Introduction

Endothelial cells (ECs) make up the lining of blood and lymphatic vessels. These cells have critical functions in vascular homeostasis, such as regulating angiogenesis, permeability, thrombosis, and vascular tone1. Dysfunction of these cells is seen in the beginning stages of a broad range of disease states like atherosclerosis, inflammation, kidney failure, etc2. Therefore, it is important to study the different states of ECs, e.g. healthy versus dysfunctional, to fully understand the pathogenesis of those disease states and to identify novel therapeutic and/or diagnostic targets.

Docosahexaenoic acid (DHA), an omega-3 polyunsaturated fatty acid (PUFA) primarily found in marine sources, is well-known for its cardioprotective effects and its anti-inflammatory properties3. In many primary EC lines and EA.hy926 cells, DHA improves the activity of endothelial nitric oxide synthase (eNOS)4,5,6, a key gate-keeper of endothelial function7. As a PUFA, DHA is also a known ligand for peroxisome proliferator-activated receptors (PPARs)8, nuclear receptors that can affect the transcription of many genes.

Real time-quantitative polymerase chain reaction (RT-qPCR) is regarded as the gold standard for sensitive gene expression quantification and is usually needed for validating RNA sequencing (RNA-seq) results. However, its accuracy, and specifically, the more commonly used relative quantification methods, depend on stable internal controls, i.e. reference genes9. Due to poor reporting of RT-qPCR methods and results historically, the Minimum Information for Publication of Quantitative Real-Time PCR Experiments (MIQE) guidelines were created to increase reliability and repeatability of results10. The MIQE guidelines mandate that reference genes be experimentally validated rather than assuming they are invariant10. Inappropriate conclusions may be made if an inadequate reference gene is used, which underscores the importance of appropriate selection and validation of reference genes11. Housekeeping genes (HKGs) refers to genes constitutively expressed to sustain fundamental cellular functions and are usually present ubiquitously across all cell types and conditions. Historically, HKGs were used as reference genes for expression studies, but now it is recognized that traditionally used HKGs do not always exhibit stable expression9. For simplicity, in this paper, we will use the term “HKGs” throughout.

Commonly used HKGs, including GAPDH, RNA18S, ACTB, and others, have shown variation across endothelial cell types, and in response to various stimuli and culture conditions12,13,14,15. One study of Busulfan‑injured microvascular EC found that YWHAZ and ALAS1 were more stable under these conditions than GAPDH and ACTB12. In another study using human umbilical vein endothelial cells (HUVECs) under stress from homocysteine, GAPDH was found to perform better than ACTB, though further validation is warranted13. Also, existing studies create an incomplete patchwork of validated HKGs for ECs. Currently, no work has systematically investigated HKGs for ECs in different growth states, or the healthy versus dysfunctional in vivo context. This gap limits the reliability of qPCR studies across vascular biology.

The intent of this study was to identify and validate a set of HKGs that exhibit stable levels of expression in ECs in different growth states and in response to DHA treatment. The greater the degree of stability the better suited an HKG would be as a reference gene under these varying culture conditions. Using transcriptomic data from a human endothelial cell line, EA.hy926, cultured to different growth states without and with exposure to DHA, a list consisting of 18 candidate stable genes was identified. Together with 8 other commonly used HKGs and 1 from the HRT Atlas database, the expression of these 27 candidate HKGs was validated by RT‑qPCR and supplied to 5 widely used gene-expression-stability-assessing algorithms: deltaCt16, BestKeeper17, geNorm18, NormFinder19, and RefFinder20. Based on the comprehensive ranking provided by these 5 algorithms, CAPZB, FBXO7, and SMU1 were superior to commonly used RNA18S and ACTB. Their effect as HKGs, in combination or alone, was validated by testing their utility in relation to the expression of selective endothelial markers under different growth states and DHA treatment.

Results

Selection of candidate HKGs

The candidate HKGs were selected from the RNA-seq data obtained from EA.hy926 human endothelial cells grown on Matrigel (MG) for 4 days (the growing state) or 10 days (the quiescent state) and either harvested as such or after treatment for 8 h with 20 µM DHA. Genes expressed at a high level (DESeq2-normalized counts >100) and not differentially expressed between the different conditions (growth state ± DHA treatment) were ranked by either coefficient of variation (CV) or the dispersion value acquired from DESeq2 analysis. The top 20 genes with lowest CV or dispersion were shortlisted, resulting in a merged list of 33 genes. As we have shown previously, EA.hy926 cells in different growth states respond differently to 20 µM versus 125 µM DHA21. Therefore, genes that showed differential expression in response to 125 µM DHA treatment were excluded. Completion of this process left 18 genes from the list as candidate HKGs for follow-up validation, along with 1 gene selected from the RefEx and HRT Atlas databases (RHOA) and 8 commonly used HKGs (Table 1).

Table 1 List of candidate reference genes. The top 20 candidates were selected from our RNA-seq results based on CV and/or dispersion from the DESeq2 package. Some reference genes (identified from databases) recommended for both aorta and/or vein and/or HUVEC were also included.
Full size table

Since the RNA-seq library was prepared with rRNA depletion, no data for RNA18S can be drawn from the RNA-seq data. It was noticed that the variation of commonly used HKGs (by CV and/or dispersion) was generally larger than those selected from the transcriptomic data (Table 1). This may be the first sign that those common HKGs are not good choices when studying ECs in different growth states.

To ensure the validity of RT-qPCR results for subsequent stability tests, the specificity of the primers was verified by melt curve analysis and agarose gel electrophoresis to establish the products are single bands at corresponding base-pair positions (Supplementary Fig. 1 and Supplementary Table 1). The efficiency of the primers was analysed with the serial dilution method (including non-template control), where the acceptable efficiency range should be within 80% to 120%, and optimally at 90% to 110% (Supplemental Table 1).

Stability of candidate HKGs

Although endothelial cells are known to exhibit contact inhibition, we published previously that EA.hy926 cells will only enter a quiescent state when grown on MG-coated plates22. In contrast, the rate of DNA synthesis of cells grown to confluence on non-coated plates remains high22, which resembles the dysfunctional state in vivo. To study the stability of candidate HKGs (listed in Table 1) across all endothelial growth states, RT-qPCR employed RNA isolated from EA.hy926 cells grown to subconfluence (day 4/D4) and confluence (day 10/D10) on both MG-coated (+MG) or non-coated (-MG) 6-well plates. The Cq values from 4 biological replications run as technical duplicates were supplied to 5 commonly used stability test algorithms: deltaCt, BestKeeper, normFinder, geNorm, and RefFinder.

The expression profiles of the candidate HKGs from RT-qPCR (Fig. 1) revealed that YWHAZ appeared to have one of the narrowest overall inter-quartile ranges (IQR, Fig. 1a) and one of the smallest variations in expression among growth conditions (Fig. 1b and Suppl Fig. 2). On the other hand, RNA18S, together with genes such as APMAP, CDIPT and PRKCSH, had the largest IQRs (Fig. 1a). Additionally, there were clear differences in mean Cq between D4 and D10 samples for those genes (Fig. 1b and Suppl Fig. 2), which indicates poor suitability as HKGs for normalization against different growth conditions.

Fig. 1
figure 1

Candidate reference gene expression profiles from RT-qPCR. a) Overall, b) By growth state; n = 4/growth state (day 4 or 10 with or without Matrigel (MG)).

Full size image

Figure 2 displays the results from 4 individual algorithms: deltaCt (Fig. 2a), BestKeeper (Fig. 2b), normFinder (Fig. 2c, d), and geNorm (Fig. 2e, f), while the results of the integrative algorithm, RefFinder, are presented in Table 2. Genes with lower scores are more stable in expression across all conditions tested, and are thus better choices for reference HKGs. Both deltaCt and normFinder methods ranked CAPZB as the most stable gene, while FBXO7 and SMU1 were among the top 5. FBXO7 and SMU1 were ranked first by geNorm. geNorm also provided pairwise variation (V) values, which can be used to instruct how many HKGs are needed for good normalization18. Specifically, a threshold of Vn/n+1 < 0.15 indicates that adding an (n + 1)th gene would not significantly improve normalization18,23. In our case, the first pairwise variation value V2/3 equaled 0.0033 (Suppl Fig. 3), indicating that two HKGs are sufficient. BestKeep produced very different results compared to the other 3 algorithms and ranked YWHAZ as the most stable gene in this data set. Although there were discrepancies in the exact rankings among different algorithms and different R packages used, RNA18S and/or HPRT1 were the least stable genes for all cases.

Fig. 2
figure 2

Stability results for the candidate reference genes from the 4 methods: a) deltaCt, b) BestKeeper from ctrlGene package, c) normFinder from source code, d) normFinder from NormqPCR package, e) geNorm from ctrlGene package, and f) geNorm from NormqPCR package.

Full size image
Table 2 RefFinder results for the candidate reference genes. RefFinder used Raw Cq values to generate rankings based on the other 4 algorithms, and then calculated their geometric means to give the final overall ranking. The top 3 ranked genes were CAPZB, SMU1, and FBXO7, which were ranked first by at least one other algorithm. However, due to lower than optimal primer efficiency, SMU1 was not included for further validation. SD: standard deviation, R: Ranking.
Full size table

RefFinder used raw Cq values to calculate rankings for the other 4 algorithms (Table 2), which matched well with those calculated individually by us for deltaCt (Fig. 2a), and by BestKeeper (Fig. 2b) as well as geNorm from the ctlGene package (Fig. 2e). Then RefFinder computed the overall ranking based on the geometric mean of rankings from the other 4 algorithms (Table 2). The top 5 most stable genes from geNorm (NormqPCR package, Fig. 2f) were the same as the other 2 sources (Fig. 2e; Table 2), although detailed rankings differed to a degree. However, the results for normFinder showed slight discrepancies among all 3 sources (Fig. 2c, d, and Table 2). What was relatively consistent, though, is the first few and last few genes in ranking for normFinder from the 3 sources. Therefore, the evidence behind the overall ranking by RefFinder should be strong: CAPZB, SMU1, and FBXO7 were the top 3 most stable HKGs from the candidate list, while RNA18S and ACTB were the least stable ones among the commonly used HKGs.

Next, validation was done by normalizing RT-qPCR results against the 3 most stable HKGs (sHKGs) versus the 3 least stable commonly used HKGs (cHKGs). Although CAPZB, SMU1, and FBXO7 should be the top 3 sHKGs, the primer efficiency of SMU1 was suboptimal (Supp Table 1), thus, CAPZB and FBXO7 were selected together with YWHAZ, which was ranked highest by BestKeeper (Fig. 2b) and had the least variable mean Cq (Fig. 1a & b). The 3 cHKGs selected were RNA18S, ACTB, and HPRT1, which were ranked the last by deltaCt (Fig. 2a), normFinder (Fig. 2c), and geNorm (Fig. 2e).

The effect of different HKG combinations on the relative expression of marker genes

Certain genes are known as endothelial markers and may differentially express during the growth of ECs. Here we chose CD36 (critical for endothelial fatty acid uptake24), CDH5 (VE-cadherin or CD144), PECAM1 (platelet endothelial cell adhesion molecule-1 or CD31), and VWF (von Willebrand factor). In addition, given the number of days in culture, HIF1A (hypoxia-inducible factor 1-α) and PCNA (proliferating cell nuclear antigen, a marker of proliferation) were used for validation. Day 10 cells should suffer more from hypoxia than day 4 cells, and have less room for proliferation. Hence, HIF1A expression should be higher in day 10 cells while PCNA expression should be higher in day 4 cells. Their expression levels relative to different HKG combinations are shown in Fig. 3.

Fig. 3
figure 3

RT-qPCR results for endothelial marker genes normalized to various combinations of HKGs. Normalized expression results for genes a) CD36, b) CDH5, c) HIF1A, d) PCNA, e) PECAM1, and f) VWF. Each panel is organized into 2 rows with the top row left to right representing cHKGs individually or in the indicated combinations: 18 S, 18 S + ACTB, and 18 S + ACTB + HPRT1, and the bottom row left to right representing sHKGs individually or in the indicated combinations: CAPZB, CAPZB + FBXO7, and CAPZB + FBXO7 + YWHAZ. n = 6–8 per gene per growth state (a growth state is day 4 or day 10 with or without Matrigel (MG)). A bar with asterisks (*: p < 0.05; **: p < 0.01; ***: p < 0.001; ****: p < 0.0001) indicates significant differences between specified groups; a bar with a p value indicates a trend (p = 0.05 to 0.1) between specified groups.

Full size image

As the most stable candidate HKG in Table 2, CAPZB outperformed RNA18S, the least stable HKG in the list. When normalized to CAPZB, we were better able to detect statistical significance among all 6 validation genes (Fig. 3). In some cases, we also observed large differences in expression when normalizing to CAPZB or RNA18S. For example, day 10 with MG samples (D10 + MG) showed 6.97-fold higher expression of CD36 than day 4 with MG samples (D4 + MG) when normalizing against RNA18S, but CD36 expression was only 3.92-fold higher when using CAPZB as the HKG (Fig. 3a).

According to MIQE guidelines, the use of 2, and even 3 or more reference genes is preferred10. Therefore, we compared the normalization results against different combinations of HKGs: 18 S + ACTB, 18 S + ACTB + HPRT1, CAPZB + FBXO7, and CAPZB + FBXO7 + YWHAZ. As shown in Fig. 3, even though the cHKGs ranked among the lowest in Table 2, the use of 3 cHKGs (18 S + ACTB + HPRT1) improved the statistical power compared to that with 2 cHKGs (18 S + ACTB), and almost matched the results with 2 sHKGs (CAPZB + FBXO7). With certain genes (CDH5, HIF1A, and VWF; Fig. 3b, c, f), normalizing against 2 sHKGs (CAPZB + FBXO7) seemed to provide slightly better statistical power than that against 3 cHKGs (18 S + ACTB + HPRT1). This could demonstrate the superior suitability of the top 2 ranked HKGs as RT-qPCR reference genes for EA.hy926 cells assessed in different growth states. However, when including YWHAZ as the third reference gene for the sHKG group (CAPZB + FBXO7 + YWHAZ), the outcome was no better than that obtained with CAPZB + FBXO7, especially for CD36 (Fig. 3a) and PCNA (Fig. 3d). This may be due to greater variability in the normalized expression level as evident by the box plots and the spread of the data points. The poorer performance of CAPZB + FBXO7 + YWHAZ compared to CAPZB + FBXO7 may be related to the lower ranking of YWHAZ in Table 2. Overall, the selected HKGs (CAPZB and FBXO7), i.e. those ranked higher in Table 2, had better performance as reference genes across EC growth states compared to commonly used HKGs (RNA18S, ACTB, and HPRT1), which ranked lower in the list.

Next, we validated the stability of these 3 sHKGs and 3 cHKGs in response to DHA treatment. Although most of the literature has used 10 to 80 µM DHA on ECs4,5,6, plasma DHA content can range from 7.2 µM to 237.5 µM in healthy young adults25, and can surge up to 588 µM after supplementation in certain subpopulations26. In addition, we have demonstrated previously that the effects of DHA at 125 µM on EA.hy926 cells were different from those observed at 20 µM21. Therefore, DHA at both 20 and 125 µM concentrations was used for this validation.

As seen in Fig. 4a, except for RNA18S, the other 5 genes exhibited a big shift in Cq value, i.e. lower expression in day 10 cells in the absence of MG (D10-MG) after 125 µM DHA treatment. However, under other conditions, variations in the Cq values were generally smaller for CAPZB and FBXO7 than other HKGs. Therefore, besides the 6 HKG combinations in Fig. 3, combinations with RNA18S and other sHKGs were added to examine their influence on the variation in D10-MG cells treated with 125 µM DHA.

Fig. 4
figure 4

RT-qPCR results for DHA-responsive genes normalized to various combinations of HKGs. a) Mean Cq values of the 6 HKGs individually by docosahexaenoic acid (DHA) treatment and growth state. Normalized expression results for genes: b) BTG1, c) NOS3 (arranged by growth states), d) NOS3 (arranged by DHA treatment), e) PCNA (arranged by growth states), and f) PCNA (arranged by DHA treatment); each panel is organized into 2 rows with the top row left to right representing HKGs individually or in the indicated combinations: 18 S, 18 S + ACTB, 18 S + ACTB + HPRT1, and CAPZB + 18 S, and the bottom row left to right representing HKGs individually or in the indicated combinations: CAPZB, CAPZB + FBXO7, and CAPZB + FBXO7 + YWHAZ, and CAPZB + FBXO7 + 18 S. n = 4 per gene per DHA treatment (0, 20, 125 µM DHA) per growth state (day 4 or day 10 with or without Matrigel (MG)). A bar with asterisks (*: p < 0.05; **: p < 0.01; ***: p < 0.001; ****: p < 0.0001) indicates significant differences between specified groups; a bar with a p value indicates a trend (p = 0.05 to 0.1) between specified groups. D4: day 4; D10: day 10; MG: Matrigel.

Full size image

BTG1 belongs to an anti-proliferation gene family that regulates cell growth and differentiation27. It is one of the few genes that was upregulated by DHA in D4 + MG cells but downregulated in D10 + MG cells based on our RNA-seq data. This trend can be observed in Fig. 4b, in all HKG combinations. The 2 sHKGs, CAPZB + FBXO7, still showed stronger statistical power compared to other combinations, evident by the statistical results in the CAPZB + FBXO7 panel compared to other panels in Fig. 4b. However, it only failed to detect the difference due to 125 µM DHA treatment in the D10-MG condition compared to other treatments under these conditions. The downregulation of BTG1 by 125 µM DHA compared to control and/or 20 µM DHA samples in D10-MG was detected when including RNA18S as one of the HKGs with sHKGs (CAPZB + 18 S or CAPZB + FBXO7 + 18 S).

NOS3 encodes eNOS, which plays a crucial role in endothelial function7. Our RNA-seq data suggested that 125 µM DHA may downregulate NOS3 transcripts compared to control. Similar to Fig. 4b, when normalizing to sHKGs, NOS3 expression after 125 µM DHA treatment of D10-MG cells appeared to be higher than in other samples under the same growth state, as well as other growth states treated with 125 µM DHA (Fig. 4c). However, NOS3 expression was lower than in other samples when normalized to 18 S or CAPZB + 18 S. Only when normalizing to 18 S did 125 µM DHA treatment significantly reduce NOS3 expression compared to VC and the 20 µM DHA samples (Fig. 4d), matching our previous report that eNOS total protein level was lowered by 125 µM DHA21. In that paper, we also reported that eNOS protein levels were higher in D10 + MG cells compared to D4 + MG cells21. This finding supports the statistical significance found between D10 + MG cells and D10-MG or D4-MG cells when RNA18S was included as an HKG (i.e. 18 S + ACTB, 18 S + ACTB + HPRT1 and CAPZB + 18 S panels in Fig. 4c). Although the CAPZB + FBXO7 + 18 S panel had p = 0.058 for the main effect of growth state, its post-hoc test results were significant for D10 + MG vs. D10-MG (p = 0.0066) and D10 + MG vs. D4-MG (p = 0.0154). Except for the 125 µM DHA treated D10-MG cells, sHKGs (CAPZB + FBXO7) were still better in detecting statistically significant differences in other conditions. In summary, RNA18S should be considered a supplemental HKG when dealing with extreme DHA treatment situations.

This ability to better detect statistical significance for 125 µM DHA-treated samples, especially in D10-MG cells, by including RNA18S as an HKG may not be necessary when the difference due to 125 µM DHA treatment is small. This was the case for PCNA, which is involved in DNA replication and repair28. PCNA transcripts were downregulated by 125 µM DHA and declined further under Day10 + MG conditions compared to Day4 + MG conditions based on the RNA-seq data. As shown in Fig. 4f, 125 µM DHA significantly lowered PCNA expression compared to the other 2 samples when normalized to 18 S + ACTB, CAPZB + FBXO7, as well as CAPZB + 18 S and CAPZB + FBXO7 + 18 S. However, only normalization to CAPZB + FBXO7 showed similar results to those in Fig. 3, where PCNA expression in day 10 cells was lower than in day 4 cells (Fig. 4e). This provided further evidence that the CAPZB + FBXO7 combination is a better choice as an HKG for ECs across different growth states and DHA treatment conditions, except for high DHA concentrations. When dealing with samples exposed to high DHA concentrations, it may be valuable to include RNA18S as an HKG to complement the combination of CAPZB and FBXO7.

Discussion

In this study, a systematic approach was used to identify a set of RT-qPCR reference genes, or HKGs, for EA.hy926 ECs across different growth states and in response to DHA treatment. This study is also the first to validate RT-qPCR reference genes for EA.hy926 cells under these specific conditions. The candidate HKGs came from 3 sources, our transcriptomic dataset, online databases that incorporated data from various human tissues and cell lines, and commonly used HKGs from the literature. Based on the transcriptomic data, the variation in expression of almost all candidate genes from different sources was known, which was a useful first approach to screen candidate HKGs. Also, given the large and increasing amount of omics data available online, the approach we describe here, not the actual results, can be robustly applied, at relative low cost, to many other cell types or tissues under specific conditions where HKGs for RT-qPCR have yet to be identified and validated according to the MIQE guidelines.

At least 14 of the 18 candidate HKGs from our RNA-seq data were neither identified nor validated as reference genes in any previous context. DDB1 and FBXO7 had been identified as candidate reference genes for human pluripotent stem cell-derived cardiomyocytes, with DDB1 selected into the final reference gene panel29, whereas NCOA4 was listed in the top 10 candidate reference genes for several cancer types30. The candidate genes identified via RNA-seq in this study also belong to different functional classes; for example, RNA processing (PCBP2, STAU1, and SMU1), ubiquitination (DDB1 and FBXO7), metabolism (PRKCSH, APMAP, and CDIPT), cytoskeleton (CAPZB and FNBP1), and signal transduction (GNB1 and CSNK2B). Therefore, the possibility of co-regulation in response to the same experimental conditions should be low.

From our list (Table 2), the top-ranked HKGs, CAPZB, FBXO7, and SMU1, were rarely mentioned in other similar studies11,12,13,14,15,23,31 let alone validated. SMU1 is a spliceosome factor whose expression data are largely unknown. Given its essential role in constitutive splicing function, SMU1’s expression might be stable. CAPZB is an actin-capping protein regulating cytoskeleton dynamics, especially in muscle32. Actin-cytoskeleton dynamics can be sensitive to cellular stress and other situations like angiogenesis33, thus possibly making expression of these genes vary greatly. For instance, the expression of TAGLN, an actin-crosslinking protein, was found to increase during HUVEC elongation in an angiogenesis model34. Therefore, supposedly, CAPZB’s expression should be variable across the 4 EC growth conditions. Similarly, FBXO7, an E3 ligase adaptor, has exhibited differential expression in certain cancers and is involved in cell cycle regulation and inflammation35. FBXO7’s disease-associated variability and regulatory roles preclude its stable expression in theory. However, context is paramount, especially in assessing reference gene stability. That is why MIQE guidelines emphasize validating reference genes for each specific experimental condition10. In this study, we used an unbiased approach to screen transcriptomic data as a means of identifying potential HKGs for our specific EC culture conditions, since RNA-seq mining often identifies non-canonical, novel candidates29,31. The algorithm output consensus across deltaCt, geNorm, NormFinder, and RefFinder provided confirmation that the validation was robust. Also, the genes revealed in this study would be stably expressed if the growth conditions and the treatment, DHA in this case, did not affect the underlying processes that modulate transcription of the genes. For instance, although cytoskeleton regulation is a very dynamic process that is sensitive to cellular stress, it has been reported that regulation of actin-capping protein (CapZ) activity in response to stimuli largely involves modulation of binding partners and/or post-translational modifications36,37,38 rather than gene expression. Similar to CapZ, the expression of FBXO7 is unregulated upon stress39. Furthermore, FBXO7’s activity is mainly regulated via protein-protein interactions40 and sub-cellular localization39. Most reports on the transcriptional changes of FBXO7 are related to diseases like cancer, which is quite distinct from the conditions of our study. Thus, context-specific regulation may help to explain the paradox here. However, if different conditions were being examined (e.g. treatment with cytokines instead of fatty acids), the suitability of CAPZB and FBXO7 as stable reference genes will need to be re-evaluated and validated. Of note, exposure to 125 µM DHA may represent a high stress condition, and this may explain the large up-shift in the expression of both CAPZB and FBXO7 seen in the D10-MG cells (Fig. 4a), a response that is not triggered by the other conditions examined.

Descriptive statistics of gene expression, such as either CV or dispersion, as in Table 1, or IQR, as in Fig. 1, provides a preliminary approach for evaluating HKG stability. For instance, FBXO7 was among the top 5 genes with respect to dispersion (Table 1), while RNA18S had one of the largest IQR (Fig. 1a). However, the accuracy of this simple method is restricted to a limited degree, and this simplicity may often obscure the underlying biological and technical variability. Therefore, more robust methods accounting for various parameters in gene stability assessment are needed. The 5 tools used in this paper, namely deltaCt, geNorm, NormFinder, BestKeeper, and RefFinder, are widely used in various HKG identification papers31,41. They all employ different methodologies when assessing gene stability, hence resulting in discrepancy of varying degrees in the final ranking. Another potential confounding factor determining the accuracy of such assessments is the primer efficiency of each gene. BestKeeper and RefFinder ignore primer efficiency in their algorithms, and mis-ranked genes tend to be those whose efficiency deviates considerably from 100%42. Therefore, although SMU1 was among the top 3 stable genes identified by RefFinder, its suboptimal primer efficiency compromises the reliability of its position in the ranking. Hence, SMU1 was not included for further validation in this study. Nonetheless, SMU1 remains a promising candidate for EA.hy926 cells, especially those in different growth states, but further investigation with improved primers would be needed to confirm its suitability.

Interestingly, since the primer efficiencies of our candidate genes are, for the most part, within the optimal range of 90 ~ 110% (Supplemental Table 1), it is unlikely that primer efficiency alone can explain the different BestKeeper results that were obtained in relation to the other 3 algorithms (Fig. 2; Table 2). However, compared to the variance modeling approach of geNorm and NormFinder that favours genes with a similar expression pattern across samples18,19, BestKeeper’s algorithm ranks genes using standard deviation (SD) of raw Cq as the basis for computing pairwise correlation metrics for each gene17. This approach favours genes with high expression levels and low technical noise43,44. For this reason, it may be advisable to exclude genes with SD >1 when using BestKeeper15,17. These differences in the metrics and mathematical criteria used by these algorithms likely explain the discrepancies in our results. The resulting differences in HKG ranking between BestKeeper and other algorithms has been commonly reported in other papers43,44,45. Therefore, some studies may base their HKG selection on the other algorithms only, not BestKeeper43. This is true in this study as well. CAPZB and FBXO7 were among the top HKGs ranked by all 4 other algorithms, while YWHAZ was ranked first only by BestKeeper (Table 2). According to our validation results, the addition of YWHAZ to CAPZB and FBXO7 did not improve performance, and at times resulted in even worse statistical outcomes (Figs. 3 and 4).

There is a limited number of published reports describing the validation of RT-qPCR reference genes for ECs. Most studies have only investigated commonly used HKGs like GAPDH, ACTB, B2M, HPRT1, RNA18S, TPB, and YWHAZ13,14,15, which we also included in our study for completeness. However, the rankings of those cHKGs differed among the various conditions and cell types. For statin-treated HUVECs, HPRT1 and YWHAZ were the most stably expressed genes among the 8 HKGs analyzed15, similar to the HKG rankings we obtained with BestKeeper, which showed that YWHAZ and HPRT1 were the most stable (Fig. 2b). For homocysteine-treated HUVECs, on the other hand, GAPDH was the most reliable HKG among the 10 cHKGs tested13. For EA.hy926 cells and primary human coronary artery ECs grown under hypoxic conditions, the top 2 most stable HKGs tested were B2M/TBP and PPIA/RPLP1, respectively14. These results for EA.hy926 cells differ slightly from the most stable cHKGs in our study where TBP was the most stable cHKG overall, whereas B2M was ranked the third last among the 8 cHKGs (Table 2). Under hypoxic conditions, the least stable HKGs were RNA18S and ACTB for EA.hy926 cells14, which is highly analogous to our cHKG rankings (Table 2). Those results indicate there is a serious need to validate RT-qPCR reference genes for each cell type as well as the specific conditions to which they are exposed. This would help to minimize the overgeneralization that often occurs when comparing one condition to another.

In addition to the many studies that have examined the suitability of cHKGs, only 2 others have utilized RNA-seq data to identify potential HKGs for endothelial-like cells, specifically human endothelial colony forming cells11 and induced pluripotent stem cells-derived endothelial cells31, in relation to cHKGs. In all 3 cases, which includes our data presented in Table 2, the most stably expressed reference genes were discovered in the RNA-seq data, and most cHKGs were ranked much lower. This indicates that it should not be assumed that cHKGs will make useful reference genes for RT-qPCR analysis in most cases. Rather, the use of RNA-seq data to unbiasedly identify candidate reference genes, as shown herein, provides a much better approach.

Conclusions

The primary outcome of our investigation was the identification of a panel of candidate HKGs for human EA.hy926 ECs in different growth states. Among them, CAPZB and FBXO7 were the top 2 stable genes based on the analysis made, which remained valid when the cells were treated with DHA. This careful evaluation and validation of HKGs prior to the actual RT-qPCR with more than 1 HKG selected adheres closely to the MIQE guidelines. We also found that, in ECs, most of the commonly used HKGs exhibited poor stability as RT-qPCR reference genes, especially when used alone. However, further validation would be needed if these genes are to be employed as HKGs for other conditions or cell types.

Methods

Cell culture and treatment

Human EA.hy926 endothelial cells (ATCC, CRL 2922) were cultured in Dulbecco’s modified Eagles’ medium (DMEM, Gibco 12100061, Waltham, MA, USA) supplemented with 20 mM HEPES (MilliporeSigma 391333, Darmstadt, Germany), 100 units/mL penicillin/streptomycin (Gibco 15140122), and 10% fetal bovine serum (FBS, Gibco 16000069). The cells were sub-cultured at ~ 80% confluency, and only cells within passage 20 were used.

For experiments, EA.hy926 cells were seeded onto 6-well plates coated with or without growth factor reduced MG (Corning® 356231, Corning, NY, USA) as previously described22. Taking the seeding day as day 0, the cells were harvested at day 4 (sub-confluent, growing state), and day 10 (quiescent state)21,22. Prior to harvesting, the cells were treated with or without 20 or 125 µM DHA (bound to 5% fatty acid-free bovine serum albumin (BSA)-PBS) for 8 h. Ethanol in 5% BSA-PBS was added as VC.

RNA isolation and sequencing

Total RNA was extracted from harvested cells using the Monarch® Total RNA Miniprep kit (NEB T2010S, Ipswich, MA, USA) according to the manufacturer’s protocol with DNase I treatment. RNA concentrations and quality were measured with the NanoDrop™ One instrument (Thermo Scientific, Waltham, MA, USA). Aliquots of day 4, and day 10 samples treated with or without DHA (20 µM, n = 2) were sent to Genome Quebéc for RNA-seq with an Illumina NovaSeq 6000 system (100 bp paired-end reads, 50 million reads per sample). Additional RNA samples (n = 4–6) were generated for RT-qPCR to validate the findings.

RNA-seq data analysis and candidate reference gene (HKG) selection

The RNA-seq data (deposited to the NCBI Sequence Read Archive (SRA) with accession number PRJNA1304531) were processed using the 4SeqGUI platform46. Briefly, the raw reads were trimmed by Skewer (v. 0.2.2)47, and then aligned against human genome GRCh38 (Ensembl release 101) using RSEM (v. 1.3.3) + STAR. The outputs from RSEM were imported into R studio (v. 4.2.2) using tximport (v. 1.26.1)48 and further processed with DESeq2 (v. 1.38.3)49 for gene expression count normalization and differential gene expression analysis.

HKG candidates were selected from the RNA-seq data according to the following steps: (1) filter out the non-differentially expressed genes defined as those with Benjamini-Hochberg (BH) adjusted p value (padj) greater than 0.1 and absolute fold change (|FC|) less than 1.5 (i.e. |log2FC| or |LFC| < 0.5); (2) compute the CV and dispersion of the normalized counts of these genes, and select the top 20 with smallest CV and/or dispersion as candidate genes. Common reference genes for human aorta, vein, and/or HUVEC cells were extracted from databases RefEx (https://refex.dbcls.jp/) and HRT Atlas (https://housekeeping.unicamp.br). Commonly used HKGs were selected from the literature, namely GAPDH, RNA18S, ACTB, B2M, GUSB, HPRT1, YWHAZ, and TBP.

RT-qPCR

RNA was converted to cDNA using the iScript cDNA Synthesis kit (Bio-Rad 1708891, Hercules, CA, USA). Then, RT-qPCR was performed using the iTaq Universal SYBR Green Supermix (Bio-Rad 1725121) on the CFX Connect platform (Bio-Rad) with technical duplicates. A list of primer sequences is shown in Supplemental Table 1. Primer specificity and efficiency were checked by the melt curve and serial dilution methods, respectively.

Statistical analysis

Five commonly used algorithms were employed to assess the expression stability of the candidate HKGs: deltaCt16, BestKeeper17, geNorm18, NormFinder19, and RefFinder20 using R or their web source. The deltaCt method compares the mean SD of a candidate gene’s deltaCt (∆Ct) to all other candidates (scripts for these calculations are available at https://github.com/SunnyXS-phd/EAcell-HKG). R package, ctrlGene (v. 1.0.1), was employed to conduct assessment using Bestkeeper and geNorm methods; NormqPCR package (v. 1.6.0) was used for the geNorm and NormFinder methods. The source R code for NormFinder was downloaded from https://www.moma.dk/software/normfinder and carried out in R. RefFinder, a web-based tool, was accessed via https://www.ciidirsinaloa.com.mx/RefFinder-master/ to give a comprehensive stability ranking of candidate genes based on the four algorithms mentioned.

The expression of genes being validated (expression is different among growth states and/or treatment based on RNA-seq and/or the literature) was compared against various reference gene combinations. One-way ANOVA was carried out for analysis of the effects of growth state, while two-way ANOVA (or Generalized Least Squares or the ART model with robust ANOVA, depending on data normality and homogeneity) was employed for the interactive effects of growth state and DHA treatment, followed by post-hoc testing. Before applying appropriate statistical tests, extreme outliers (defined as those more than 3× IQR from the group mean) were removed before analysis. Log transformation was used for data sets that failed the Shapiro-Wilk test. Detailed statistical analyses and R scripts are available at https://github.com/SunnyXS-phd/EAcell-HKG. Statistical significance was defined as p < 0.05, except for the interaction effect in a two-way ANOVA (growth state × DHA treatment) where p < 0.1 was deemed significant.