RNA sequencing and immunohistochemistry jointly improve tumor biomarker interpretation

Abstract

This study aimed to assess the correlation between RNA sequencing (RNA-seq) and immunohistochemistry (IHC) in detecting key cancer biomarkers across solid tumors, and then, to establish RNA-seq thresholds that accurately reflect clinical IHC classifications. Expression levels of nine biomarkers—ESR1, PGR, AR, MKI67, ERBB2, CD274, CDX2, KRT7, and KRT20—were analyzed in 365 formalin-fixed, paraffin-embedded samples from breast, lung, gastrointestinal, and other solid carcinomas. Correlations between RNA-seq data and IHC scores were determined using Spearman’s correlation coefficients, with RNA-seq cut-offs established to distinguish positive from negative IHC scores. The results revealed strong correlations for most biomarkers, with coefficients ranging from 0.53 to 0.89. RNA-seq thresholds were confirmed across internal and external cohorts, demonstrating high diagnostic accuracy (up to 98%) and precision in identifying biomarker expression levels. The analysis also highlighted the influence of tumor microenvironment and purity, particularly in the moderate correlation of 0.63 observed for PD-L1. Our study demonstrates that RNA-seq can serve as a robust complementary tool to IHC, offering objective and high-throughput biomarker assessment. The RNA-seq thresholds established provide a reliable method for determining biomarker positivity, supporting the integration of RNA-seq in clinical diagnostics to enhance precision, especially where tumor purity and microenvironment factors are significant.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-12780-3.

Introduction

RNA-based biomarkers are becoming increasingly valuable in modern oncology as robust prognostic tools with high sensitivity and specificity reflecting intricate regulatory processes within cancer cells. RNA biomarkers can also serve as predictive parameters for treatment decision-making^1–3. Evaluation of biomarker expression by RNA-seq can potentially circumvent drawbacks encountered in immunohistochemistry (IHC), the current standard in biomarker testing from tissues, such as variability across antibody clones, a lack of a universal scoring system applicable within and across cancer types, and the subjective nature of assessment by pathologists^4–6. While IHC remains the gold standard for routine clinical pathology due to its accessibility and cost effectiveness, RNA-seq offers standardized quantitative assessment across multiple biomarkers in a single assay. This uniformity is particularly valuable in translational research settings, where multiplexed and exploratory biomarker validation requires cross-marker comparability on the same tissue section. IHC application is often limited by the amount of tissue needed to measure important diagnostic or therapy-associated tissue markers, tested in clinical trials. Yet, to leverage RNA-seq effectively in clinical settings as a complement to IHC, the correlation between mRNA levels and their corresponding protein expression must be demonstrated. By improving methodological approaches for comparing and assessing these factors, including the impact of the tumor microenvironment (TME) and tumor purity on their correlations, we aim to provide a robust framework for integrating RNA-based biomarkers into cancer research and routine clinical diagnostics.

Several studies linking simultaneous measurement of RNA biomarkers to IHC data^7–11 revealed correlations between RNA and protein expression for well-known biomarkers such as steroid hormone receptors, human epidermal growth factor receptor (ERBB2, HER2), proliferation marker Ki-67, and the immune checkpoint molecule programmed death-ligand 1(CD274, PD-L1). However, other studies demonstrated considerably varying correlations, did not establish clear expression cut-offs for RNA-seq to reproduce established clinical IHC scoring thresholds, and did not compare data for common diagnostic markers, such as caudal-related homeobox transcription factor CDX2 and cytokeratins CK7 or CK20, critical for tumor origin identification and differential diagnosis of carcinomas^12–14. The application of RNA-seq for these biomarkers could enhance diagnostic accuracy and be helpful for pathologists as a screening tool.

In our study, we observed high correlations between IHC and RNA-seq analyses that led to the establishment of RNA-based cut-offs for key tumor predictive and diagnostic biomarkers: estrogen receptor (ER), progesterone receptor (PR), androgen receptor (AR), Ki-67, HER2, PD-L1, CDX2, CK7, and CK20. Selection of these specific biomarkers for evaluation was informed by their established roles in distinct cancer types and for supporting decisions about treatment (e.g., ESR1, ERBB2, CD274) and diagnostic utility (e.g., KRT7, KRT20, CDX2), representing a spectrum of clinical applications across solid tumors^15–18. A binary classifier was developed for each biomarker using mRNA expression level cut-offs to predict IHC score-based classifications of biomarker negativity and positivity in an internal cohort. The selected RNA-based cut-offs demonstrated high F1 scores, effectively discriminating between negative/low and positive/high expression, affirming their potential predictive accuracy. The performance of these cut-offs was consistent across diverse carcinoma cohorts, and their associations with clinical outcomes closely recapitulated those of IHC-derived predictions, underscoring the robustness of our approach and its promise in clinical utility with further validation. Our approach leverages RNA-seq to establish quantifiable thresholds for key biomarkers, potentially overcoming the subjective variability of traditional IHC assessments. This methodology not only enhances the precision and reproducibility of biomarker evaluation but also facilitates the integration of molecular data into clinical decision-making, setting a new standard in the pursuit of personalized medicine.

Methods

Internal cohort and tumor processing

The internal cohort consisted of formalin-fixed, paraffin-embedded (FFPE) tissue blocks from 365 patient samples (Supplementary Table 1). For 313 patients, FFPE blocks from the clinical program were provided initially and used for both RNA-seq and IHC testing. To confirm the diagnoses, pathologists examined hematoxylin and eosin (H&E) slides¹⁹ from the FFPE tumor blocks retrospectively to select samples with neoplastic cellularity of higher than 20%. Available IHC results for ER, PR, AR, Ki-67, HER2, PD-L1, CDX2, CK7, and CK20 were obtained retrospectively from CLIA-certified laboratories submitted to the BostonGene clinical program or stained and analyzed in the BostonGene laboratory from the FFPE blocks used for RNA-seq. These assays were performed as laboratory-developed tests (LDTs) validated under institutional protocols.

For the other 52 patients, tissue samples were originally provided as fresh-frozen tissue (FF) specimens (purchased from Cureline Tissue Bank [USA]), which were used for RNA-seq directly, and FFPE blocks for IHC testing were prepared in the BostonGene laboratory. PD-L1 results from external pathology reports were based on 22C3 clone (Dako pharmDx, Agilent Technologies, USA) or a validated LDT.

Each patient provided informed consent. The use of clinical samples was conducted in accordance with the Declaration of Helsinki and has been granted exemption from ethics approval by the Biomedical Research Alliance of New York (BRANY) Institutional Review Board (IRB; #22-12-938-853).

Public datasets

Public RNA-seq data were retrieved from the TCGA cohorts (phs000178)²⁰: BRCA (n = 1,056), COAD (n = 106), READ (n = 246), PRAD (n = 479), LUSC (n = 480), ESGA_AC²¹ (n = 468), BLCA (n = 398), PAAD (n = 150), SARC (n = 249), LUAD (n = 487), KIRC (n = 491), and OV (n = 360). Apart from the TCGA cohorts, we also used data from CPTAC Breast (phs000892, n = 109)²², PRJEB25780 (n = 42)²³, and SU2C-MARK (n = 64)²⁴ with corresponding clinical and available IHC data. Annotation for molecular subtypes of breast cancer was taken from Thennavan et al.²⁵.

The expression of biomarkers in isolated cells of the tumor microenvironment (TME) was assessed in a collection of BostonGene’s sorted cells used to create the Kassandra algorithm²⁶. The expression of biomarkers in tumor cells was evaluated in the collection of cell lines previously obtained from public databases, including ArrayExpress²⁷ and GEO²⁸. The full list of dataset IDs can be found in Supplementary Table 5.

Immunohistochemistry

IHC was performed using a fully automated research stainer (Leica BOND RX, Leica Biosystems, USA) with specific primary antibodies according to manufacturing guidelines (Supplementary Table 3) validated as LDTs. Each run included positive and negative controls. All stained slides and matching H&E sections were scanned with a Vectra Polaris (Akoya Biosciences, USA) scanner at ×20 magnification.

We utilized QuPath (version 0.3.2) with the positive cell detection algorithm to quantify nuclear immunostains (ER, PR, AR, Ki-67, etc.). The parameters were set to QuPath’s default for DAB chromogen, with adjustments for cell size and optical density thresholds calibrated on the control slides²⁹ (Supplementary Fig. 1). Membrane biomarkers PD-L1 and HER2 were scored visually by two pathologists^30,31 according to the clinical IHC cut-off guidelines^15,16,30,31 (Supplementary Table 3). Since CDX2, CK7, and CK20 do not have standardized diagnostic cut-offs, they were scored as either positive or negative with a cut-off of 1% tumor cells (TC) based on visual pathology assessment. Two pathologists reviewed each digital slide multiple times, discussed their observations and interpretations, and arrived at the final decision.

The IHC scores for ER, PR, AR, Ki-67, CDX2, CK7, and CK20 were calculated as the percentages of positively stained TC. HER2 clinical IHC scores were categorized as negative (0), low expression (1+, 2+), or high expression (3+). PD-L1 expression was assessed as either a combined positive score (CPS) or a percentage of positive TC, depending on the cancer type (Supplementary Table 3).

RNA sample preparation and sequencing

All tumor specimens from the internal cohort were prepared and sequenced in the BostonGene laboratory. For RNA isolation from FFPE samples, 10 µm-thick paraffin slices were trimmed from each FFPE tissue block using a microtome. RNA was extracted from 10 paraffin slices with the RNAeasy mini kit (Qiagen, Germany). The library was prepared using target enrichment with the SureSelect XT HS2 RNA kit (Agilent Technologies, USA). The SureSelect Human All Exon V7 + UTR (Agilent Technologies, USA) exome probe set was used for RNA hybridization and capture.

RNA from 52 FF tissue samples was extracted with AllPrep DNA/RNA Mini Kit (Qiagen, USA). Libraries were prepared with TruSeq^® Stranded mRNA Library Prep (Poly-A mRNA; stranded; Illumina, Inc., USA).

Libraries were sequenced on NovaSeq 6000 (Illumina, Inc., USA) as paired-end reads (2 × 150) with a targeted coverage of 50 mln reads per sample.

RNA-seq processing

Bulk RNA-seq fastq files were processed by Kallisto version 0.42.4 for Linux³². The Kallisto index file was downloaded from the Xena project to be consistent with the TCGA expression data³³. The protein-coding transcripts as well as IGH/K/L- and TCR-related transcripts were retained, while the noncoding RNA as well as histone- and mitochondria-related transcripts were removed, resulting in 20,062 protein-coding genes. Gene expression was quantified as transcripts per million (TPM) and log2-transformed³⁴. A pseudo count of one was added to the TPM to avoid negative values for expression levels.

Batch correction

For the internal cohort samples with mRNA captured using target enrichment, the batch effect was eliminated using Procrustes³⁵. After batch correction, biomarker expression levels in the internal cohort corresponded to those observed in poly(A)-enriched FF samples and could be compared to the TCGA cohort (Supplementary Fig. 3B).

Assessment of tumor proliferation signature

The functional gene expression signature (FGES) associated with tumor proliferation was estimated using the ssGSEA method³⁶. The signature was composed of the following genes: AURKA, AURKB, BUB1, CCNB1, CCND1, CCNE1, CDK2, CETN3, E2F1, ESCO2, MCM2, MCM6, MKI67, MYBL2, PLK1³⁷.

DNA sample preparation and whole exome sequencing (WES)

DNA was extracted from FF solid tumor samples using AllPrep DNA/RNA Mini Kit (Qiagen, USA), from FFPE solid tumor samples using AllPrep DNA FFPE Kit (Qiagen, USA), and from normal tissue (whole blood, PBMC, saliva) using QIAmp DNA Blood Mini Kit (Qiagen, USA). The library from FFPE tissue was prepared using the exome capture SureSelect XT HS2 DNA kit (Agilent Technologies, USA). The All Exon V7 exome probe (Agilent Technologies, USA) was used for DNA hybridization and capture. In total, 316 samples that underwent WES had RNA-seq expression data and IHC scores for at least one marker.

Purity estimation and copy number alteration (CNA) calling using WES

The reads were filtered and stripped from adapters using Cutadapt version 3.4³⁸ and aligned to the human reference genome GRCh38 (GRCh38.d1.vd1 assembly) using BWA version 0.7.17³⁹. Base quality scores were recalibrated by BaseRecalibrator and ApplyBQSR from GATK version 4.1.2.0⁴⁰. Duplicate reads were removed using the MarkDuplicates algorithm of the Picard tool version 2.20.7⁴¹.

Coverages were extracted from BAM files using the snp-pileup utility program from the FACETS tool version 0.5.14⁴². CNA for ERBB2 were calculated by FACETS. Tumor purity was estimated for each tumor sample using Sequenza version 2.1.2⁴³.

Statistical analysis

Data were statistically analyzed and plotted using Python version 3.7.5. F1 scores and other performance metrics for classification were calculated with the sklearn package version 1.1.1 and internal functions. Statistical data are provided in the text, figure legends, and figures. P values between groups were calculated by the Mann–Whitney U test from SciPy package version 1.8.1⁴⁴. The relationship between RNA-seq and IHC expression was assessed by Spearman’s rank correlation coefficient using the spearmanr function from SciPy package version 1.8.1. We used Spearman’s rank correlation coefficient assuming a monotonically increasing relationship between RNA expression in the bulk sample measured as log₂(TPM + 1) and protein expression measured as the percentage of stained cells on the slide. To determine survival curves for each sample, we used KaplanMeierFitter in lifelines version 0.27.3⁴⁵. Survival differences were assessed using the log-rank test in lifelines version 0.27.3. All graphs were plotted using Matplotlib version 3.5.2⁴⁶ and Seaborn version 0.11.2⁴⁷.

Development of RNA-based single-gene classifiers

A binary classifier was developed for each marker using mRNA expression levels of single genes to predict the groups of IHC scores. For each RNA expression level increment of 0.5 log₂(TPM + 1), we assessed the classification performance by calculating the F1 score. The RNA-seq thresholds were determined by maximizing the F1 score for each biomarker in our internal cohort using iterative cutoff selection. We selected the expression thresholds for ESR1, PGR, AR, MKI67, ERBB2, CD274, CDX2, KRT7, and KRT20 with the highest F1 scores indicating the highest concordance with respective IHC scores. These thresholds were then validated in external cohorts (TCGA, SU2C).

Results

Correlations between IHC and RNA-seq data of biomarkers in solid carcinomas

We first investigated mRNA expression distributions and correlations between levels of single genes measured by RNA-seq (Fig. 1A) and corresponding protein expression levels defined as IHC scores across multiple cancer types (Fig. 1B–C, Supplementary Fig. 2). The correlations between IHC scores and mRNA expression levels were comparable across IHC performed in-house and through external pathology laboratories (Supplementary Fig. 3A). Hence, we combined both datasets into the internal cohort for subsequent analyses (n = 365, Supplementary Table 1). Batch effects were corrected for mRNA expression data for adequate downstream comparisons with validation datasets (Supplementary Fig. 3B).

Fig. 1.

Overview of the internal cohort. (A) Heatmap illustrating the RNA expression levels (measured in log₂(TPM + 1)) of nine key biomarkers—ESR1, PGR, ERBB2, AR, MKI67, CD274, CDX2, KRT7, and KRT20—across a diverse set of solid tumors, including breast carcinomas, gastrointestinal adenocarcinomas, non-small cell lung cancer (NSCLC), and other malignancies for 365 samples. The samples were sorted based on ESR1 expression inside a group of diagnoses, with each column representing a single sample and each row corresponding to one biomarker. The color gradient indicates the gene expression level based on RNA-seq in log₂(TPM + 1) with varying minimum and maximum values. It effectively contextualizes the relationship between RNA expression levels and specific cancer diagnoses, showcasing the broader patterns and trends in biomarker expression within the studied cohort. (B) The distribution of IHC-stained slides across biomarkers, color-coded by diagnosis Panels A and B share the same color palette indicating the specific diagnosis groups: breast carcinomas samples are highlighted in pink, gastrointestinal adenocarcinomas in blue, NSCLC in green, and other cancers in grey. (C) The distribution of stained slides with estimated IHC scores according to IHC scores is displayed, categorized into negative or 0 for HER2 (purple), positive or 3 + for HER2 (red), and 1–10 for ER or 1 + or 2 + for HER2 (orange).

We divided tumor samples of the internal cohort into four main groups: the breast group, which included breast carcinomas independent of hormone receptors and HER2 expression status; the gastrointestinal (GI) group, which included colorectal, gastric, and gastroesophageal adenocarcinomas; and the non-small cell lung carcinoma (NSCLC) group, which included adenocarcinomas and squamous cell carcinomas of the lung. All samples, including those that did not fit in any of the three main groups, were pooled into the solid cancer group and analyzed together to measure the overall correlation for solid tumors (Table 1, Supplementary Table 2).

Table 1.

Correlation between IHC and RNA-seq data for ER, PR, AR, Ki-67, HER2, and PD-L1.

Protein/gene	Spearman’s rank correlation coefficients of RNA-seq (log2 TPM) and IHC score (% positive tumor cells or combined positive score for PD-L1)
	Breast	Solid
ER/ESR1	0.7 (n = 136, P < 0.001)	0.81 (n = 167, P < 0.001)
PR/PGR	0.89 (n = 120, P < 0.001)	0.88 (n = 148, P < 0.001)
AR/AR	0.82 (n = 18, P < 0.01)	0.59 (n = 87, P < 0.001)
Ki-67/MKI67	0.86 (n = 98, P < 0.001)	0.86 (n = 166, P < 0.001)
HER2/ERBB2	0.68 (n = 182, P < 0.001)	0.68 (n = 230, P < 0.001)
PD-L1/CD274	0.53 (n = 42, P < 0.001)	0.63 (n = 112, P < 0.001)

Steroid hormone receptor biomarkers: ER, PR, AR

ER, PR, and AR are paramount in identifying hormone-driven cancers, such as breast, endometrial, and prostate cancer, respectively^15,48–50. Their expression levels not only dictate the potential responsiveness to hormone therapy, but also inform on the tumor’s biological behavior and prognosis^15,27,45,51. Therefore, we included samples from the breast and solid cancer groups in the correlation analysis of these steroid hormone receptors (Fig. 2A–C).

Fig. 2.

Establishing mRNA expression-based cut-offs to predict IHC classes of biomarker positivity and negativity. (A–F) Сorrelation between IHC and RNA-seq expression of each biomarker for breast and solid cancer samples in the internal cohort (first column). ROC AUC metrics show the performance of each RNA-seq-based single-gene classifier in distinguishing the IHC classes, i.e., positive or negative expression according to IHC-based clinical thresholds for each biomarker (second column), for all Solid cancers of the internal cohort. The selected RNA-seq cut-offs with the highest F1 scores represent the mRNA expression levels that best recapitulated the IHC classes scored with nucleic (A–D) or membrane (E, F) staining in tumor and immune cells (solid cancer group, third column, Mann–Whitney U test). Validation of RNA-seq cut-offs on public datasets: PRJEB25780 for PD-L1 and TCGA for the other biomarkers (last column, Mann–Whitney U test). Box plots represent the median and the 25th and 75th percentiles, and whiskers extending to the smallest and largest value within the 1.5x interquartile range.

A strong correlation was observed between ER protein levels and mRNA expression in the breast group (ρ = 0.7, n = 136 samples). Combined solid cancer samples demonstrated a very strong correlation (ρ = 0.81, n = 167) between IHC and RNA-seq-based ER expression. For PR, very strong correlations were observed in both the breast and solid cancer groups (ρ = 0.89, n = 120 and ρ = 0.88, n = 148 respectively). AR exhibited a very strong correlation in the breast group (ρ = 0.82, n = 18) and a moderate correlation in the solid cancer group (ρ = 0.59, n = 87; Table 1).

Proliferation marker: Ki-67

IHC scores and mRNA levels for proliferation marker Ki-67 demonstrated a positive linear relationship across all analyzed cancer types. The breast and solid cancer groups showed higher correlation coefficients (ρ = 0.86 for both, P < 0.001, n = 98 and n = 166, respectively; Table 1; Fig. 2D, Supplementary Fig. 4A) than the GI samples (ρ = 0.6, n = 24, P < 0.01; Supplementary Fig. 4A).

HER2 and PD-L1

Clinical IHC scores for growth factor receptor HER2 consistently correlated with ERBB2 mRNA expression levels in the solid cancer and breast groups (ρ = 0.68 for both; P < 0.001; n = 230 and n = 182, respectively; Table 1; Fig. 2E), as well as in the GI group (ρ = 0.6; Supplementary Fig. 4A). IHC CPS values and mRNA levels for the immune checkpoint molecule PD-L1 (CD274) correlated moderately to strongly across all groups: breast group, ρ = 0.53 (P < 0.001, n = 42; Fig. 2F); GI group, ρ = 0.6 (P < 0.001, n = 36); NSCLC group, ρ = 0.69 (P = 0.04, n = 9); and solid cancer group, ρ = 0.63 (P < 0.001, n = 112; Supplementary Fig. 4A). Tumor cell values for PD-L1 IHC in the NSCLC group also demonstrated a moderate correlation with mRNA expression levels (ρ = 0.62, n = 21; Supplementary Fig. 4B).

Diagnostic biomarkers: CDX2, CK7, CK20

Protein and mRNA levels of CDX2 in the solid cancer group demonstrated a significant positive correlation (ρ = 0.82, n = 33, P < 0.001). In the GI group, the correlation was weak (ρ = 0.38, n = 10, P = 0.3; Supplementary Fig. 4C). The RNA-seq and IHC data for CK7 correlated strongly in both the solid cancer (ρ = 0.84, n = 62, P < 0.001) and GI groups (ρ = 0.87, n = 24, P < 0.001; Supplementary Fig. 4D). CK20 expression levels also strongly correlated in both the solid cancer (ρ = 0.79, n = 46, P < 0.001) and GI groups (ρ = 0.89, n = 21, P < 0.001; Supplementary Fig. 4E).

Establishing RNA-seq thresholds to classify positivity or negativity of biomarkers in tumors

Next, we aimed to establish RNA expression-based thresholds to distinguish between clinically significant IHC-based biomarker levels in carcinomas. A binary classifier was developed for each marker to predict the IHC score class of a tumor sample from RNA-seq data. IHC served as the reference standard, providing the ground truth for determining the levels of biomarkers in carcinomas. For all markers except HER2, we established thresholds differentiating between negative IHC scores (low expression, below the threshold) and positive IHC scores (high expression, above the threshold; Table 2, Supplementary Table 3). For HER2, we selected two thresholds: one to distinguish negative (IHC 0) from positive cases (IHC 1+, 2+, 3+), and the other to distinguish samples with negative (IHC 0) and low protein expression (1+, 2+) from those with high expression (3+). Detailed classification performance metrics of the RNA expression thresholds, such as F1 scores, accuracy, error rate, precision, negative predictive value (NPV), sensitivity, and specificity, are outlined in Supplementary Table 4. We selected the RNA-seq-based cut-off with the highest F1 score for each marker to accurately reproduce the IHC score classification (Supplementary Fig. 4F). Overall, the classifiers for all six predictive and prognostic biomarkers demonstrated high performance in separating IHC classes, with ROC AUC scores of 0.84–0.99 (Fig. 2A–F).

Table 2.

Concordance between IHC clinical thresholds and RNA-seq gene expression.

IHC clinical score for specific diagnoses	IHC threshold	Corresponding RNA-seq threshold, log2(TPM + 1)	F1 score among specific diagnoses	F1 score among all solid group diagnoses
ER+ breast carcinomas	≥ 1% TC	> 2.5	0.96 (n = 136)	0.94 (n = 167)
ER-high breast carcinomas	≥ 10% TC	> 3.5	0.98 (n = 136)	0.98 (n = 167)
PR+ breast carcinomas	≥ 1% TC	> 2	0.9 (n = 120)	0.9 (n = 148)
AR+ breast carcinomas	≥ 10% TC	> 3.5	0.97 (n = 18)	0.86 (n = 87)
Ki67-high breast carcinomas	≥ 20% TC	> 3.5	0.93 (n = 98)	0.92 (n = 166)
HER2+ breast carcinomas	> 0	> 6	0.83 (n = 182)	0.81 (n = 230)
HER2-low breast carcinomas	< 1+, 2+ <	> 6 <8.5	0.8 (n = 182)	0.78 (n = 230)
HER2-high breast carcinomas	= 3+	> 8.5	0.86 (n = 182)	0.85 (n = 230)
PD-L1-high breast carcinomas	≥ 10 CPS	> 2.5	0.61 (n = 42)	0.69 (n = 112)
PD-L1-high NSCLC	≥ 50% TC	> 3.5	0.67 (n = 22)	0.62 (n = 39)

We observed distinct distributions of RNA-seq data between tumor samples with positive and negative IHC scores for nuclear biomarkers ER, PR, AR, and Ki-67 in the internal cohort (P < 0.001 for all comparisons, Fig. 2A–D). Importantly, RNA expression-based cut-offs efficiently differentiated between the two IHC score classes, with high F1 scores for the internal cohort (F1 scores: 0.98, 0.90, 0.86, and 0.92, respectively). The analysis of ESR1 expression showed that the mRNA cut-off of 3.5 log₂(TPM + 1), which separated the TC 0–10% versus 10–100% IHC classes (F1 score = 0.98, Fig. 2A), performed better than the 2.5 log₂(TPM + 1) cut-off, which separated TC 0-1% versus 1-100% (F1 score = 0.94, Supplementary Fig. 4G, H). RNA-seq thresholds for each biomarker not only achieved high F1 scores, but also demonstrated exceptional accuracy and precision pivotal for clinical diagnostics. For instance, ER and PR thresholds displayed impressive accuracy rates of 0.97 and 0.90, respectively, with corresponding precision (positive predictive value, PPV) rates of 0.99 and 0.98, ensuring the reliability of these tests in identifying hormone receptor statuses. The AR biomarker, essential for prostate cancer and certain breast cancers subtypes, showed a precision of 0.83 and an accuracy of 0.93, reinforcing its diagnostic validity (Supplementary Table 4).

ERBB2 mRNA expression distributions differed significantly between samples with negative (IHC HER2 0) and positive (IHC HER2 1 + or higher) scores (P < 0.001), and between samples with negative/positive-low (IHC HER2 0, 1+, 2+) and positive-high (3+) scores (P < 0.001). The classifier performed well across all analyzed samples; F1 scores were high in the internal cohort (0.81 and 0.85, respectively, Fig. 2E).

For PD-L1, F1 scores were markedly lower, indicating reduced discriminative performance based on RNA-seq-derived cut-offs (Fig. 2F, Supplementary Fig. 4F). Remarkably, the RNA-seq threshold of > 2.5 log₂(TPM + 1) consistently discriminated against PD-L1 expression across a range of IHC thresholds, including CPS 1, CPS 10, and TC 1%. This observation suggests that 2.5 log₂(TPM + 1) CD274 expression may serve as a uniform cut-off for PD-L1 expression across these clinical IHC benchmarks. However, at the higher IHC threshold of 50% TC, the RNA-seq cut-off increased to > 3.5 log₂(TPM + 1), which corresponded to F1 scores of 0.67 for NSCLC and 0.62 for other solid tumors (Supplementary Fig. 4B, Table 2). RNA-based boundaries were less distinct in the PD-L1 and HER2-low groups compared to other biomarkers. This was likely due to the significant expression of these biomarkers in the TME, including epithelial, endothelial, immune, and fibroblast cells (Fig. 3D), which can contribute to the detected RNA levels in tumor samples.

Fig. 3.

Applicability of the derived RNA-seq cut-offs and effect of cellular composition. (A–B) Application of the derived RNA-seq cut-offs (gray horizontal line) for ESR1, PGR, AR, MKI67, ERBB2, and CD274 on data from public TCGA cohorts (A) and cancer cell lines (B). (C) Correlation of WES-derived tumor purity and RNA expression for ESR1, PGR, AR, MKI67, ERBB2, and CD274 (Spearman’s rank correlation coefficient, ρ). (D) Application of the selected RNA-seq cut-offs to sorted cells. Box plots represent the median, 25th and 75th percentiles, and whiskers extending to the smallest and largest values within the 1.5x interquartile range.

The IHC analysis of CDX2 (n = 33) showed a distinctive expression pattern in gastrointestinal adenocarcinomas, particularly in colorectal cancer⁵². We identified > 1 log₂(TPM + 1) as the cut-off for positive RNA expression with an F1 score of 0.99, indicating a high diagnostic performance of the test reflecting the IHC classification (Supplementary Fig. 4C). The difference between the IHC-negative and -positive groups reached statistical significance (P < 0.0001), confirming the robustness of the selected cut-off. CK7 is traditionally associated with carcinomas originating from the lung, breast, or upper GI tract¹². We identified the > 5 log₂(TPM + 1) cut-off for a positive score in the test cohort (n = 62) with F1 = 0.89 (Supplementary Fig. 4D). The analysis of CK20 (n = 46), which is commonly expressed in colorectal carcinomas⁵³, revealed that the RNA cut-off of > 3.5 log₂(TPM + 1) was highly congruent with positive staining (F1 = 0.99; Supplementary Fig. 4E).

In our study, RNA-seq was evaluated as a screening tool to assess its accuracy in biomarker detection across various cancer types (Supplementary Table 4). Among 167 samples, the RNA-seq threshold of 3.5 log₂(TPM + 1) accurately classified 110 out of 115 ER-positive cases (1% of TC and higher) and 44/52 ER-negative cases, showing a promising accuracy of 92%. Similarly, the RNA-seq threshold of 8.5 log₂(TPM + 1) correctly identified 11/13 cases of HER2-high (3+) and 215/217 negative/HER2-low cases (0, 1+, 2+), achieving an accuracy of 98%. However, challenges arose in distinguishing HER2-low cases (1+, 2 + only); the threshold range of > 6 < 8.5 log₂(TPM + 1) correctly identified only 103/146 HER2-low cases, achieving an accuracy of 74%, with 43 false positives and 16 false negatives (Supplementary Table 4). The Ki67 proliferation index with an RNA-seq threshold of > 3.5 log₂(TPM + 1) correctly classified 101/103 Ki67-high cases (more than 20% of TC) and 48/63 Ki67-low cases, yielding 92% accuracy. In contrast, PD-L1 detection (10 CPS and more) presented a mixed outcome; the threshold of 2.5 log₂(TPM + 1) tested on 112 cases presented an overall accuracy of 80%; only 24/39 high-expression and 66/73 low-expression cases were correctly identified. The diagnostic accuracy for CDX2, CK7, and CK20, markers that may support determination of tumor origin, was promising, ranging from 87 to 100% (Supplementary Table 4).

These findings highlight the potential of RNA-seq to accurately replicate IHC findings, suggesting its viability as a high-throughput and objective biomarker assessment tool in oncology. Further refinement and validation are required for clinical application.

Applicability of RNA-seq thresholds in validation datasets

After defining our RNA expression-based thresholds for the internal cohort, we tested them on public datasets, including on TCGA cohorts (Figs. 2A-E and 3A, Supplementary Figs. 4B-E), and CPTAC-BRCA (Supplementary Fig. 4H). Similar to their performance with the internal cohort, the RNA-seq thresholds for all nine biomarkers separated negative and positive protein expression groups for TCGA and CPTAC Breast cohorts, predominantly within F1 scores of 0.71–0.99.

We also applied the RNA-seq cut-offs to the biomarker expression data of TCGA cohorts for different cancer types (Fig. 3A) and cancer cell lines (Fig. 3B, Supplementary Table 5). We observed distinct boundaries of low and high expression levels for ESR1, PGR, and AR across different tumor types. The highest expression (above RNA-seq cut-off) of the three steroid receptors was mainly seen in breast carcinomas, as well as in prostate carcinomas for AR (Fig. 3A), in agreement with the expression levels observed in cancer cell lines (Fig. 3B) and as expected for these cancer types^48,54. The ERBB2 RNA-based lower cut-off aiming to separate negative and positive HER2 IHC scores delineated between low and high mRNA expression, whereas the higher cut-off for separating cases with very high protein levels effectively segregated the outliers. The distributions of MKI67, ERBB2, and CD274 expression varied more considerably across different tumor types (Fig. 3A). Biological underpinnings, such as sex-specific expression of the steroid receptors and tissue-specific or TME-derived expression of biomarkers, also contributed to the observed RNA levels in different malignancies, underscoring the significance of expression profiles that accurately characterize cancer subtypes and tumor compartments.

Effect of cellular composition in tumors on total mRNA expression

The assessment of IHC scores for ER, PR, AR, Ki-67, and HER2 is based exclusively on tumor cell counts, whereas RNA-seq gene expression measurements represent an average of contributions from tumor cells, the TME, and normal cells in a biopsy. This discrepancy may affect the comparability of IHC scores with RNA expression levels. To ascertain how the cellular composition of samples affects RNA-seq data, we examined the relationship between tumor purity, reflecting malignant cell content, and gene expression levels (Fig. 3C).

We observed that ESR1 mRNA expression in ER-positive samples exhibited a positive correlation with tumor purity (ρ = 0.35, P < 0.001), while ER-negative samples showed a negative association (ρ = − 0.53, P = 0.002). This suggests that the observed mRNA levels came from combined ESR1 expression in tumor, normal mammary epithelial, and TME cells (Fig. 3D). Conversely, in ER- and PR-negative breast cancers, an increase in non-tumoral cellular components in samples with lower tumor purity corresponded to elevated ESR1 expression. Such dynamics may hamper the delineation between ER-negative and -positive cancers, particularly in samples with lower tumor purity (< 40%), highlighting the need for sophisticated algorithms to deconvolute gene expression derived from malignant and non-malignant cells in bulk RNA-seq analyses.

No significant associations were detected between tumor purity and expression levels for PR (ρ = 0.06, P = 0.54) and AR (ρ = 0.05, P = 0.74). A modest positive correlation was observed for MKI67 (ρ = 0.25, P = 0.0062) and ERBB2 (ρ = 0.32, P < 0.001), which could marginally influence the assessment of the RNA-seq expression status. Notably, CD274 expression was inversely correlated with tumor purity (ρ = − 0.37, P = 0.0016), implicating immune cells as significant contributors to the gene expression, especially as tumor purity diminished (Fig. 3C). Given that the CPS scoring system accounts for the presence of PD-L1 on both tumor and immune cells, it may provide a more accurate reflection of the RNA-seq data than individual scoring systems based on TC and immune cell content.

To further delineate the contributors to gene expression within samples, we analyzed mRNA expression in isolated cells of the TME (Fig. 3D). ESR1, PGR, and ERBB2 expression detected in fibroblasts exceeded our established cut-offs, indicating their potential role in affecting expression level quantification of these markers. The contribution of immune cells to the expression of MKI67, ERBB2, and AR was less pronounced. Only CD274 showed increased expression attributable to the presence of macrophages, T cells, and neutrophils. Our findings underscore the variable impact of immune cell infiltrates on the expression of different markers in tumor biopsies.

Comparison of Ki-67 IHC scoring with tumor proliferation gene signature

To refine the correlation analysis between RNA-seq and IHC-derived expressions of MKI67, we expanded our investigation to include a FGES encompassing a suite of genes associated with cell-cycle regulation and tumor progression (Supplementary Fig. 5A). We aimed to mitigate potential technical errors arising during sequencing through the incorporation of a broader set of genes. Contrary to expectations, the inclusion of this composite signature did not enhance the correlation (ρ = 0.85, P < 0.001), which remained consistent with that of the MKI67 gene alone (ρ = 0.86, P < 0.001, Fig. 2D). This observation suggests that the FGES, despite its composite nature, can be employed comparably to Ki-67 IHC testing to assess the proliferation index of tumor cells. Furthermore, our results support the hypothesis that the mRNA expression cut-off of the MKI67 gene alone can emulate a broader subset of proliferation-associated genes within the FGES.

Concordance between CNA events in ERBB2 RNA-seq cut-offs and IHC thresholds

CNA events in the ERBB2 gene aligned well with both the IHC clinical scores and mRNA expression-based cut-offs (Supplementary Fig. 5B). The IHC 3+ group was characterized by ERBB2 amplifications and significant mRNA overexpression, and did not show any gene deletions. Similarly, none of the samples with RNA expression levels exceeding 8.5 log₂(TPM + 1) had gene deletions. Moreover, the lower RNA-based cut-off of 6 log₂(TPM + 1), reflecting the separation between negativity and low positivity of HER2 (1+, 2+ based on IHC), effectively distinguished ERBB2 deletions and neutral events from amplifications and gains. This serves as an additional validation for the accuracy of the ERBB2 thresholds.

Clinical applicability of the RNA thresholds

To assess the prognostic and predictive values of the six biomarkers routinely tested in clinical practice through IHC, we analyzed the selected RNA-seq cut-offs on public cohorts. The established mRNA cut-offs for ER and PR, which are critical in determining hormone receptor status in breast cancers, and for HER2, whose overexpression is also pivotal in guiding clinical management, stratified TCGA samples of stage III and IV breast carcinomas into groups that corresponded to the IHC surrogate subtypes (Fig. 4A). According to St Gallen Consensus⁵⁵, the IHC molecular surrogate subtypes of breast cancer are identified by specific IHC cut-offs for hormone receptor (HR) positivity, HER2 status, and Ki-67 proliferation index as lower or higher than 20% of tumor cells⁵⁶: Luminal A (HR+/HER2– with low Ki-67), Luminal B (HR+/HER2– with high Ki-67 or HR+/HER2+), HER2-enriched (HR–/HER2+), and Triple-Negative (HR–/HER2–) subtypes. The RNA-defined subtyping of the breast tumors (ESR1+ PGR+ ERBB2+; ESR1+ PGR+ ERBB2–; ESR1– PGR– ERBB2+; ESR1– PGR– ERBB2–) strongly correlated with the IHC-defined subtypes (P < 0.00001, Supplementary Fig. 5C). Moreover, breast cancer subtypes defined by IHC (P = 0.003) and RNA-seq (P = 0.05) exhibited similar overall survival trends for stage III and IV tumors (Fig. 4A).

Fig. 4.

Clinical applicability of RNA-seq cut-offs in comparison with IHC thresholds. (A) RNA-seq cut-off effectively identified breast cancer subtypes in the validation dataset (stages III and IV, TCGA-BRCA, n = 237, multivariate logrank test), yielding overall survival (OS) results similar to IHC. Hormone receptor (HR) status corresponded to ER/ESR1 and PR/PGR expression. Negative IHC annotation for ER, PR, and HER2, as well as the negative mRNA cut-offs for the corresponding genes (ESR1, PGR, and ERBB2) were used to define triple-negative breast cancer (TNBC). (B) Disease-free intervals in ESR1+ ERBB– MKI67-high versus ESR+ ERBB– MKI67-low breast cancer patients (Stages I–III) classified by the elevated RNA cut-off of 4 log₂(TPM + 1). Statistical analysis was performed using the multivariate logrank test. (C) The CD274 (TC) RNA-seq threshold of 3.5 log₂(TPM + 1) showed similar differences in survival between the PD-L1–low and PD-L1–high groups as the 50% TC IHC threshold in lung cancer (SU2C-MARK lung cohort, n = 64, multivariate logrank test). (D) The CD274 (CPS) RNA-seq threshold of 2.5 log₂(TPM + 1) significantly diversified responders from non-responders to pembrolizumab in metastatic gastric cancer (PRJEB25780, n = 42). SD: stable disease; PD: progressive disease; CR: complete response; PR: partial response.

Ki-67 was specifically evaluated in breast cancer subtypes where its expression is particularly indicative of proliferation rate⁵⁷, and therefore, prognosis and potential treatment paths. The RNA-based ESR1+ ERBB2– MKI67-high subtype corresponded to the Luminal B surrogate subtype characterized by ER+ HER2– and Ki-67 of 20% or more TC by IHC scoring. To improve the accuracy of our subtype classification and its prognostic utility, we adjusted the RNA-seq threshold to 4 log₂(TPM + 1), which significantly separated the disease-free intervals (DFI) between the Luminal subtypes within the ESR1+ ERBB2– cohort (P = 0.01, Fig. 4B). However, the 3.5 log₂(TPM + 1) cut-off demonstrated insufficient statistical power to reliably discern between MKI67-low (Luminal A) and MKI67-high Luminal B (Supplementary Fig. 5D). Specifically, the ESR1+ ERBB2– MKI67-high group showed a shorter DFI compared to the MKI67-low group (77.8% versus 88.1%, respectively, for the 4 log₂(TPM + 1) cut-off; and 83.4% versus 86.9%, respectively, for the 3.5 log₂(TPM + 1) cut-off), illustrating the critical impact of precise RNA-seq thresholding on clinical outcome prediction.

We tested the clinical performance of the obtained cut-off for PD-L1 in the SU2C-MARK lung cohort (n = 64) and the PRJEB25780 gastric cancer cohort (n = 42) in predicting survival and response to immune checkpoint inhibitors, respectively. Beforehand, we compared PD-L1 expression profiles between these cohorts, other public datasets, and the internal cohort used for establishing the RNA-based cut-off, to ascertain that the distributions of CD274 expression levels were comparable and, therefore the use of the thresholds sustainable (Supplementary Fig. 5E).

In the SU2C-MARK lung cohort, we compared progression-free survival (PFS) after the initiation of PD-1/PD-L1 therapy in patients with high CD274 expression versus low and negative CD274 expression using a cut-off of 3.5 log₂(TPM + 1), which corresponded to the IHC-based threshold of 50% TC. Categorizing patients into RNA-defined low (< 3.5 log₂(TPM + 1)) and high (> 3.5 log₂(TPM + 1)) gene expression groups resulted in significant stratification of PFS (P = 0.04; Fig. 4C). The PFS observed with the RNA expression-based stratification closely reapitulated that seen with the IHC-defined low (< 50% TC) and high (≥ 50% TC) protein expression groups, and both stratification approaches yielded significant separation between the two groups (P = 0.04 and P = 0.006, respectively).

We also tested the predictive value of RNA-seq cut-off for CD274 in the PRJEB25780 cohort of metastatic gastric cancer. The obtained threshold of 2.5 log₂(TPM + 1), which corresponded to CPS 10, significantly differentiated between responders and non-responders to pembrolizumab (P = 0.006, Fig. 4D, Supplementary Fig. 5F). Altogether, our established RNA expression-based cut-offs for biomarker positivity assessment in tumors hold promise in supporting clinical evaluations.

Discussion

Findings from this study support the value of RNA-based biomarkers in the clinical toolset by elucidating correlations between protein and mRNA expressions based on IHC and bulk RNA-seq, respectively, for key clinical and diagnostic biomarkers in pathology practice. The selection of biomarkers was guided by their routine application in clinical diagnostics, either as predictive markers of therapeutic relevance (ER, PR, HER2, PD-L1, AR, Ki-67) or as lineage/classification markers frequently used in differential diagnosis (CK7, CK20, CDX2). While variability in conventional IHC can lead to inconsistent treatment decisions, limited comparisons between studies, and conflicting results, bulk RNA-seq provides a high-throughput alternative for quantitative measurement of gene expression. We established single-gene classifiers using mRNA expression cut-offs selected by the highest F1 scores for ESR1, PGR, AR, MKI67, ERBB2, CD274, CDX20, KRT7, and KRT20 that were concordant with the respective IHC thresholds in the internal and validation cohorts. Our findings show the potential of RNA-seq in standardizing biomarker evaluation and the generalizability of using RNA-seq data to predict low/negative and high/positive biomarker expression.

However, some biomarkers did not exhibit high correlation between their RNA and protein expression levels. While steroid hormone receptors and Ki-67 demonstrated robust correlations, PD-L1 exhibited only a moderate correlation, and some genes in specific tumor groups, such as CDX2 in gastrointestinal cancers, displayed no significant correlation. This variability could be attributed to the influence of the TME that is typically avoided in IHC score assessments but can significantly impact total RNA expression in biopsies. We found that RNA expression was dependent on the TME and tumor purity in some cases; for instance, tumor purity significantly affected the expression of CD274, and its expression was considerably high in immune cells. Neoplastic cellularity and heterogeneity as well as tissue complexity underscored by the qualitative and quantitative variances in transcriptomes and proteomes among different TME cell populations, present a distinctive challenge in leveraging RNA-seq for genome-scale transcript quantification. Our observations highlight the need for advanced analytical methods like RNAscope^10,58 or new bioinformatics algorithms to separate tumor signals from those of the surrounding microenvironment in RNA-seq data, a challenge compounded by tumor heterogeneity that may affect the predictive accuracy of RNA expression levels.

Furthermore, discrepancies between RNA-seq-based classifications and IHC scoring may result from differences in sampling, tumor heterogeneity, post-translational protein modification, and antibody clone selection and staining in IHC^5,59. Namely, alternative splicing contributes to proteomic complexity through the generation of isoform proteins that are not detectable by IHC antibodies⁶⁰.

Our previous findings^61,62 were in line with other studies, demonstrating moderate to strong correlations between RNA expression levels and IHC staining results^10,63–66. Kang et al.⁶⁷ found a strong correlation between CD274 mRNA expression levels measured by a quantitative reverse transcription-polymerase chain reaction and PD-L1 IHC staining in gastric and urothelial carcinoma samples. Similarly, Sorokin et al.⁷ reported strong and statistically significant correlations between RNA-seq and IHC measurements for several biomarker genes in breast and lung cancer. Conroy et al.⁶⁸ also demonstrated that CD274 mRNA expression measured by RNA-seq was comparable to PD-L1 expression assessed by IHC, both analytically and clinically, in predicting immune checkpoint inhibitor response.

The delineation of the IHC HER2 1+/2+ cut-off (low expression) from negativity (0) and 3+ emerges as a critical clinical tool, particularly in the context of HER2-targeted therapies. However, challenges arose in distinguishing HER2-low cases (1+, 2+ only); the threshold range of > 6 < 8.5 log₂(TPM + 1) correctly identified only 103/146 cases with score 1+, 2 + and 68/84 cases with score 0, 3+, achieving a sensitivity of 87% and specificity 61%. By adjusting the lower threshold to > 5 log₂(TPM + 1), we were able to increase the sensitivity to 98% (Supplementary Table 4) and reduce the number of false negatives to 2. Conversely, for confirmatory testing where specificity is paramount, increasing the lower threshold to > 7 log₂(TPM + 1) resulted in a greater specificity of 95%, reducing the number of false positives to 5. These adjustments emphasizes that the one-size-fits-all approach of optimizing for F1 does not necessarily result in optimal performance for both screening and confirmatory tests. However, our additional threshold adjustment experiments demonstrate the feasibility of tailoring this approach during test design and development to optimize RNA-seq performance for specific settings. These are key considerations illustrating the flexibility of RNA-seq applications, particularly when considering implementation in clinical practice.

The discrepancies in IHC scoring, especially between HER2 0 versus HER2 1+/2+, emphasize the utility of RNA-seq in standardizing evaluations for this subtle yet clinically significant classification^17,64. There is clinical evidence that ERBB2 mRNA levels correlate with response to TDM-1⁶⁹ and T-DXd¹⁰ HER2-targeted drug conjugates. We observed a strong correlation between gene copy number, mRNA expression, and protein expression and established RNA cut-offs for negative, HER2-low, and HER2-high expressions that are consistent with recent genomic and IHC findings⁷⁰. However, de Oliveira et al.⁷¹ showed that IHC HER2 equivocal cases in gastroesophageal cancer could be a potential pitfall for evaluation with ERBB2 mRNA expression. The same findings for equivocal cases were demonstrated by Carretero-Barrio et al. in breast carcinomas⁷². By defining the specific clinical application—screening versus confirmation—we can better tailor RNA-seq threshold optimization to enhance diagnostic accuracy and utility.

Indeed, these findings highlight the potential of bulk RNA-seq as a versatile, dependable tool in personalized medicine, supporting clinicians in treatment decisions and assisting payers in optimizing resource utilization. In providing high-throughput capabilities for reliable quantitative measurement of gene expression, RNA-seq may overcome the deficiencies of conventional IHC, whose variability can lead to inconsistent and suboptimal treatment decisions. It is important to note that variability in RNA-seq results can arise from differences in normalization methods or pipeline settings, potentially affecting absolute expression values and correlation metrics. Similarly, although digital pathology can improve reproducibility in IHC assessment, it still faces limitations related to staining quality, algorithm calibration, and threshold selection⁷³. Both clinicians and payers would appreciate the potential of RNA-seq to standardize and improve biomarker evaluations, as better diagnostic approaches may lead to the improved outcomes desired by both clinicians and payers.

The retrospective design and small sample size of our tumor cohorts could limit the generalizability of our results. The number of evaluated cases per biomarker varied, reflecting differences in clinical testing frequency and archival availability. We recognize that the predominance of breast cancer cases in our dataset may influence threshold generalizability. A particular challenge was the lack of directly comparable Ki-67 data, leading us to rely on external datasets for subtype analysis in breast cancer, notably subtypes Luminal A and Luminal B. The absence of matched diagnostic IHC and RNA-seq data for certain biomarkers compelled us to confirm our findings using TCGA cohorts, making some of our gene expression comparisons more inferential. We note that RNA-based thresholds for PD-L1 (CD274) were calibrated against IHC assays using the 22C3 clone and CAL10 LDT antibodies. If a different PD-L1 antibody clone, such as SP142 or 28 − 8, had been used as the reference, the observed concordance and optimal RNA-seq cut-off could differ. This highlights the importance of harmonizing RNA thresholds with the specific IHC assays used in each clinical or research context. Furthermore, the RNA cut-offs we established, although insightful, are provisional and subject to refinement. Specific RNA-seq cut-offs require further validation in clinical trials.

Brueffer et al.⁸ showed that the classification error rates for RNA seq-based classifiers for key breast cancer biomarkers were generally equivalent to conventional histopathology. RNA-seq can thus provide a reliable biomarker kit, amenable to standardization similarly to histopathology, while potentially overcoming interpretation bias and confounding variability in trials and clinical practice. We demonstrated that expression levels of tested IHC markers were reliably reproduced by RNA-seq. This study serves as a proof of concept that RNA-seq could be used to screen new biomarkers for protein target discovery, supporting oncologists and pathologists in actualizing personalized therapy for cancer patients, aiding manufacturers of IHC tests, and enhancing the rigor of clinical trials⁷³.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

V.K., A.Bag, E.P., F.M-B., and D.S. performed study concept and design; V.K., D.S., S.D., K.K, O.K., and L.B. performed development of methodology and writing, review and revision of the paper; S.D., A.T., A.Sh. K.K., B.U.Sh., S.M., A.But., and A.Bel. provided acquisition, analysis and interpretation of data, and statistical analysis; A.Bag. provided technical and material support. All authors read and approved the final paper.

Funding

This research was funded by BostonGene Corporation.

Data availability

All data generated or analyzed during this study are included in this published article and its supplementary information files. The internal dataset used and analyzed within the current study is included in Supplementary Table 1. The public datasets analyzed in this study are enumerated in Methods section. Data from the internal cohort have been deposited in NCBI’s Gene Expression Omnibus. They are accessible through GEO Series accession number GSE293591 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE293591) and available from the corresponding author on reasonable request.

Declarations

Competing interests

This research was funded by BostonGene Corporation. BostonGene authors were employees thereof at the time the study was performed. Authors declare no other competing financial interests in relation to the work described.