Background

Cancer remains one of the leading causes of human mortality worldwide. Gene expression profiling, which involves simultaneous measurement of the activity of thousands of genes, offers valuable insights into the molecular characteristics of tumors1,2. This approach has yielded significant discoveries, including biomarkers that aid in diagnosis, prognosis, and prediction of therapeutic response across various cancer types3. For instance, the Oncotype DX Breast Recurrence Score (Genomic Health, Redwood City, CA, USA), approved by the Food and Drug Administration, is widely utilized for assessing breast cancer recurrence risk. Similarly, Prolaris (Myriad Genetics, Salt Lake City, UT, USA) is employed in prostate cancer to predict therapy response and likelihood of cancer recurrence4,5. These qualitative transcriptional signatures predominantly rely on absolute measured values of gene expression.

However, the quantitative precision of gene expression measurements is often compromised by various pre-analytical variables, rendering them error-prone and uncertain6,7. These variables include sampling methods8,9, tumor sample heterogeneity5,10, fixed time delays in sample processing11,12, preservation conditions13,14, degradation levels15, library preparation kits16,17, amplification kits18,19, RNA quantity20, measuring platforms21,22, and laboratory sites23. Such variables introduce variability in risk assessment of qualitative signatures across samples from the same patient24. Thereby these limitations restrict the application of assays like Oncotype DX and Prolaris, which necessitate transporting samples to the central laboratory for stringent quality control25,26.

Previous studies have revealed the robustness of within-sample relative expression orderings (REOs) of gene pairs against substantial measurement variations induced by pre-analytical variables, including formalin-fixed paraffin-embedded (FFPE) samples27, variations in tumor epithelial cell proportions28, and low-input RNA samples29. However, many findings were derived from simulated data due to limited access to real datasets28. Moreover, these studies typically focused on single variables and neglected the combined effects of multiple variables. Consequently, there exists a dearth of comprehensive assessments regarding the impact of pre-analytical variables on gene expression analysis in biospecimen collection and preprocessing.

In this study, we assess the influence of ten pre-analytical variables spanning biospecimen collection to data generation, on gene expression measurements and REOs of gene pairs through single-variable and multi-variable analyses. These variables include sampling methods, tumor sample heterogeneity, fixed time delays, preservation conditions, degradation levels, extraction and amplification kits, RNA quantity, measuring platforms, and laboratory sites.

Methods

Datasets and preprocessing

We traversed protocols and workflows for gene expression measurements based on oligonucleotide technology or Illumina-based sequencing technology and screened expression data designed for exploring pre-analytical variables spanning biospecimen collection to data generation from public databases, including Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/), The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/), and Sequence Read Archive (SRA, https://www.ncbi.nlm.nih.Gov/sra/). To minimize the influence of unobservable variables and ensure comparability, only paired samples were selected for pre-analytical variables analysis. Finally, we collected ten pre-analytical variables, as outlined in Table 1. Detailed information on data processing is described in Supplementary Methods.

Table 1 Datasets used in this study.
Full size table

Differential expression analysis

At the individual level, we calculated the fold change (FC) of gene expression values between each paired case and control sample for each gene. The genes with a FC greater than 2 were defined as differentially expressed genes (DEGs). As described in the following formula (1):

$$\text{fold change}=\frac{\text{exper}\left(\text{B}\right)}{\text{exper}\left(\text{A}\right)}$$
(1)

where exper(A) indicates the gene expression value of the control sample and exper(B) indicates the gene expression value of the paired case sample.

At the population level, DEGs were identified between case and control groups using limma, with a false discovery rate (FDR) less than 0.05 and a FC greater than 2.

Consistency scores of REOs

The comparison of two genes in a gene pair (Gi, Gj) is regarded as an event with only two possible outcomes: the expression level of Gi is either higher or lower than that of Gj. Consequently, the REO of the gene pairs is denoted as Gi > Gj or Gi < Gj. We assumed that the REOs of gene pairs is consistent or contradictory between paired samples.

To quantify the consistency of REOs between paired samples, we utilized the following formula (2):

$$\text{Consistency score }= \frac{\text{N}}{(\text{N}+\text{M})}$$
(2)

where N and M represent the number of gene pairs with consistent and contradictory REO pattern between a paired sample, respectively.

Subsequently, all genes in a control sample were ranked based on their expression levels in ascending order. The rank difference for each gene pair was calculated using the following Eq. (3):

$$R_{ij} = \left| {R_{i} } \right. - \left. {R_{j} } \right|$$
(3)

where Ri and Rj denote the ranks of genes Gi and Gj in control samples, respectively, and Rij represents the absolute rank difference between the two genes. Gene pairs with the smallest Rij were considered to have the closest expression levels.

For all gene pairs within each control sample, 10% and 20% of gene pairs with the smallest expression differences were excluded. The remaining gene pairs were used as the standard to calculate the consistency score of these gene pairs in the paired case sample using formula (2).

At the population level, significant reversal gene pairs, defined as gene pairs with significant different distribution of REO patterns between two groups, were identified using Fisher’s exact test with a FDR less than 0.05.

Results

Effect of single pre-analytical variable on gene expression measurements and relative expression orderings

To evaluate the impact of pre-analytical variables on gene expression measurements and REOs of gene pairs, separate gene expression comparative analyses of case samples were conducted against paired control samples, as presented in Table 1. The average number of genes exhibiting twofold changes in expression levels and the average consistency scores of REOs (see Methods) were calculated for paired case and control samples extracted from the same patient.

  • 1. Sampling methods

Surgical, biopsy, and cytological samples are commonly sampling methods in clinical settings for diagnostic purposes. Accurate molecular testing of tumors necessitates sufficient tissue acquisition30. Thus, biopsy and cytological samples were classified as case samples, while surgical samples were classified as control samples. Twenty paired biopsy and surgical samples from patients of esophageal cancer were extracted from GSE32701. On average, 3286 genes exhibited twofold changes in expression values in biopsy samples compared to paired surgical samples. Conversely, an average consistency score of REOs exceeding 86% was observed in these samples. Moreover, after excluding 10% of gene pairs with the closest gene expression levels, the average consistency score for the remaining gene pairs increased to 89.90% (Fig. 1A). Similar results were observed in 116 paired cytological and surgical samples of breast cancer collected from GSE129559 (Figure S1A).

  • 2. Tumor sample heterogeneity

Fig. 1
figure 1

Effect of sampling methods and tumor sample heterogeneity on gene expression measurements and REOs. Scatter plots depict the consistency scores of REOs (pink and green dots, left vertical axis) and the number of DEGs (purple dots, right vertical axis) between biopsy and surgery samples from GSE32701 (A), among samples with varying tumor proportions from GSE116782 (B).

Full size image

In molecular oncology testing, the proportion of malignant cells in tumor tissue is pivotal for obtaining reliable results31. The over 70% of the tumor cells were deemed suitable for subsequent sequencing32. Therefore, samples containing high proportions of tumor epithelial cells (93%–74%) in GSE116782 were designated as the control group, whereas those with low proportions (73%–14%, 35%–14%) were classified as case groups. On average, 5707 genes with twofold changes in expression values, but an average consistency score of REOs with 89.24% was observed in paired case samples. Furthermore, after excluding 10% of gene pairs with the closest gene expression levels, the average consistency score for the remaining gene pairs increased to 92.46% (Fig. 1B). Additionally, similar results can be found in three paired colon cancer samples collected from GSE84984 (Figure S1B).

  • 3. Fixed time delays, Preservation conditions and Degradation levels

RNA is highly susceptible to degradation, particularly at room temperature, posing a challenge in the collection and preservation of biospecimens in routine clinical practice15,33. To evaluate the impact of RNA degradation on gene expression measurements, separate gene expression comparative analyses were conducted for fixed time delays of samples (the duration between sample aspiration and preservation at room temperature), preservation conditions (fresh-frozen (FF) and FFPE), and different degradation levels (intactness, degradation, and high degradation), as summarized in Table 1.

For fixation time delays of samples, six paired bone marrow samples with varying fixed time delays within the same patients were collected from GSE1347. Compared to their paired 0-h sample, on average, 2113 and 2970 genes exhibited twofold changes in expression measurements, but the average consistency scores of REOs were 88.94% and 85.63% in 24-h and 48-h samples, respectively. After excluding 10% of gene pairs with the closest expression levels, the average consistency scores for the remaining gene pairs increased to 92.27% and 88.84%, respectively (Fig. 2A).

Fig. 2
figure 2

Effect of fixed time delays, preservation conditions and degradation levels on gene expression measurements and REOs. Scatter plots depict the consistency scores of REOs (pink, green and blue dots, left vertical axis) and the number of DEGs (purple dots, right vertical axis) between samples with different fixed time delays from GSE1347 (A), among samples with different preservation conditions from GSE17558 (B) and TCGA_BRCA (C). Boxplots depict the distribution of (D) number of DEGs and (E) consistency scores of REOs between samples with different degradation levels from SRP097611.

Full size image

For preservation conditions, eight and six paired FFPE and FF samples of lung cancer and breast cancer were extracted from GSE17558 and TCGA_BRCA. On average, 5009 and 10,388 genes exhibited twofold changes in expression values, but the average consistency scores of REOs were 86.42% and 84.64% between FFPE and FF samples, respectively (Figs. 2B-C). Furthermore, after excluding 10% and 20% of gene pairs with the closest expression levels, the average consistency scores for the remaining gene pairs increased from 89.80% and 87.87% to 92.69% and 90.65%, respectively (Figs. 2B-C). Similar findings were observed in additional samples of breast cancer (GSE51124, GSE113976, GSE93338) and diffuse large B-cell lymphoma (GSE19246) (Figures S2A-D).

For different degradation levels, on average, 5180 and 3743 genes exhibited twofold changes in expression values, but the average consistency scores of REOs were 89.50% and 87.25% between 72 paired intact and degraded samples (samples-size peak at ~ 800 bp) and between 48 paired intact and highly degraded samples( sample-size peak at ~ 200 bp) collected from SRP097611 (Fig. 2D-E). The consistency scores of REOs decreased as delay time increased (Fig. 2A) and as degradation levels rose (Fig. 2E). Furthermore, the average consistency score of the REOs was higher in the degraded samples than in the highly degraded samples (Wilcoxon rank-sum test, p < 0.05) (Fig. 2E).

  • 4 RNA quantity

Adequate RNA quantity is essential for accurately measuring gene expression levels in high-flux experiments. However, multiple rounds of pre-amplification conducted before measuring low-input RNA samples may introduce significant bias29. Thus, we compared the gene expression of low-input samples with those of high-input samples in this study, as detailed in Table 1.

We obtained 32 paired low-input and high-input samples of human reference RNA from GSE124198. On average, 4577 genes exhibited twofold changes in expression levels. Conversely, an average consistency score of REOs was 83.81% (Fig. 3A). Furthermore, after eliminating 10% and 20% of gene pairs with the smallest absolute expression levels in each high-input sample, over 90% of REOs of gene pairs remained stable in low-input samples (Fig. 3A). Furthermore, we collected 174 paired low-input samples (ranging from 1 to 20 ng) and high-input samples (20 ng or 100 ng) of human reference RNA from SRP097611. Using 100 ng high-input samples as control, on average, 3620, 3788, 5402, 5730, and 4785 genes exhibited twofold changes in expression values, yet the average consistency scores of REOs were 93.22%, 91.95%, 83.52%, 80.42%, and 77.57% for the 20 ng, 10 ng, 5 ng, 2 ng, and 1 ng samples, respectively (Fig. 3B-C).Similarly, an average of 2564, 2432, 3491, and 3769 genes exhibited twofold changes in expression values, yet the average consistency scores of REOs were 94.54%, 94.15%, 84.72%, and 78.65%, for the 10 ng, 5 ng, 2 ng, and 1 ng samples, respectively, when using 20 ng high-input samples as the standard (Fig. 3B-C). The consistency scores of REOs showed a downward trend as the RNA input decreased (Fig. 3C).

Fig. 3
figure 3

Effect of RNA quantity on gene expression measurements and REOs. Scatter plots depict the consistency scores of REOs (pink, green and blue dots, left vertical axis) and the number of DEGs (purple dots, right vertical axis) among samples of different RNA quantity from GSE124198 (A). Boxplots depict the distribution of (B) number of DEGs and (C) consistency scores of REOs among samples of different RNA quantity from SRP097611.

Full size image

Similar results were observed in 28 paired samples of ovarian cancer from GSE18769 and GSE17572, as well as 8 paired samples of breast cancer from GSE113976 (Figures S3A–C).

  • 5 Measuring platforms

Due to differences in probe design principles, substantial variability in gene expression levels may occur in samples measured by various platforms34. To assess the effect of measuring platforms on gene expression measurements, we analyzed a set of 55 paired samples of breast cancer measured by the Illumina beadchip platform (GPL10558) and Illumina HiSeq platform (GPL11154) from GSE60789. On average, 9833 genes exhibited twofold changes in expression values, but the average consistency score of REOs was 73.28% between the paired samples. (Fig. 4A). Furthermore, an average of 75.44% and 77.43% of REOs of the remaining gene pairs in the samples measured by the beadchip platform were stable after removing 10% and 20% of the gene pairs with the smallest absolute expression levels from samples measured by the HiSeq platform (Fig. 4A).

  • 6 Laboratory sites

Fig. 4
figure 4

Effect of measuring platforms and laboratory sites on gene expression measurements and REOs. Scatter plots depict the consistency scores of REOs (pink, green and blue dots, left vertical axis) and the number of DEGs (purple dots, right vertical axis) among samples measured by different platforms from GSE60789 (A); among samples measured at different laboratory sites from GSE48035 (B) .

Full size image

Samples tested by different laboratories may exhibit batch effects, leading to systematic variations in gene expression values35. To assess the effect of laboratory sites on gene expression measurements, we obtained 12 paired samples measured by different laboratory sites from GSE48035. On average, 6062 and 5983 genes exhibited twofold changes in expression values, but the average consistency scores of REOs were 80.59% and 80.64% in samples measured at site R and site V compared to samples measured at site L (Fig. 4B). Furthermore, after excluding the bottom 10% and 20% of gene pairs, the average consistency scores increased from 82.98% and 83.01% to 84.86% and 84.83% in samples measured at site R and site V, respectively (Fig. 4B).

  • 7 Library preparation kits

The Illumina TruSeq kit and RiboZero kit have demonstrated strong performances in library preparation and are widely considered as the “gold standard” 3,16. Additionally, Sven Schuierer et al. demonstrated that the TruSeq kit produced superior experimental results for intact samples with various RNA quantities, whereas the RiboZero kit performed well for intact and degraded samples, and the RNA Access kit was suitable for highly degraded samples6. Accordingly, the samples were divided into control groups and case groups, as outlined in Supplementary Table S1.

Comparative analyses of gene expression values from six different groups of library preparation kits from GSE124198 revealed 5656 genes with twofold changes in expression values and an average 82.21% consistent REOs of gene pairs (Fig. 5A) However, two groups exhibited an average consistency score of REOs below 80%, potentially influenced by the effects of various RNA types (Total RNA vs. mRNA). When excluding the two groups, the average consistency score was above 85%, and increased to 88.38% and 90.80% after excluding the bottom 10% and 20% of gene pairs (Fig. 5A).

Fig. 5
figure 5

Effect of library preparation kits on gene expression measurements and REOs. Scatter plots depict the consistency scores of REOs (pink, green and blue dots, left vertical axis) and the number of DEGs (purple dots, right vertical axis) among samples using different library preparation kits from GSE124198 (A). Boxplots depict the distribution of (B) number of DEGs and (C) consistency scores of REOs among samples using different library preparation kits from SRP097611.

Full size image

Furthermore, we evaluated the effect of different kits in degraded and highly degraded samples from SRP097611. Samples with three degradation levels (intactness, degradation, and high degradation) using various library preparation kits were divided into five groups (Supplementary Table S1). On average, 5366, 7230, and 4220 genes exhibited twofold changes in expression values in case samples with three degradation levels (Fig. 5B). Conversely, the average consistency score of REOs were 77.85%, 75.56%, and 81.23% in case samples with three degradation levels (Fig. 5C).

Similar results were observed in five paired samples of breast cancer using different library preparation kits from GSE113976 and 16 paired samples of human reference RNA using different kits from GSE48035 (Figures S4A-B).

  • 8 Amplification kits

Research has indicated that the Nugen Ovation FFPE WTA System Kit, with a modified deparaffinization process, yields better results than Affymetrix IVT kit36,37, and that Nugen WT-Ovation FFPE System offers certain advantages over the standard Affymetrix IVT kit38. Thus, samples amplified using the Ovation system were designated as the control group, as shown in Table 1.

We analyzed 59 paired case and control samples within the same patients of diffuse large B-cell lymphoma from GSE19246. Between paired samples, there was an average of 315 genes with twofold changes in expression measurements and 81.81% of gene pairs with stable REOs in case samples. After eliminating 10% and 20% of the gene pairs with the closest expression levels, we found that the average consistency scores were 88.35% and 91.06%. (Fig. 6A). Furthermore, we collected eight paired samples within the same patients of colon cancer from GSE73883. On average, 1256 genes exhibited twofold changes in expression values, but the average consistency scores of REOs were 77.68% between case and control samples (Table1). After excluding the lowest 10% and 20% of gene pairs, the average consistency scores increased to 80.33% and 82.83% (Fig. 6B). Similar results were observed in 8 paired samples of breast cancer from GSE93338 (Figure S4C).

Fig. 6
figure 6

Effect of amplification kits on gene expression measurements and REOs. Scatter plots depict the consistency scores of REO (pink, green and blue dots, left vertical axis) and the number of DEGs (purple dots, right vertical axis) among samples using different amplification kits from GSE19246 (A) and GSE73883(B).

Full size image

Effect of multiple pre-analytical variables on gene expression measurements and relative expression orderings

All the aforementioned results derived from single-variable analyses. Furthermore, we assessed the robustness of gene expression measurements and REOs under the influence of two or three variables. The control group comprised high-quality samples with high-input, intactness, and/or those processed using kits demonstrating excellent performances, as outlined in Supplementary Table S2.

We evaluated the effect of preservation conditions combined with RNA quantity, library preparation kits or amplification kits on gene expression measurements and REOs based on paired samples from GSE113976, GSE93338 and GSE19246, respectively. On average, 7386, 10,590, and 706 and 222 genes exhibited twofold changes in expression values, yet the average consistency scores were 74.66%, 73.82%, 71.76%, and 78.69% in four comparisons with three combinations of two variables (Figs. 7A-C, Figure S5A, Supplementary Table S2). Upon removal of 10% and 20% of the gene pairs with the closest expression levels in the control samples, the average consistency scores for the remaining gene pairs increased from 75.76%, 75.41%, 73.87% and 81.41% to 76.88%, 76.41%, 75.93% and 83.91%, respectively (Figs. 7A-C, Figure S5A).

Fig. 7
figure 7

Effect of the combination of two pre-analytical variables on gene expression measurements and REOs. Scatter plots depict the consistency scores of REOs (pink, green and blue dots, left vertical axis) and the number of DEGs (purple dots, right vertical axis) among samples with different preservation conditions and RNA quantity from GSE113976 (A); among samples with different preservation conditions and library preparation kits from GSE113976 (B); among samples with different preservation conditions and amplification kits from GSE93338 (C); among samples with different RNA quantity and library preparation kits from GSE113976 (D) and GSE124198(E) .

Full size image

Furthermore, we assessed the influence of the combination of RNA quantity and library preparation kits or degradation levels based on paired samples from GSE113976, GSE124198 and SRP097611, respectively. On average, 7274, 9445 and 5780 genes exhibited twofold changes in expression values in case samples with combinations of two variables (Figs. 7D-E, Fig. 8A, Supplementary Table S2). Conversely, the average consistency scores of REOs were 74.94%, 77.70% and 85.26% respectively, and increased steadily after excluding the bottom 10% and 20% of the gene pairs in the case samples (Figs. 7D-E, Fig. 8B).

Fig. 8
figure 8

Effect of the combination of RNA quantity and degradation levels on gene expression measurements and REOs. Boxplots depict the distribution of (A) number of DEGs and (B) consistency scores of REOs among samples with different RNA quantity and degradation levels from SRP097611.

Full size image

Likewise, similar findings were observed when evaluating the effect of measuring platforms in conjunction with library preparation kits on gene expression measurements and REOs based on paired samples from GSE73883 (Figures S5B).

Moreover, we assessed the effect of the combination of three variables: RNA quantity, preservation conditions, and library preparation kits on gene expression measurements and REOs based on paired samples from GSE113976. On average, 10,191 genes exhibited twofold changes in expression values. The average consistency score of REOs was 73.54% in the case samples and increased to 75.21% and 76.31% after excluding the bottom 10% and 20% of the gene pairs, respectively (Fig. 9A). We also used 102 paired samples from SRP097611 to evaluate the effect of the combination of three variables: RNA quantity, degradation levels, and library preparation kits on gene expression measurements and REOs. We found that an average of 7525 genes exhibited twofold changes in expression values in case samples (Fig. 9B). However, we observed that the average consistency score was 77.25% in these samples (Fig. 9C).

Fig. 9
figure 9

Effect of the combination of three pre-analytical variables on gene expression measurements and REOs. Scatter plots depict the consistency scores of REOs (pink, green and blue dots, left vertical axis) and the number of DEGs (purple dots, right vertical axis) among samples with different RNA quantity, preservation conditions and library preparation kits from GSE113976 (A). Boxplots depict the distribution of (B) number of DEGs and (C) consistency scores of REOs among samples with different RNA quantity, degradation levels and library preparation kits from SRP097611.

Full size image

Higher Robustness of REOs of gene pairs than gene expression measurements

To quantify the effect of pre-analytical variables on gene expression measurements and REOs of gene pairs, we calculated five indices for each single-variable and multi- variable analysis, as shown in Table 2. All results were summarized in Fig. 10, Supplementary Table S1 and S2. Individual analysis of paired samples revealed Prop-FC was greater than Prop-REO in over 85% comparisons. Besides, these differences shown statistically significant in both single-variable and multiple-variable analyses (Wilcoxon rank-sum test, p < 0.05).

Table 2 Indices for quantifying the effect of pre-analytical variables.
Full size table
Fig. 10
figure 10

Comparative analysis of the effect of pre-analytical variables on gene expression measurements and REOs of gene pairs.

Full size image

Population analysis between control and case group revealed Prop-DEG was significantly higher than Prop-SRP (Wilcoxon rank-sum test, p < 0.05) in single-variable analysis. We failed to identify DEGs in 32.5% comparisons but reversal gene pairs in 76.25% comparisons. The latter completely encompasses the former. Excluding comparisons mentioned above, Prop-DEG was still significantly higher than Prop-SRP (Wilcoxon rank-sum test, p < 0.05). Group comparisons in multi-variable analyses could identify DEGs in 89% of case but reversed gene pairs in only 23% of cases. Similarly, Prop-DEG was significantly greater than Prop-SRP in other comparisons.

Furthermore, we also calculated the proportion of significant reversal gene pairs among all gene pairs involving DEGs, defined as Prop-DEG&SRP, in comparisons where both DEGs and reversal gene pairs were identified. The mean of Prop-DEG&SRP were 18.00% and 21.13% in single-variable and multi-variable analyses, respectively. And, the Prop-DEG&SRP remained below 20% in over 60% instances.

In conclusion, the above results demonstrate that both single and multiple pre-analytical variables significantly influence the expression measurements of thousands of genes. However, the REOs of gene pairs exhibit higher robustness against these pre-analytical variables.

Discussion

In this study, we systematically analyzed gene expression measurements affected by ten variables, namely, sampling methods, tumor sample heterogeneity, fixed time delays, degradation levels, preservation conditions, library preparation kits, amplification kits, RNA quantity, measuring platforms, and laboratory sites, from biospecimen collection to data generation. Both single and multiple variable analyses revealed that these pre-analytical variables induce twofold changes in expression measurements in thousands of genes, posing a significant challenge in transitioning from the basic to clinical research for biological biomarkers39,40. For the transcriptional signatures relying on the risk scores summarized from the expression of the signature genes, applying risk score thresholds directly to different samples proves problematic41,42,43. In contrast, we observed that most gene pairs maintained consistent REO patterns under the effect of single or multiple pre-analytical variables. The robustness of within-sample REOs of gene pairs across various pre-analytical conditions suggests that REO-based signatures could offer more reliable biological classification in translational medicine.

Variations in gene expression measurements have been extensively studied. The Micro Array Quality Control (MAQC) project revealed that the median coefficient of variation of gene expression measurements in replicate samples ranged from 5 to 15% within the same test sites, but increased to 10% to 20% across different test sites44. Additionally, the Sequencing Quality Control (SEQC) project found that variations of expression values arose from library construction45. Guan et al. demonstrated significant measurement variations in gene expression even with low-throughput PCR-based technologies25. Built on these findings, we expanded our analyses to include 10 pre-analytical variables, which helped to gain a more comprehensive understanding of the variation in gene expression measurements and within-sample REO-based analyses.

In principle, these pre-analytical variables inevitably have an effect on gene expression measurements and REOs of gene pairs. Fewer pre-analytical variables sound fewer DEGs and higher consistency of REOs, The consistency scores of REOs decrease as the number of DEGs increases. But our results have demonstrated higher robustness of REOs against these pre-analytical variables, which is probably attributed to qualitative character of REOs. While REO-based signatures might overlook subtle quantitative information in gene expression. It’s worth considering that subtle quantitative information of gene expression measurements tends to be unreliable25,41,42. Therefore, the perceived limitation of REOs analysis can actually be viewed as a unique advantage due to its robustness. Considering the sufficient number of gene pairs in a specific analysis, we can identify gene pairs that are robust to multiple variables46,47. These robust gene pairs can be identified offering potential for further clinical research purposes.

If researchers must rely on measurements of gene expression for biological research, it is essential to control for various pre-analytical variables or exclude genes affected by these factors. Failing to do so may lead to spurious associations or misinterpretations. We also recommended validating these findings across multiple datasets or cohorts to ensure their robustness and generalizability in different contexts. By implementing these strategies, researchers can strengthen the robustness and reproducibility of their gene expression studies, even in the presence of various pre-analytical variable influences.

Since all data utilized in this study were measured from human tumor tissues or reference RNA, the findings and conclusions are therefore confined to the domain of cancer research. Extending these results to other areas of study would require further analysis. Owing to limited data, multiple variable analyses could not be conducted for all variable combinations. Therefore, in the future, appropriate experiments should be designed. Furthermore, the sample sizes in the analysis of individual variables were limited, necessitating greater sample numbers in future research.

Conclusions

Single and multiple variable analyses showed that ten pre-analytical variables, spanning biospecimen collection to data generation, lead to twofold changes of expression measurements in thousands of genes. Conversely, REOs of gene pairs exhibit higher robustness under the influence of these pre-analytical variables, suggesting that REO-based signatures hold promise for clinical practice.