2 Table), we found that most orthologs had a similar behavior in both species (59 similar versus 15 opposite gene-tissue pairs; P = 1.277 × 10⁻⁷ one-sided binomial test; Fig 3C), consistent with their shared diurnal regimes. In contrast, core clock genes in mouse (a preferentially nocturnal mammal) largely showed the opposite behavior than their human counterparts (13 similar versus 38 opposite gene-tissue pairs; P = 3.105 × 10⁻⁴ one-sided binomial test; Fig 3C and S7 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Day/night differential expression of the core clock genes in human, baboon, and mouse.

(A) Day/night differential expression of the 16 core clock genes in GTEx tissues. Cells in the matrix are colored according to the log₂ fold-change obtained with the voom-limma pipeline, from yellow (day) to blue (night). Genes without significant effects were colored in grey. In addition, we labelled each gene (“Main gene time”) as day (night) if it was up-regulated in more tissues during the day (night) than during the night (day). We labelled the tissues similarly (“Main tissue time”) depending on the number of genes that were up-regulated during the day (night) on that tissue. One gene, RORA, and one tissue, the aorta, were up-regulated in the same number of times during the day and the night and have not been labelled. Genes and tissues have been sorted according to (i) their main time and (ii) their number of significant effects. (B) Expression of NR1D2 in the artery—tibial (red) and in the brain—cortex (yellow) GTEx samples at the time of death of the GTEx donors (in hours). Curves were obtained from the Z-score of the expression TPMs using the geom smooth function from ggplot2 in R with the “loess” method. (C) Comparison of human and baboon (left) and human and mouse (right) orthologous core clock genes in common tissues. Only gene-tissue pairs that are significant for both compared species are shown. Cells are separated into two with (i) on the bottom left, the day-night classification of the gene in the tissue in human, and (ii) on the top right, the baboon/mouse phase obtained from Mure and colleagues [2] or Li and colleagues [26], respectively. (D) Top 10 enriched Gene Ontology Biological Process annotations for day and night high-confidence genes separately. See Methods for details. The data underlying this figure can be found in S1 Data.

https://doi.org/10.1371/journal.pbio.3001986.g003

Next, we aimed at identifying a set of “high-confidence day-night genes” from which to obtain further functional insights. Although the highly sensitive approach that we implemented above is well suited to reveal tissue-level patterns, it might be oversensitive to generate reliable individual gene-level insights. Therefore, we defined as high-confidence day-night genes a set of 445 genes with a consistent day (282) or night (165) pattern in multiple tissues (P ≤ 0.05 two-sided binomial test across all tissues, brain regions, and/or across non-brain tissues; see Figs 2B and S6 and S8 Table and Methods). High-confidence day-night genes included most clock genes (with the exception of TIMELESS, RORA, and RORB) and were significantly enriched for circadian and photoperiodism-related Gene Ontology (GO) terms (Fig 3D and S9 Table). Moreover, this gene set significantly overlapped those annotated as circadian in the Circadian Gene DataBase in humans [28] (87 genes, P = 3.074 × 10⁻⁸ one-sided Fisher test) and substantially expands the set of genes known to be varying during the day-night cycle. Additional GO terms were enriched among high-confidence day-night genes, including apoptosis and cell cycle regulation (genes peaking during the day) or fatty acid metabolism, cellular respiration, and various signaling pathways (genes peaking during the night) (Fig 3D and S9 Table). An example of high-confidence day-night gene with previously unknown circadian variation is THRA (S7A Fig), a thyroid hormone receptor that peaks at night in 15 tissues (including the thyroid), the expression of which was found disrupted in the hypothalamic structures of rats in constant darkness or lighting [29]. Other examples of high-confidence day-night genes present in a large number of tissues are the ribosomal protein RPS26 (day in 16 tissues), the nuclear pore complex protein NPIPB5 (day in 16 tissues), and the transcription corepressor TRIM22 (day in 13 tissues) (S7B–S7D Fig).

One of the physiological signatures of the day-night rhythmicity is the sleep–wake cycle. Therefore, we next focused on a set of 254 protein-coding genes expressed in GTEx and that were previously reported to increase the risk of insomnia or to be associated with other sleep traits in humans [30] (S10 Table). From this gene set, 186 (73.2%) exhibited a significant day-night variation in at least one tissue, closely matching the genome-wide behavior across tissues (S8 Fig). Moreover, none of them belonged to the core clock, and they did not significantly overlapped with our set of high-confidence day-night genes (7 genes in common; P = 0.49, one-sided Fisher test) or with those annotated as circadian in the Circadian Gene DataBase in human [28] (P = 0.31, one-sided Fisher test over expressed genes). These results thus suggest that genes involved in sleep traits do not appear to be particularly impacted by circadian or day-night gene expression patterns. However, four sleep genes were high-confidence day-night (S9 Fig) and annotated as circadian in the Circadian Gene DataBase [28] either in human (PC) or in other species (PITPNC1 and PDE4B in mouse and QSOX2 in Arabidopsis thaliana), pointing at a potential relevance, in specific cases, on day-night regulation of sleep behaviors. In contrast, we find a substantial overlap with genes involved in circadian rhythm sleep disorders using MalaCards [31]. Specifically, we found that 11/34 genes associated with advanced sleep phase disorder and 10/25 associated with delayed sleep phase disorder are part of the high-confidence day-night dataset. While 10 of these correspond to core clock genes, we also find one additional candidate, CATSPER2, in our high-confidence day-night gene set.

Tissue-dependent seasonal variation in gene expression

Next, we analyzed the seasonal variation of gene expression. To make the analyses consistent with those of day-night patterns and to minimize the impact of GTEx reporting only the season and not the actual day of death (see Methods), we set out to identify season-specific genes, i.e., genes that are differentially more highly or more lowly expressed in one particular season versus the others. This minimizes some of the impact of measurements taken around the seasonal boundaries. Including as covariates day-night variation as well as sex, BMI, age, and postmortem interval, and using a comparable definition of differential gene expression (raw P ≤ 0.05 and absolute log₂ fold-change ≥ 0.1; see Methods), we found that 16,408 (91.1%) of all expressed protein-coding genes were differentially expressed in at least one season in at least one tissue (hereafter seasonal genes; Supp Dataset 2 - https://doi.org/10.6084/m9.figshare.21906255.v1). There were no large differences in the number of seasonal genes across seasons, ranging from 12,026 genes in summer to 13,192 in fall. Per tissue and per season, the average number of seasonal genes were similar among seasons and comparable to day-night patterns (5.3% in spring, 4.9% in summer, 6.5% in fall, and 5.3% in winter, compared with 5.5% day-night genes), but there were more unique seasonal genes than day-night genes per tissue when all seasons were considered together (17.7%), as expected from having four versus one comparisons. The effect sizes of seasonal genes were also similar to those observed in day-night genes (S10 Fig).

As in the case of the day-night variation, the goal of this first high-sensitivity analysis was to reveal differences and similarities between patterns of seasonal variation across tissues. In this regard, in stark contrast to day-night patterns, the tissues with the highest proportion of genes showing seasonal changes included testis, skin, and multiple brain subregions, among others (Figs 4A and S11). As in the case of day-night variation, the tissue specificity of these seasonal patterns was strongly supported by randomizations (S12 Fig) and using FDR ≤ 0.1 as a cutoff for the differential expression analysis (S13 Fig). Most of these tissues did not show a clear bias in the direction of the expression changes (i.e., the number of season-specific up- or down-regulated genes was similar; S11 Fig and S11 Table). Testis, however, exhibited a massive gene up-regulation in fall and down-regulation in spring.

Although the overlap between seasonal and day-night genes was statistically significant for most tissues, most differentially regulated genes were unique to one or the other type of variation (S12 Table). Moreover, consistent with the distinct set of top varying tissues, the number of day-night and seasonal genes across tissues did not correlate (Spearman’s rho = −0.12; Figs 4B and S14). Tissues with larger day-night than seasonal variation included various tissues from the thoracic cavity (e.g., lung and the heart’s left ventricle), which may reflect changes in heart rate and breathing patterns between day and night [32]. On the other hand, tissues with more seasonal than day-night genes included most brain subregions and gonadal tissues, likely mirroring the involvement of brain-gonadal axis in regulating seasonal physiology and behavior [33] (Fig 4C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Distribution of the seasonal genes in the GTEx tissues.

(A) Number of unique genes found as seasonal (x-axis), i.e., genes differentially expressed in at least one season when compared to the others, per tissue (y-axis). The numbers in parentheses represent the percentage of unique seasonal genes in a given tissue over the number of expressed genes in that tissue. (B) Number of unique seasonal vs. day-night genes per tissue. Statistics from a Spearman’s correlation is shown. Tissue colors correspond to the GTEx color panel. (C) log₂ ratio between the number of day-night genes vs. seasonal genes for each tissue and season separately. The higher the log₂ value, the more day-night genes compare to the number of seasonal genes. Tissues were clustered using Euclidean distance and the Ward’s method. The data underlying this figure can be found in S1 Data.

https://doi.org/10.1371/journal.pbio.3001986.g004

Recurrent seasonal variation of gene expression across tissues

On average, seasonal genes showed seasonal expression in a number of tissues comparable to that of day-night genes: 2.5 in spring, 2.5 in summer, 3 in fall, and 2.5 winter, compared with 2.6 tissues for day-night genes. These genes also tended to be consistently up- or down-regulated across tissues. Therefore, as per day-night variation, we defined sets of high-confidence seasonal genes that varied in a consistent manner across multiple tissues for each season separately (S13 Table; see Methods). In total, we identified 1,748 unique genes: 308 in spring (138 up and 170 down), 361 in summer (158 up and 203 down), 1,072 in fall (691 up and 381 down), and 322 in winter (89 up and 233 down) (Fig 5A and S13 Table). The top enriched gene functions were largely specific for individual seasons, although some regulatory categories such as transcription and translation were shared by multiple sets (Fig 5B and S14 Table). As expected, multiple immune-related terms were enriched in fall and winter. Moreover, 122 high-confidence day-night genes were also among the high-confidence seasonal gene sets (Simpson index = 0.27; S15 Table), including three core clock genes: PER3 (fall, summer, winter), PER1 (winter), and NR1D2 (fall).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Seasonal genes and associated functions.

(A) Distribution of the number of tissues (x-axis, log₁₀ scale) in which genes were classified as seasonal (green: spring, red: summer, light blue: fall, dark blue: winter). The names of the genes found as seasonal in a large number of tissues for a given season are depicted, and the font color corresponds to their respective season. Black diamonds show the number of high-confidence seasonal genes per number of tissues. (B) Enriched Gene Ontology Biological Process annotations for up and down high-confidence seasonal genes separately per season (green: spring, red: summer, light blue: fall, dark blue: winter). Top terms with a raw P < 0.005 are shown; see Methods for details. (C) Gene Ontology enrichment of Biological Processes of the set of 192 strongly seasonal genes as computed by Enrichr [77,78]. (D) Seasonal log₂ fold-change of the five hormone-coding genes included in at least one high-confidence seasonal gene set. Genes without significant effects are colored in grey. The data underlying this figure can be found in S1 Data.

https://doi.org/10.1371/journal.pbio.3001986.g005

An interesting example of high-confidence seasonal gene is GLTSCR1, a component of the SWI/SNF chromatin remodeling complex also known as BICRA, which increased in fall in 16 tissues (e.g., in cerebellar hemisphere; inset of Fig 5A). Similarly, RTF1, a component of the RNA polymerase II transcription-associated PAF1 complex, decreased in fall in nine tissues (S15A Fig). This complex is deeply conserved across eukaryotes, and it has been described to be involved in the regulation of flowering time in plants [34,35] and to be required for induction of heat shock genes in animals [36,37]. PAF1c has been proposed to establish an antiviral state to prevent infection by incoming retroviruses: in case of infection by influenza A strain H3N2, PAF1c associates with viral NS1 protein, thereby regulating gene transcription [38]. Other examples include C4A, decreasing in spring in 21 tissues (S15B Fig), which localizes to the histocompatibility complex, and KRT1, which decreases in 24 tissues in winter (S15C Fig) and encodes a keratin gene that is a downstream effector of the corticotropin hormone release pathway [39]. Krt1 mutations in mouse have been associated with various phenotypic defects, ranging from abnormal circulating interleukin [40] to aberrant pigmentation of the epidermis [41]. Seasonal pigment variation is well known in mammals from the northern hemisphere [42], which might suggest a conserved role of the corticotropin release pathway in pigment seasonal variation.

Finally, we focused on a set of 192 genes (1.1% of all expressed genes) that exhibited the strongest quantitative seasonal expression differences (at least a two fold-change in expression in one tissue-season pair). These were usually highly tissue specific, since only seven of these genes belonged to the set of 1,370 genes with recurrent seasonal patterns. These genes were enriched for functions related to epidermal differentiation and immunity (Fig 5C), consistent with previous results in blood cells [14]. In line with the enrichment for multiple immune-related functions, these genes also exhibited significant overlap with gene sets that have been associated with Coronavirus Disease 2019 (COVID-19), including genes whose expression changes upon Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) infection and genes predicted to be functionally related to ACE2 (S16 Fig), although ACE2 itself did not show a strong seasonal pattern. In particular, 24 out of the top 200 genes among the latter predictions (12.5%) were strongly seasonal, with 20 of them being up-regulated in the intestine specifically in the winter.

Seasonal variation of gene expression of hormone genes

Hormones [43,44] have been described to broadly regulate the body’s seasonal physiology [7–9]. Thus, to explore whether genes that encode peptide hormones undergo particularly strong seasonal changes, we used a list of 73 genes with hormone-encoding capability based on Mirabeau and colleagues [45] (S16 Table). First, we compared the mRNA expression patterns of genes encoding pituitary hormones in the pituitary with the actual hormone levels, measured in blood, reported by Tendler and colleagues using medical records [43]. In general, we found a good correspondence between mRNA and hormone levels, despite the marked different approaches, especially for those peaking in summer (i.e., POMC [ACTH] and TSHB [TSH]) (S17 Fig). The main exception was FSHB [FSH], whose hormone levels had a low variation range and different behaviors in males and females [43].

Next, we assessed seasonality across all hormone genes and found 42 genes (58% of all hormone genes) to be seasonal in at least one tissue (S18 Fig), a significant depletion respect to the whole genome (P = 3.27 × 10⁻⁵, two-sided proportion test). Nevertheless, this depletion may be expected given the high degree of tissue-specific expression of most hormones (S19 Fig). Despite this, five hormone genes were included in at least one high-confidence seasonal gene set (Fig 5D). Among these, POMC, encoding the well-known photoperiodic hormone ACTH, was seasonal in 26 tissue-season pairs, being mainly up-regulated in summer, as in its main tissue of expression, the pituitary (see above). POMC expression has also been shown to be dependent on longer-term photoperiod in the Siberian hamsters [46]. Other seasonal hormones also have well-known roles, mainly in the cardiovascular system and growth: UCN, a corticotropin for stress response and appetite regulation [47,48], and ADM, which is important for vasodilation [49]. GNRH1, the gonadotropin-releasing hormone 1, which influences seasonal changes in other mammals [50], is seasonal in two artery tissues: the aorta and the tibial artery. Mutations in GNRH1 have been shown to be related to ischemic heart disease, which shows a seasonal pattern [51,52]. Interestingly, leptin (LEP), a hormone involved in seasonal food-seeking behavior, thermoregulation [53], and obesity [54,55], was regulated only in winter in three tissues: adipose visceral, nerve, and blood (S18 Fig).

Seasonal changes in brain cytoarchitecture

Various studies have shown seasonal histological variation in different brain regions from several mammals, including anterior cingulate cortex in shrews [56], dendritic spines in amygdala in response to short days (i.e., in fall) in white-footed mice [57], and the volume of suprachiasmatic nucleus in humans [58]. To investigate potential seasonal changes in human brain’s cytoarchitecture, we used gene expression profiles of cell type–specific markers for a variety of brain cell types [59]: astrocytes, neurons, oligodendrocytes, microglia, and endothelial cells (S17 Table). We found that astrocyte markers significantly increased their expression in fall and decreased in summer (Fig 6; see Methods for statistical analysis). In particular, we observed significant increased expression of astrocyte markers in the hypothalamus and frontal cortex in the fall and a decrease in the cerebellum and frontal cortex in the summer. Oligodendrocyte markers, in contrast, tended to decrease expression in the fall, in particular, in nucleus accumbens and anterior cingulate cortex (Fig 6). In winter, all oligodendrocyte markers increased their expression in nucleus accumbens. Moreover, we observed a significant increase in the expression of neuronal markers in winter, particularly in the hypothalamus and spinal cord, and a global decrease in the fall (Fig 6). Similar enrichments were obtained using a total of 627 gene markers for 16 cell type clusters from the suprachiasmatic nucleus [60] (S20 Fig). Overall, these results are thus consistent with previous histological studies in humans and other mammals [56–58] that showed that the relative volume or cytoarchitecture of astrocytes, oligodendrocytes, and neurons change with the season in a subregion-specific manner.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Seasonal variation in cell type–specific markers across human brain regions.

Percentage of the marker genes down-regulated (blue) or up-regulated (red) for five brain cell types: astrocyte, endothelial, microglia, neuron, and oligodendrocytes, in brain subregions (depicted in the brain scheme above, downloaded and modified from the GTEx web portal). All significant markers for a cell type for a specific season in a subregion were differentially expressed in the same direction, over- or underexpressed. Markers without significant effects were colored in grey. P values were obtained by randomizing the observed number of up- and down-regulated markers per cell type across regions and seasons conservatively allowing only up- or down-regulated makers for a given region–season combination (see Methods for details). *** P < 0.001, ** 0.001 ≤ P < 0.01, * 0.01 ≤ P < 0.05. The data underlying this figure can be found in S1 Data.

https://doi.org/10.1371/journal.pbio.3001986.g006

Potential impact of day-night and seasonal variation on drug targets

There is a growing body of evidence that circadian rhythms impact drug efficiency, including variation in the expression of targets and of transporters [61]. For example, appropriate daily timing of anticancer drugs has been suggested because of circadian regulation of cell cycle within tumors [62]. Similarly, administration of drugs to lower cholesterol levels has been proposed to adapt to the diurnal variation of this lipid [63]. Here, we have compared our set of 445 high-confidence day-night genes with the gene targets for known drugs in the Drug–Gene Interaction database (DGIdb 4.0; [64]). We found that 91 (20%) of these genes are targeted by 1,077 drugs (see S18–S20 Tables). Less is known about the impact of seasonal variation in drug administration. We have found that 307 high-confidence seasonal genes (18%) were present in DGI targeted by 2,632 drugs (see S21–S23 Tables). Interestingly, among the 18 drugs targeting at least 10 high-confidence seasonal genes, we observed 11 anticancer drugs, including bortezomib (16 target genes), cisplatin (15), sorafenib (12), doxorubicin (12), doxorubicin (12), and ME-344 (12) (in clinical trial).

RNA-seq datasets

The RNA-seq data were generated, mapped, and quantified by the GTEx consortium (GTEx v8) [15]. Tissues with less than 100 donors were discarded from the analyses (Kidney—Medulla, Kidney—Cortex, Cervix—Ectocervix, Fallopian Tube, Cervix—Endocervix, and Bladder) as well as two cell lines (Cells—EBV-transformed lymphocytes and Cells—Cultured fibroblasts). For the Whole Blood tissue, all premortem samples were discarded from the analysis for homogeneity with the other tissues. After filtering for samples with available covariates for the differential analyses, we employed 16,151 samples from up to 46 tissues for 932 individuals. GTEx individuals are biased toward old males (median age = 55 years, 67% male). GENCODE v26 [73] was used for GTEx as well as the annotation for the protein-coding genes.

Metacycle

The function meta2d from the R package MetaCycle (v.1.2) was runned for each 46 tissues separately using the default parameters. Genes with a median TPM < 1 per tissue were filtered out and the time of death of the individuals were converted into hours.

Classification of the time of death as day or night and by season

Circadian patterns had been previously analyzed using the GTEx data, but inferring the time of death from the expression of marker genes [3]. Here, we use, instead, the actual time of death as provided by the GTEx consortium. Using this time of death, individuals had been classified between dead during the day or dead during the night if their time of death was falling into the following intervals [08:00 to 17:00) and [21:00 to 05:00), respectively (Fig 1A). Other times of death have been discarded for the day-night analysis to avoid taking into account any RNA-seq samples coming from people where the day-night status was unsure or uncertain, given the variation in day-night times along the year. We refer to these uncertain points as twilight. Using this classification, 351, 315, and 266 individuals have been classified as day, night, and twilight, respectively, of which only the 666 day and night individuals (11,527 samples) were used for differential expression for the day-night cycle. The time corresponding to day and night was manually curated from the Boston (Massachusetts) sunrise and sunset intervals during the year (using https://www.timeanddate.com/sun/usa/boston). The season of death was provided by the GTEx consortium with 190, 221, 282, and 239 donors that died in spring, summer, fall, and winter, respectively. Note that for the season, no people were discarded regarding the time of death. We also provide the information about the month in which the RNA was extracted for each season of death (S17 Table).

Differential gene expression between day and night

Differential expression between day and night was performed separately on the 46 tissues using samples from the 666 individuals classified as day or night, ranging from 98 samples (Uterus) to 560 samples (Muscle—Skeletal). Genes were filtered per tissue, removing all genes with a median TPM < 1 over the day and night samples, leading to 18,022 protein-coding genes expressed in at least one tissue (31,530 genes including all biotypes). The analyses were run using R v3.6.1, the TMM normalization method from edgeR (with the calcNormFactors function) [74,75], and the voom-limma pipeline (with the voom, lmFit, and eBayes functions) [19,76] using default parameters. The significance of the time of death was assessed correcting for the following covariates: sex if the tissue was not sex-specific; age; BMI; the postmortem interval; the season of death. All genes with an associate P ≤ 0.05 and an absolute log₂ fold-change ≥ 0.1 were considered as day-night. Results, including all biotypes, are available in Supp Dataset 1 - https://doi.org/10.6084/m9.figshare.21906252.v1 (including P values, adjusted P values, and log₂ fold-changes). The number of differentially expressed genes across experiments was not affected by the number of samples (S21A Fig), and no effect was observed regarding the ratio of day/night samples and the number of differentially expressed genes between day and night (S21B Fig).

Definition of circadian-like genes using a sliding window approach

To identify circadian-like genes, we set up pairs of 6-hour sliding windows separated by 6 hours, i.e., [i,i+6) and [i+12,i+18) for i = 0…11, and we performed a differential gene expression analysis between both intervals for each i. This differential expression analysis was done using the same method as the one used to define our day-night genes, except that we adjusted the nominal P value per gene per tissue for the number of windows (n = 12) using the FDR procedure.

Season-specific gene expression

Differential expression between seasons was performed separately on the 46 tissues using samples from the 932 individuals, ranging from 139 samples (Brain—Substantia nigra) to 789 samples (Muscle—Skeletal). Genes were filtered as described in the section “Differential expression between day and night.” This resulted in a set of 18,018 protein-coding genes expressed in at least one tissue (31,517 genes including all biotypes). The effect of each season was assessed by comparing one season against all the others, leading to four differential expression analyses, one for each season (see “Differential expression between day and night” for details). This approach was taken to make the analyses more consistent with that of day-night variation (i.e., using the exact same method) and to minimize the impact of the fact that only the season, but not the actual day, of death is known in GTEx (i.e., two individuals could have died one day apart but in two different seasons, or 90 days apart but still within the same season). The covariates used were the following: time of death classified as day, night, or twilight; sex if the tissue was not sex-specific; age; BMI; postmortem interval. All genes with an associate P ≤ 0.05 and an absolute log₂ fold-change ≥ 0.1 were considered as seasonal. Results, including all biotypes, are available in Supp Dataset 2 - https://doi.org/10.6084/m9.figshare.21906255.v1 (including P values, adjusted P values, and log₂ fold-changes). The number of seasonal genes was not related to the number of samples per the tissue (S22A Fig), and no relation was observed between the proportion of samples available for a given season over the total and the ratio of up- and down-regulated genes. (S22B Fig).

Testing tissues for day or night

To test if day-night genes in two tissues or more are more prone to be either day or night, we computed for each gene the ratio of day tissues over the total number of tissues for this gene. Genes will be considered to have a “consistent” ratio if their ratio is <0.25, mainly night tissues, or >0.75, mainly day tissues. For each number of tissues, from two to ten tissues, we calculate the percentage of genes that are considered to have a consistent ratio and the expected probability to be in the consistent ratio distribution following a binomial distribution with a probability of success (day tissue) equal to 0.5 and a number of draws equal to the number of genes that are day-night for this number of tissues. We then performed a binomial test for each number of tissues. The results are available at S5 Table.

Day-night and season labels permutation

For each tissue, the day, night, and twilight labels were shuffled before any filtering, preserving the original number of each label. The twilight samples were then removed and the genes were filtered according to their median TPM (median TPM ≥ 1). The differential gene expression was performed as described above in “Differential gene expression between day and night.” The permutation was done 1,000 times per tissue. A similar method was performed for the seasonal genes by permuting, per season, the labels of the season of interest and the other seasons before the median TPM filtering. The differential gene expression was performed as described above in “Season-specific gene expression.” The permutation was done 1,000 times per tissue per season. To obtain the empirical cumulative distribution function for protein-coding genes with a non-adjusted P ≤ 0.05 (or a non-adjusted P ≤ 0.05 and an absolute log₂ fold-change ≥ 0.1), the proportion of time these genes have a non-adjusted P ≤ 0.05 (or a non-adjusted P ≤ 0.05 and an absolute log₂ fold-change ≥ 0.1) over the 1,000 permutations was computed, with all tissues pulled together. To compare the number of protein-coding genes that are defined as day-night or seasonal between the original data and the 1,000 permutations, the median per tissue (or per tissue and season) was computed (S2 and S12 Figs).

Definition of the high-confidence gene sets

To define the five high-confidence gene sets (one for day-night genes and four for the seasonal genes), we perform three binomial tests per gene: (i) using all tissues; (ii) using only non-brain tissues; and (iii) using only brain regions. The number of draws was the number of tissues in which the gene was day-night or seasonal and the number of successes the number of tissues in which the gene was day or up. Only two of the tested genes have divergent behavior between non-brain and brain tissues: NR1D2 (11 night non-brain tissues and 9 day brain tissues) and RGS1 (6 day non-brain tissues and 6 night brain tissues). No divergent behavior was observed for the seasonal genes.

Comparison of human, baboon, and mouse core clock genes

The baboon results were downloaded from Mure and colleagues [2]. We extracted the significant genes using the same threshold as Mure and colleagues (P ≤ 0.05). Common tissues between GTEx and baboon have been manually curated (S2 Table). The mouse circadian genes were downloaded from the CirGRDB database [26], which include genes already defined as circadian, and we only selected the genes found by RNA-seq and from the publication with the PubMed ID 25349387, corresponding to the publication by Zhang and colleagues [1]. Common tissues between GTEx and the mouse tissues list had been manually curated (S7 Table). This dataset was used to compute the number of occurrences of Clock and Npas2 in mouse tissues.

To perform the comparison of the core clock genes in human versus baboon, we extracted the core clock genes in both species in the 20 common tissues between the GTEx and the baboons’ tissue. From this, we analyzed only genes that were significant in both species, i.e., we discarded the genes found significant in either baboon or human but not in both. The same analysis was done to compare human and mouse.

Functional enrichment analyses

The high-confidence sets for day-night and seasonal genes, separated by day/up- and night/down-regulated genes, were used as input for the enrichment analysis tool Enrichr [77,78]. Results for GO Biological Process were retrieved and are provided fully in S9 (day-night) and S14 (seasons) Tables. To simplify visualizations in Figs 3D, 5B and 5C, we disregarded GO terms with fewer than 10 total genes. In addition, we removed redundant categories by excluding those terms sharing at least 80% of genes with a category with a lower P value. For COVID-19-related gene sets, gene sets with an adjusted P value greater than 0.05 were discarded.

Circadian and hormone gene sets

The protein sequences, along with the IDs, of proteins annotated as circadian, either experimentally or by orthology, in human were downloaded from the Circadian Gene DataBase website (http://cgdb.biocuckoo.org/) [28] the 01/26/2021. The correspondence between the Ensembl protein ID and the UniProtKB ID with the Ensembl gene ID was downloaded using BioMart from Ensembl [79]. The list of hormone genes was downloaded from Mirabeau and colleagues [45]. The Ensembl peptide IDs were linked to their respective Ensembl gene ID using R and the biomaRt package [80,81]. The gene IDs for hormones with deprecated peptide IDs were retrieved manually using the Ensembl website (http://www.ensembl.org) and the ones that were obsolete were removed. Two hormone genes were added manually: GH1 and LEP. The list is available at S16 Table.

Statistical assessment of cell type–specific marker changes across seasons

For each cell type, we evaluated a fixed number of markers (Astrocyte: 10, Endothelial: 10, Microglia: 9, Neuron: 11, Oligodendrocytes: 10) from [59] (S17 Table). We assessed the significance of the number of up- or down-regulated markers for each cell type in each specific region-season pair (each cell in Fig 6) as well as for all regions together (top row in Fig 6). For this purpose, we first counted the total number of instances in which markers for a given cell type are up- and down-regulated across the entire set of region-season pairs (Astrocyte: 42 up, 36 down; Endothelial: 12 up, 20 down, Microglia, 9 up, 8 down; Neuron, 27 up, 12 down; Oligodendrocytes, 18 up, 27 down). Then, we performed 1,000,000 randomizations of the up and down instances across seasons and regions. We conservatively allowed only up or down markers in a given region-season pair for each cell type, since this is what we have observed in the real data (i.e., no cell type markers showed contradictory patterns in any region-season pair). To calculate the P value for each region-season pair, we simply obtained the number of randomizations in which the number of up- or down-regulated markers was equal or higher than the tested region-season pair and divided it by 1 million. These P values were Bonferroni corrected for multiple testing (260 tests). To calculate the P value across regions, we sum all up and down instances across regions and again obtained the number of randomizations in which the number of up- or down-regulated markers was equal or higher than the sum across regions and divided it by 1 million. These P values were also Bonferroni corrected for multiple testing (20 tests).