Automated image quantification of immunohistochemistry-stained immune cells in triple-negative breast cancer

doi:10.21203/rs.3.rs-3138809/v1

Download PDF

Research Article

Automated image quantification of immunohistochemistry-stained immune cells in triple-negative breast cancer

https://doi.org/10.21203/rs.3.rs-3138809/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background:

Breast cancer is a molecularly heterogenous disease for which the composition of the tumour microenvironment (TME) is acknowledged with an increasing role in treatment response and prognosis. In triple-negative breast cancer (TNBC) tumour infiltrating lymphocytes (TILs), representative of a general immune response, have been associated with a favourable prognosis. With growing number of TME cell type markers being analysed by conventional IHC or other in situ methods combined with need of spatial marker relationship analysis digital image analysis tools are needed to facilitate broader in situ characterisation of the breast cancer TME.

Methods:

A TMA comprising 218 patients with TNBC, enrolled in the Sweden Cancerome Analysis Network – Breast (SCAN-B) study, with complementary clinicopathological, WGS, and RNA-sequencing data were used. The TMA was stained using immunohistochemistry for p53, CD3, CD4, CD8, CD20, CD68, FOXP3, and PD-L1 (SP142 antibody), with available pathology scoring for CD20, PD-L1 and TILs. An open-source digital image analysis pipeline, Tissue microarray MArker Quantification (TMArQ), for analyses of single marker IHC images was developed implementing starDist segmentation. Primary pipeline output was the number of positive cells based on IHC staining.

Results:

TMArQ’s cell counts for analysed immune markers were on par with results from more advanced trained machine learning algorithms and consistent with both estimates from human pathology review, different quantifications/classifications derived from RNA-sequencing as well as known prognostic patterns of immune response in TNBC. When combined with somatic genetic information (TP53-mutation and homologous recombination deficiency, HRD) the pipeline demonstrated consistency in p53 protein expression versus TP53 variant type and superior patient outcome for the combination of high CD3 counts with HRD-positivity in patients with adjuvant standard-of-care chemotherapy.

Conclusions:

TMArQ is an easy-to-use open-source automated pipeline for IHC-based cell detection and quantification to be used as an exploratory tool in cancer image analysis. Digital analysis tools will likely greatly facilitate further characterisation of the breast cancer TME in novel ways and allow for a more precise linking of TME features and molecular alterations detected by large-scale omics methods, thereby deepening our understanding of breast cancer.

Breast cancer

Triple-negative breast cancer

Tumour microenvironment

Immune markers

Digital Image Analysis

IHC

Homologous recombination deficiency

patient outcome

Breast cancer is the most common malignancy in women worldwide (1). At the molecular level, breast cancer is a heterogenous disease, a fact ultimately reflected in disparate patient outcomes that are dependent on a complex interplay between biological characteristics of both the patient and the tumour. To guide therapy for patients in both the primary and advanced setting different biomarkers have been suggested over time, including more recently also gene expression-based assays. In current clinical management of breast cancer three treatment predictive markers however still represent the cornerstones for disease stratification, the oestrogen-receptor (ER), the progesterone receptor (PR), and overexpression/gene amplification of the ERBB2/HER2 oncogene (HER2). The status of these markers is typically defined by immunohistochemistry (IHC) on formalin fixed paraffin embedded (FFPE) tissue sections, complemented with in situ hybridization methods for ERBB2/HER2 gene amplification. Together, these three markers define four important clinical subgroups of breast cancer associated with different treatment options: I) ER-positive and HER2-negative tumours, II) ER-positive and HER2-positive tumours, III) ER-negative and HER2-positive tumours, and IV) triple-negative tumours (TNBC) that are negative for all three markers (2).

Historically, patients with TNBC have often had a poor prognosis due to aggressive disease without available targeted therapies, leaving chemotherapy as the systemic therapy and radiotherapy for local disease control as the main treatment options. The clinical subgroup of TNBC today constitute around 10% of all breast cancer, and have been associated with younger age at diagnosis, early relapses, highly rearranged genomes, and a strong association to germline alterations in high penetrant breast cancer predisposition genes like BRCA1 and BRCA2. The latter links TNBC with DNA repair deficiency, typically through homologous recombination deficiency (HRD) based on BRCA1 and BRCA2’s role in the repair of double-strand breaks. HRD is likely a key driver of tumorigenesis in early-stage TNBC, predisposing for the formation of complex tumour genomes, but has also been linked to improved prognosis after adjuvant chemotherapy (3, 4). Besides HRD, there is also a well-established prognostic association between a general immune response and improved patient outcomes in early-stage TNBC patients treated with and without (neo)adjuvant chemotherapy (5–7). Here, the infiltration of different immune cell types in the tumour microenvironment (TME) may represent the systemic anticancer immune response towards tumour characteristics that are yet not fully understood. The composition and density of different immune cell populations in the TME likely influence tumour progression and success of anti-cancer therapies profoundly, although our understanding of these processes is still far from complete. A general immune response in a tumour tissue sample may be estimated in several ways. Among these are morphological assessment of the level of stromal tumour infiltrating lymphocytes (sTILs) on H&E-slides (5), IHC-staining of cell type specific markers, immune cell type directed flow cytometry, or genomic methods such as mRNA expression analyses (8). Compared to TNBC, the other clinically defined breast cancer groups (excluding ERBB2/HER2-amplified disease) are often considered to be less immunogenic (cold), and for these groups, early clinical trials using immune checkpoint inhibitor therapy have not demonstrated a significant therapeutic effect (9). Still, within TNBC there is a large heterogeneity in the level of immune response observed across individual tumours (10, 11). This immune heterogeneity has also been recognized in proposed transcriptional TNBC subtypes though an immunomodulatory (IM) subtype/phenotype (12). For instance, in 2011 Lehmann et al. proposed the IM subtype as one of six intrinsic TNBC subtypes, but later refined this subtyping scheme to comprise four subtypes (BL1, BL2, M, LAR) where the IM subtype can be superimposed on the former as an independent feature of the tumour (13).

To better understand interactions between somatic tumour alterations and the TME, improvements in methods for the characterization of the TME cell composition and cell type interactions are needed. Ideally, this involves both a quantitative identification of cell types using e.g., cell type specific markers, but also the ability to target the interrelationship between different markers expressed both in malignant and non-malignant cells based on their spatial location. Technology advancements have made it possible to analyse multiple markers simultaneously in a single FFPE section, but also through flow cytometry, mass spectrometry, or even more complex spatial profiling methods based on nucleic acid or protein detection (8, 14). While these technologies provide improved resolution and multiplex capability, they are often costly, can have specific tissue requirements, and have a limited throughput. Thus, in routine diagnostics of breast cancer and research studies, single marker (single-plex) IHC analysis on FFPE whole sections or tissue microarray (TMA) sections still represent the mainstay option. As a consequence of this, large amounts of archival slides and scanned image data are available for retrospective analysis. In contrast, many cutting-edge technologies rely on fresh sectioning, and analysis pipelines built around these are often not amenable to archival image data.

While throughput of both single-plex and multiplex methods are increasing, data analysis remains a bottleneck, especially if trained human grading/scoring is required. In routine practice IHC stains are often manually scored by a trained pathologist or researcher. Despite standardization efforts this often requires subjective assessments to be made which in turn introduce bias. The existence of this type of subjective bias has been repeatedly demonstrated in breast cancer for the commonly assessed clinical markers (15, 16). To circumvent this limitation many digital image analysis tools have been developed, including tools that include deep-learning for different classification/scoring applications such as TIL-estimation (17, 18) and PD-L1 scoring (19). Still, challenges remain to be solved for digital image analysis methods, including optimal segmentation of individual cells in processed images to allow correct digital quantification, and transferability of trained deep learning (DL) models to other IHC data. For the former, advancements in convolutional neural networks have improved the extraction of cellular metrics that are essential for definition of the spatial tissue architecture, like the shape, size, position, and count of individual cells (20).

In the current study we aimed to develop a minimally supervised open-source digital image analysis pipeline that can facilitate more quantitative and reproducible analyses of single marker IHC images in a high-throughput and spatially resolved manner. Our method also aims to facilitate analyses like those that can be performed on state-of-the-art platforms such as the PhenoImager instrument (Akoya Biosciences) on single-marker IHC-stains through in-silico overlaying of multiple IHC stains from separate sections. We apply our method to a large TMA data set comprising over 200 TNBCs from a molecularly very well profiled, population representative, TNBC cohort of patients from South Sweden diagnosed 2010–2015 (3) that were stained for multiple immunological TME markers. Through comparisons with matched genomic data as well as pathologist and DL scoring, we demonstrate that our pipeline generates a deterministic quantification of the immunological microenvironment in TNBC. We show it is on par with gold standard and more advanced methods and is also capable of recapitulating known prognostic associations of the immune TME, which can in turn be further improved through combination with e.g., DNA repair deficiency status. These results demonstrate the value and potential of our method for enhancing our understanding of the interplay between genotype and phenotype in TNBC.

Patient cohort, clinical data, and ethics approval

All included patients with primary TNBC were enrolled in the Sweden Cancerome Analysis Network – Breast (SCAN-B) study (ClinicalTrials.gov ID NCT02306096) (21, 22)), approved by the Regional Ethical Review Board in Lund, Sweden (Registration numbers 2009/658, 2010/383, 2012/58, 2016/742, 2018/267 and 2019/01252) as previously described (3). All patients provided written informed consent prior to enrolment. All analyses were performed following local and international regulations for research ethics in human subject research. Clinicopathological data were obtained from Staaf et al. (3), with updated patient outcome data for invasive disease-free survival (IDFS) of 6.51 years for censored patients (no event).

TMA construction, IHC staining and TIL estimates

For 218 patients from the cohort reported by Staaf et al. (3) a TMA (five blocks) was constructed with two 1mm cores / tumour as previously described (11). The TMA was stained for p53 and a set of immune markers: CD3, CD4, CD8, CD20, CD68, FOXP3, and PD-L1 (SP142 antibody, Roche). TIL scoring on whole section H&E-slides, CD20 IHC staining, and PD-L1 IHC staining and associated pathology scorings have been described previously (11). Pathology scorings for CD20 and PD-L1 were performed on the same stainings used for the automated pipeline, while the TIL scoring was performed on a whole slide section taken from the same FFPE tissue block prior to TMA core sampling. For PD-L1, 1% staining in tumour-infiltrating immune cells (termed IC) was used as cut-off for PD-L1 positivity in line with guidelines for the SP142 antibody. Remaining IHC stainings were performed as outlined in Supplementary Table 1.

RNA-sequencing and whole genome sequencing data

RNA-sequencing data, including PAM50 molecular subtypes by nearest centroid correlation, for included cases was obtained from Staaf et al. (23). Expression values were summarized to transcripts per million (TPM) estimates for this study. Entrez ID from the Gencode27 metadata was used as gene identifiers. CIBERSORT (24) immune cell deconvolution scores and Lehmann TNBCtype mRNA subtypes (BL1, BL2, M, and LAR) (13) were obtained from the study by Aine et al. (11) for matching samples. The Lehmann IM class (negative/positive) was defined using a cut-off in correlation of 0.17 to the original IM-centroid correlation. Matched whole genome sequencing (WGS) data for analysed cases were obtained from Staaf et al. (3) and included in addition to somatic alterations also mutational signature scores, and HRD predictions by the HRDetect algorithm (4).

Quantification of immune antibodies in IHC

We analysed single-plex IHC stains for p53 and a set of 7 immune cell antibodies using our pipeline TMArQ, as illustrated in Fig. 1A. The pipeline allows users to specify which images should be analysed via a configuration file. Input data currently consists of png images in RGB format of individual cores extracted using standard dearraying methods. Briefly, in the pipeline TMArQ detects the core area using Hough circle detection as implemented in scikit-image version 0.19.2 (25). The detected diameter and centre of the core are recorded in a text file, which is used to centre the core in the image for downstream analysis. If no circle is detected the image is considered empty or the tissue of too poor quality (containing heavy folding or large gaps e.g.) and is flagged, allowing it to be excluded from further analysis.

Next, hematoxylin and DAB stains are separated using colour deconvolution, as described in (26) and implemented in scikit-image (25). For cell detection the pipeline uses starDist segmentation version 0.8.1 for general cell detection in the hematoxylin layer (27, 28). TMArQ then determines the presence of DAB stains by thresholding the DAB layer using triangle thresholding on an image-by-image basis to reduce the effects of non-uniform staining across cores and TMAs. It then combines the coordinates of the cells detected by starDist and the thresholded DAB-layer for positive/negative cell detection. TMArQ places every cell at the centre of a 16x16 pixel grid. If the grid in the thresholded layer contains more than 10% positive pixels we consider the cell to be positive for the specific marker. This cut-off is based on the CD3 marker (Supplementary Fig. 1), showing that cut-off variations between 5–25% inferred little correlation differences compared to CD3 TPM expression values from RNA-sequencing or pathology estimated TIL percentages. For further detailed information on the pipeline, please refer to https://github.com/StaafLab/TMArQ, where TMArQ is publicly available. The automated TMArQ IHC-based cell counts used in this study are available in Additional file 2 (Supplementary Table 2).

Integration of IHC quantifications with additional data

All statistical analysis was carried out in R, version 4.0. Automated IHC-based cell counts were log2 transformed before analyses. We calculated correlations between automated cell counts and RNA-sequencing data using the Pearson correlation. Correlations between estimated TIL percentages and cell counts were computed using the Spearman correlation. Survival analyses were performed in R using the survival and survminer packages with IDFS as the clinical endpoint for patients treated with adjuvant chemotherapy. Survival curves were estimated using the Kaplan-Meier method and compared using the log-rank test. Hazard ratios were calculated through univariable Cox regression using the coxph R function.

Comparison of automated TMArQ pipeline cell counts to available pathology scores

TMArQ is an automated digital analysis pipeline based on open-source software as outlined in Fig. 1. To assess the pipeline’s performance and validity we applied it for automated IHC scoring of p53 and six immune cell antibodies as well as PD-L1 (SP142) in a large TNBC TMA cohort (n = 218) described in Table 1. Following the application of the pipeline, we first compared the extracted log2 transformed cell counts to different marker score estimates made by breast cancer pathologists as outlined in our previous study (11). This included scores for PD-L1 and CD20 (B-cell marker) on TMA core level (same stainings as for the automated pipeline), and for TIL percentages on whole-slide H&E-stained tumour sections from the same FFPE tissue blocks that were used for TMA construction. Supplementary Fig. 2A-G shows cell counts per core versus TMA blocks.

Table 1

Sample and patient characteristics for the SCAN-B study cohort.
	Samples	Unique samples
TMA marker
CD3	436	218
CD4	436	218
CD8	436	218
CD20	436	218
CD68	436	218
PDL1 (sp142)	436	218
FOXP3	436	218
p53	436	218
TMA block
1	104	52
2	96	48
3	96	48
4	50	25
5	90	45
Grade
2	50	25
3	376	188
NA	10	5
HRD status
high	256	128
low/inter	176	88
NA	4	2
TNBCtype (Lehmann-4)
BL1	158	79
BL2	88	44
LAR	78	39
M	104	52
NA	8	4
Lymph node status
negative	270	135
positive	162	81
NA	4	2
PAM50 subtype
Basal	344	172
Her2	64	32
LumA	8	4
LumB	2	1
Normal	12	6
unclassified	2	1
NA	4	2

To investigate whether our automated pipeline scores could be used to capture a clinically relevant antibody cut-off, we compared automated PD-L1 cell counts using the Roche SP142 antibody to the established clinical cut-off of 1% IC staining for the antibody. Figure 2A shows the number of PD-L1 stained cells from the pipeline for individual cores versus the binned pathology groups of PD-L1-low (< 1% IC) or PD-L1-high (≥ 1% IC) as outlined in (29). The analysis shows a clear separation, demonstrating that the pipeline counts are consistent with pathologist scoring. However, we also note a degree of overlap between the PD-L1 low positive and negative distributions. We also calculated the correlation of the PD-L1 pipeline cell counts to PD-L1 percentage scores by a pathologist (Fig. 2B). Here we also observed a trend of increasing pipeline counts corresponding to higher PD-L1 pathology scores with a Spearman correlation of 0.58 (p < 0.01).

To evaluate our automated pipeline against human-derived scores on a semi-quantitative (ordinal) scale, we compared the automated pipeline counts to available pathology scores for the B-cell marker CD20. For this analysis CD20 antibody presence was previously scored using a four-level scale by a pathologist, where 0 corresponds to no/lowest CD20 staining (11). Figure 2C shows that higher pipeline cell counts correspond to increasing pathology scores. In particular, the increase in corresponding pipeline counts when moving from a score of 2 to 3 is notable.

Finally, we compared pathology-estimated TIL proportions (%) to our automated CD3 (a pan-T-cell marker) cell counts. For every tumour we compared the mean CD3 count of the two cores to the estimated TIL percentage. We found a correlation of 0.58 (p < 0.05) between an increased CD3 + cell count in the TMA cores and an increased TIL percentage from the whole slide analysis (Fig. 2D). Taken together, these comparisons support the notion that the output from of our digital pipeline is in good agreement with independently derived estimates by a human evaluator.

Comparison of TMArQ automated pipeline results to matched bulk tissue RNA-sequencing data

To investigate how the automated cell counts matched gene expression data, we compared our pipeline’s output to I) matched bulk tissue RNA-sequencing data for analysed marker genes (TPM-values), II) in silico estimated cell-type proportions by CIBERSORT from RNA-sequencing data, and III) proposed gene expression subtypes in TNBC, including an immunomodulatory (IM) class derived from mRNA expression. Figure 3A shows the automated cell counts from the TMA core versus corresponding RNA-sequencing TPM values for the analysed immune markers. The strongest correlations were observed for the CD8 and CD3 markers with Pearson correlations of 0.59 and 0.51 (p < 0.05) respectively. This indicates that cell counts, and bulk gene expression values are in agreement with each other. Regarding potential sources of confounding, we note that the RNA-sequencing and TMA data was generated from different pieces of the same tumour sample and are therefore sampled several millimetres apart.

Next, we compared the automated cell counts to the abundance of in silico estimated cell types (B-cells, T-cells, macrophages) by the CIBERSORT deconvolution method applied to the bulk RNA-sequencing data. Here, Fig. 3B shows a positive Spearman correlation of only approximately 0.3 between the fraction in silico CIBERSORT cell type fractions and matched automated in situ cell counts, a finding that extended to all analysed IHC markers as shown in the correlation heatmap in Fig. 3C. Similarly, we also observed low correlations for all markers when using the ratio of stained cells / total number of cells per core versus CIBERSORT estimates (Supplementary Fig. 3). Notably, when merging the CIBERSORT cell type fractions into a combined sum, the correlation increased substantially compared to all in situ cell counts as exemplified for CD8 in Fig. 3B and for all markers in Fig. 3C. To further investigate this finding, we compared CIBERSORT B-cell, T-cell, and macrophage estimates versus the same RNA-sequencing data as for the cell type marker genes shown in Fig. 3A. Here, we observe that Spearman correlations for the RNA-sequencing TPM estimates versus individual CIBERSORT estimates were at best in the same range as for antibody CD8 and CD3 cell counts versus CIBERSORT estimates (i.e., around 0.6 in Spearman correlation, Supplementary Fig. 3). Moreover, correlation estimates for individual genes (TPM) increased notably when compared to a summarized CIBERSORT cell type fraction (up to Spearman correlation 0.8, Supplementary Fig. 3), which is in line with the findings for the cell counts (shown in Fig. 3C).

Finally, we compared automated CD3 (a pan lymphocyte marker) cell counts with the Lehmann immunomodulatory (IM) classification and tumour classification by the four updated Lehmann TNBC subtypes (TNBCtype-4: BL1, BL2, M, and LAR) (13). The IM-positive class showed overall the highest abundance of CD3 + cells when analysed as a separate entity (Wilcoxon’s test p = 3.7e-9) and within the four Lehmann subtypes when used as a substratification (Fig. 3D). For the latter, we make several interesting observations. Firstly, the BL1 subtype appears to have similar CD3 + counts irrespective of IM-class, while the BL2 class is a mix of potentially immune hot and cold tumours. In contrast, the M (Mesenchymal) subtype appears to have an immune cold phenotype, with no IM-positive cases and correspondingly low automated CD3 + counts. For the LAR subtype, which has been proposed as a more immune cold phenotype compared to BL1 and BL2 (29, 30) IM-positive cases were associated with significantly higher automated CD3 + counts than IM-negative samples. These findings suggest that lymphocyte presence as measured by our pipeline and immune classification of TNBC by mRNA expression subtypes are in agreement with each other. Importantly though, our in situ analysis suggests clear differences in immune cell abundance depending on the underlying mRNA-based TNBCtype subtypes.

Antibody to antibody comparisons and joint antibody expression patterns

To further analyse the generated output from TMArQ we assessed core-to-core variability for markers, marker-to-marker variability, and combined patterns of marker expression versus clinicopathological variables and molecular classifications based on matched RNA-sequencing and WGS.

The TMA used in this study contained two cores for every sample taken from the same whole slide, with each core targeting different tumour rich areas. Figure 4A illustrates analysis of core-to-core variability for the CD8 and PD-L1 (SP142) antibodies, while Fig. 4B summarises all markers. Overall, core-to-core Spearman correlations ranged between 0.48–0.66, with lower values for CD20 and FOXP3. Next, we analysed the correlation of automated cell counts between different markers within the same core (stack). Here, we observed a comparably higher Spearman correlation between different antibodies within the same core. Figure 4C illustrates the correlation of CD8 to CD4 and PD-L1 cell counts within the same core, while Fig. 4D summarizes this for all marker combinations.

Next, we applied hierarchical clustering to average cell counts to explore between-sample and marker associations. Figure 4E shows the resulting two main clusters. Here we observed a general co-expression of immune markers in samples, consistent with the high within-core correlations between antibodies (Fig. 4D). Moreover, we observe a higher proportion of lymph node-positive tumours and mRNA-classified IM-positive tumours in the sample cluster with generally higher cell counts (Fig. 4E, left cluster) (Fisher’s exact test p = 0.02 for lymph node status, and p = 7e-8 for IM-status). For genetic tumour characteristics such as TP53 mutation status, HRD-status, BRCA1-deficiency due to biallelic inactivation by mutations (somatic/germline and LOH) or promoter hypermethylation it is evident in Fig. 4E that substantial proportions of tumours are found in both immune groups. For HRD-positive cases (HRD-high), proportions between heatmap clusters in Fig. 4E were 54% and 46% respectively, despite a similar tumour mutational burden between groups (Wilcoxon’s test p = 0.21, Supplementary Fig. 4A). Thus, while HRD-positive TNBCs have a generally higher in situ cell count than HRD-negative tumours, considerable heterogeneity in immune cell infiltration levels exists within the HRD-positive tumour group (Supplementary Fig. 4B).

In silico merged antigen expression patterns and non-immune cell quantification

The first step of TMArQ for analysing TMA cores involves circle detection and coordinate normalization. This step enables us to place all subsequently detected cell objects on a common coordinate grid, which allowed us to test the concept of creating in silico merged multiplexed images of separate IHC stains for our single-plex markers. For studying tumour-immune interactions, it is essential to distinguish cancer cells from normal epithelial cells and immune cells. In addition to using epithelial cell markers, somatic alterations may infer protein expression that can be used to identify malignant cells. An example of the latter is mutational inactivation of the tumour suppressor gene TP53, a frequent event in TNBC, which can confer protein overexpression in mutated cells (31, 32). The type of mutation can conceivably influence the expression pattern. As mutational inactivation occurs only in the malignant cell, p53 protein expression could be used to identify and distinguish malignant cells across the core surface. We stained our TMAs for p53 and quantified positive cells using our automated pipeline for comparison versus TP53 mutational status determined by WGS. Figure 5A (left panel) shows an example for a TP53 mutated tumour with high p53 antibody staining expression displaying a clear separation between the malignant and non-malignant cells. TP53 mutated tumours indeed had higher stained cell counts across the cohort (Fig. 5B), although protein expression in the TP53 mutated subset was highly heterogenous compared to the TP53 wildtype tumours. This heterogeneity may be due to not all TP53 mutations resulting in high p53 protein staining, as evident for another TP53 mutated tumour shown in Fig. 5A (right panel). Stratification by TP53 mutation consequence and affected p53 protein domain illustrated how the in situ p53 cell counts varied depending on the variants’ consequences and positions (Fig. 5C). For instance, TP53 deletions and TP53 nonsense and frameshift mutations result in a typically lower p53 IHC-based cell count as these variant types would typically not generate an expressed protein. Notably, p53 cell counts were especially low for tumours with nonsense mutations within the DNA binding domain (DBD) of the protein (Fig. 5C).

Given that epithelial cells can be demarked by either pan cytokeratin stainings, or somatic alterations like TP53 when present, this may allow for exploration of tumour-immune interactions using image overlays. We tested this by overlaying immune cell markers with p53 stains as a proof of concept of an in silico merged multi-marker application of the pipeline. As illustrated in Fig. 5D-E, this approach may enable the analysis of spatial interactions between markers of different cell types.

Automated TMArQ cell counts are associated with differences in patient outcome in TNBC

Immune response in the TME, as measured by for example TILs or mRNA expression of immune-related genes has repeatedly been associated with improved outcome in early-stage TNBC patients treated with and without (neo)adjuvant chemotherapy (5–7, 34). Considering the correlation of the automated cell counts with different pathology scores and RNA-sequencing estimates we analysed whether the automated cell counts also carried prognostic information for TNBC patients in our cohort treated with adjuvant chemotherapy. As seen in Fig. 6A for the CD3 marker, dividing the automated CD3 cell counts into two groups (low/high) based on the median observed cell count was associated with a difference in IDFS, with the CD3-high group showing improved patient outcome. To expand this analysis, we performed univariate Cox regression analyses for all seven immune markers (Fig. 6B). In agreement with the overall aspect of an immune response being associated with improved outcome in TNBC, we observed hazard ratios < 1 for the group of high versus low cell counts for all markers, with statistically significant results for the CD3, CD4, CD8, and PD-L1 markers, but not FOXP3, CD68, and CD20. In comparison, univariate Cox regression for pathologist-estimated whole slide TILs (with TILs as continuous values) showed a hazard ratio of 0.963 with a 95% confidence interval of 0.941–0.985 with a p-value of p = 0.001. Moreover, a median split of TIL scores (binary low/high groups using all cases) was significantly associated with IDFS in chemotherapy-treated patients with the high group having superior IDFS (log-rank p = 0.0001). Together, this demonstrates that output from the automatic pipeline is consistent with a general prognostic theme of immune response in TNBC and suggests that the signal of a clinically relevant immunological tumour response may be captured in a small sample such as a 1mm TMA core.

In TNBC, HRD-status has also been proposed as a prognostic factor in addition to immune response. Based on the marker co-expression analysis shown in Fig. 4E, which revealed that HRD-positive tumours (by WGS) were present in both the immune-high and immune-low clusters, we stratified HRD-status with the CD3-low/high cell count and performed a 4-tier classification. We then evaluated the association between this stratification and IDFS in adjuvant chemotherapy-treated patients (Fig. 6C). Notably, an excellent outcome was observed for the HRD+/CD3 high combination, while HRD+/CD3 low cases had a similar (worse) prognosis as HRD-/CD3 high cases. The poorest outcome was observed for HRD-/CD3 low cases, with a long-term IDFS of approximately 50%. Together, this analysis demonstrates how quantitative measures of cell abundance can be combined with other molecular or clinicopathological variables to explore novel associations and identify subsets of patients with different outcomes following therapy.

In the current study we developed an automated digital analysis pipeline (TMArQ) for IHC-based cell count scoring to investigate the TME in TNBC using image data from single-plex IHC stains. Using automated cell counts on individual TMA cores facilitates a more reproducible, quantitative, and accessible way to generate and test hypotheses and perform integrative multi-omics analyses. Although not the principal aim of this study, we believe our approach should be amenable to other cancer types as well and we have therefore made the underlying code available as a publicly accessible stand-alone pipeline.

We base our analyses on multiple immune marker stainings from a TMA constructed from a molecularly very well-characterised cohort of population-representative early-stage TNBC from South Sweden 2010–2015 (3, 11). While this cohort allows for in-depth comparisons versus molecular data from both RNA-sequencing and WGS, we note that the cores in the used TMA were aimed to target tumour-rich areas in the corresponding FFPE tumour tissue blocks. As such, the TMA cores may not necessarily capture the full immune phenotype of the TME in their respective tumour. This could in part explain some of the observed differences between pipeline and pathology estimates for TILs. Moreover, for each tumour case in the TMA there were two 1mm cores and the averages of these have been used in most analyses. As demonstrated in Fig. 4, core-to-core variability exists and may in part account for discrepancies compared with estimates obtained using other methods. Interestingly, we observed a general trend towards higher marker-to-marker correlation within a core for the analysed immune markers (Fig. 4). A general trend towards higher co-occurrence of all immune cell types within a given tumour was also observed (Fig. 4E). This is not surprising considering that many of the analysed immune cell types may act together in mounting a tumour immune response. Our combined analysis of different stains from the same core (stack) for a tumour emphasizes the value of a pre-planned staining order of the TMA to facilitate in silico merging of single-plex IHC stains, to avoid issues arising from stainings to be merged being several micrometres physically apart in a tissue core.

To evaluate the automated IHC-based cell counts we compared pipeline quantifications to several pathologist-derived estimates of the same stainings (CD20 and PD-L1 sp142) as well as TIL-estimates from whole slide H&E-stains taken before TMA construction. For the latter, we observed a clear correlation between automated cell counts for the CD3 pan lymphocyte marker and pathologist estimates of TILs. As a caveat we note that IHC staining was performed on cores from tumour-rich areas that may not capture the full range of the TME immune response. In comparison, for TIL-estimations, full sections are typically recommended over biopsies in guidelines (35), and TIL-estimates from whole slide evaluations have been reported to be higher than in matched TMA core analysis (36). Moreover, TIL scoring based on H&E-stained specimens does not consider the various origins and functions of immune cells (37), and like other pathology-estimated scores human TIL-scoring is associated with inter- and intra-observer variability (36–38). In support of the validity of our approach, a notable finding is that the correlation of our automated CD3 cell counts derived from TMA cores to TIL estimates generated on the corresponding whole-slide sections by an expert pathologist was close in strength to the correlation obtained when an AI-trained TIL predictor based on DL methodology was applied to the same whole section H&E-stains used by the pathologist (Spearman’s rho of 0.55 versus best AI predictor correlation of 0.63) in the study by Bai et al. (17). Moreover, based on data from the same study by Bai et al. (17) we note that for the same set of chemotherapy-treated patients as in Fig. 6A, our automated cell counts outperformed the AI-trained eTILs estimates in survival association. The latter was not significantly associated with IDFS when patients were split into low/high groups based on the median eTIL (log-rank p = 0.24, Supplementary Fig. 4C). This may indicate that simple estimates such as automated leukocyte counts on dual 1mm TMA cores can be used to probe relevant clinical correlations in a high-throughput fashion.

Estimation of PD-L1 status has become important in cancer therapy following the introduction of immune checkpoint inhibitors (ICIs) in patient care. A problematic issue in this context is that usage of an ICI is often associated with the requirement to use a specific PD-L1 antibody as an accompanying diagnostic assay to guide treatment (39). The SP142 PD-L1 antibody is the accompanying diagnostic assay for the PD-L1 inhibitor atezolizumab, which was first reported effective in combination with nab-paclitaxel in patients with unresectable, locally advanced, or metastatic triple-negative breast cancer in the phase 3 IMpassion130 trial (40). This trial invoked an accelerated FDA approval for atezolizumab that however was later retracted based on negative results from the IMpassion131 trial (41). Irrespectively, for pipeline assessment, the SP142 assay serves as an excellent example, with a positive staining call of PD-L1-positivity set as low as 1% staining of tumour-infiltrating immune cells (ICs), compared to the scoring of PD-L1 expressing tumour cells (TCs). Figures 2B-C demonstrate that the automated pipeline counts correspond well with pathology-derived binary classes of PD-L1-low/PD-L1-high, but also that there is a strong correlation to actual SP142 grading in percent. Here, the full dynamic range of the pipeline counts could enhance scoring and reproducibility at low expression levels compared to the challenging assessment of whether more or less than 1% staining of IC exists by the human eye. PD-L1 IC scoring is challenging for several reasons as discussed in (39) that together may explain discrepancies observed both for the discrete class analysis and the overall correlation analysis in this study. In the near future, it is conceivable that computer image analysis algorithms could function as an orthogonal method for PD-L1 IC scoring particularly for borderline and challenging cases (39). Finally, we also compared pipeline output to pathology scores for CD20, classified into four expression bins. Here we saw a good agreement between cell counts and score groups, but also a particularly sharp increase for the highest score bin (Fig. 2D). This sharp increase may indicate either a potential manual bias in the scoring of the highest CD20-expressing tumours, or a biological aspect of large numbers of infiltrating CD20-positive cells (e.g., as tertiary lymphoid structures) connected to some specific tumour feature. Regarding the latter hypothesis, the highest CD20 pathology group (group 3) contained 80% HRD-positive tumours, compared to 66% for group 2, 60% for group 1, and 44% for group 0. Whether this observation represents a direct consequence of a specifically rearranged genome, or merely a correlative association remains to be determined through deeper molecular analyses.

In addition to comparisons to matched pathology estimates we also compared the automated IHC-based cell counts to different outputs from matched bulk tissue RNA-sequencing. For the latter comparisons these are: I) made on different tumour pieces, and II) made on selected tumour rich areas for the IHC stainings versus RNA extracted from bulk tissue without specific regard for tumour cell content. Still, we observed strong correlations between automated cell counts and matched TPM expression for particularly CD3 and CD8 (Fig. 3A). Notably, for these markers, Locy et al. (42) also reported the highest agreement between TIL pathology scores and mRNA expression estimates based on Nanostring analysis. Overall, a good correlation between estimates by IHC or flow cytometry and bulk mRNA expression levels (in similar ranges that we observe) has been reported for certain cell types (including T-cells, cytotoxic cells, mast cells, and macrophages) in different malignancies (43). Bulk mRNA expression however cannot provide insights into the spatial distribution of cell types as in situ analyses can.

Several transcriptome-based cell-type quantification methods for immuno-oncology have to date been reported, of which CIBERSORT is one of the most commonly used (44). When CIBERSORT estimated cell type proportions were compared to automated cell counts we found an intermediate correlation for macrophages, B-cells, and T-cells between the methods for analysed antibodies (Figs. 3B-C). The correlation increased somewhat when CIBERSORT cell type fractions were summed into a more general immune cell score (Fig. 3C). Collectively, these findings imply that CIBERSORT deconvolution predominantly identifies a general trend of immune cells rather than specific cell subsets, except for outlier cases that score highly using any method. Notably, when comparing CIBERSORT B-cell proportions versus the CD20 pathology classifications we saw increasing fractions for CIBERSORT up until the highest score category, which had a lower average CIBERSORT B-cell fraction than the second highest group (Supplementary Fig. 3). This contrasts with the pattern seen for the cell counts (Fig. 2D), suggesting here that the in situ derived counts are in better agreement with manual scoring than the deconvolution method. Moreover, it should be noted that observed correlations between automated cell counts and bulk mRNA deconvolution cell estimates are likely in line with what might be expected in heterogeneous breast cancer tissue. This assertion is in line with observations using xCell (45) similar to CIBERSORT, which showed correlations between 0.5–0.7 when estimated cell proportions were compared to cell sorted blood proportions. This illustrates the difficulties in using in silico deconvolution methods for precise assessments of the TME in individual samples.

When we compared automated CD3 cell counts to a more general immune classification in TNBC, IM-class (13), we observed a consistent pattern overall with higher CD3 quantifications in IM + tumours. The IM classification is based on a general set of immune response associated genes considering the general pathway enrichments reported by Lehmann et al. (46). As such, the IM-signature may be less associated with the expression of a specific cell type, and more with a general immune cell load. While patients with LAR-classified tumours have been associated with both a generally immune cold TME (together with the TNBCtype M class (12)) and a generally poorer response to conventional systemic therapy (13, 47), it remains to be determined if the observation of a subset of TNBC LAR patients with high lymphocyte infiltration (reflected by automated CD3 cell counts, Fig. 3D) translates to a difference in patient outcome. Considering the high marker-to-marker cell count co-expression we observed for individual marker pairs (Figs. 4C-D) we also clustered pipeline cell counts for all analysed immune markers (Fig. 4E). The latter analysis revealed that most markers appeared correlated with regard to cell count levels, defining at one end an immune cold subset of TNBC tumours and at the other end an immune hot tumour subset, as reported previously (10, 12). Consistent with the established prognostic role of immune response in TNBC and the general immune marker-to-marker co-expression in TMA cores, patient stratification based on automated cell counts subdivided patients in our cohort with adjuvant systemic therapy into groups with differing prognosis in terms of IDFS (Fig. 6). This association was also observed for pathologist-derived TIL scores, but not for AI-estimated eTILs (17). As for PD-L1 scoring, the continuous counts produced by our automated pipeline could be useful for considering threshold selection for the definition of prognosis or risk groups in a discovery-validation cohort approach.

Another observation made in our cohort based on linking the cell count co-expression analysis with genomic data was that a highly rearranged tumour genome is not equivalent to an immune hot phenotype in situ. This is exemplified by WGS-classified HRD-positive tumours in our cohort, that bear highly rearranged genomes and a high mutational load (3). In our analyses, these tumours with similarly complex genomes display variable automated immune cell counts across all investigated immune markers (Fig. 4). Similar findings, albeit using more limited sequencing data, have also been noted by other studies (10). Therefore, one can infer a complex interplay between DNA repair deficiency status and the immune phenotype of the TME that may have considerable prognostic relevance. Indeed, a simple 4-tier grouping of HRD-status and dichotomized CD3 cell counts showed that patients with HRD-positive tumours with higher CD3 counts have a better IDFS after adjuvant chemotherapy compared to the other HRD/CD3 groups (Fig. 6). Moreover, we demonstrate that while p53 staining in general is associated with the presence of pathogenic TP53 mutations, this is not always the case, as highlighted by our integration of WGS-derived variant consequence and position data (Fig. 5). Together, these examples illustrate how somatic tumour features from other platforms can be combined with in situ TME estimates in breast cancer to add nuance to analyses and explore novel hypotheses.

The proposed pipeline has its limitations. We did not validate the pipeline for the use on individual patients. Curation of input images still requires a human inspection to avoid technical artefacts that may affect cores with tissue folding, considerable gaps, poor staining quality or other artefacts. However, the pipeline can be easily adapted to flag potential outliers and spot cores that appear different in quality. Moreover, the current cut-off in the pipeline is based on the CD3 marker and set to be a general-purpose cut-off. Thus, we acknowledge that it may not be the optimal cut-off for all individual markers. It should therefore, ideally, be individually examined by a user in new analyses. Regarding in silico merging of different stains we note that combining stains from a stack of sections (the z-axis of a core) may result in positional shifts when transitioning across different cell layers. To some extent, careful planning of the IHC staining order in single-plex studies may mitigate this issue. Future pipeline development could include the integration of deep learning frameworks for positive cell detection, detection of partial staining, and alternative segmentation methods for general cell segmentation. While currently applied exclusively to TMA cores the pipeline could also be expanded to whole slide analysis, including scoring of virtual TMA cores from whole-slide sections. Additionally, applying this framework to multiplex stainings on e.g., the PhenoImager-platform could allow for an even greater potential for analysing cell-cell interactions.

We present an easy-to-use open-source automated pipeline for IHC-based cell detection and quantification, providing an effective exploratory tool for cancer image analysis. The results of this approach in TNBC are comparable to those generated by more advanced machine learning algorithms and consistent with estimates from expert pathologists, quantifications derived from RNA-sequencing, and known prognostic factors. The latter includes the known association of a general immune response with favourable prognosis in TNBC, but also an exemplification of how merging an in situ TME readout such as CD3-positive cell infiltration can be combined with a genetic classification of HRD to further identify a good-prognosis subgroup of patients. Our approach also allows for the spatial positioning of identified cell objects, which can be used to align and merge single-plex staining experiments into composite images for potential cell-cell interaction analysis. While the method has its limitations when identifying individual cells, it may serve as a basis for cell annotation required for fully supervised deep learning methods. Overall, digital image analysis tools will likely facilitate the next step in TME characterization of breast cancer and other malignancies. Using a more quantitative approach to TME profiling allows for more precise and detailed integrations with omics data such as DNA methylation, gene expression profiling and WGS, which should facilitate the next step in phenotypic and integrative cancer studies

DL: Deep learning

HRD: homologous recombination deficiency

IM: immunomodulatory subtype

ICI: immune checkpoint inhibitor

IDFS: invasive disease-free survival

LAR: luminal androgen receptor subtype

SCAN-B: Sweden Cancerome Analysis Network – Breast study

TIL: tumour infiltrating lymphocyte

TME: tumour microenvironment

TNBC: triple-negative breast cancer

TPM: transcripts per million

WGS: whole genome sequencing

ETHICS APPROVAL AND CONSENT TO PARTICIPATE

All included patients were enrolled in the Sweden Cancerome Analysis Network – Breast (SCAN-B) study (ClinicalTrials.gov ID NCT02306096) (21, 22), approved by the Regional Ethical Review Board in Lund, Sweden (Registration numbers 2009/658, 2010/383, 2012/58, 2016/742, 2018/267 and 2019/01252) as previously described (3). All patients provided written informed consent prior to enrolment. All analyses were performed following local and international regulations for research ethics in human subject research.

CONSENT FOR PUBLICATION

Not applicable.

CONFLICT OF INTEREST STATEMENT

The authors declare that they have no competing interests.

AVAILABILITY OF DATA AND MATERIAL

The genomic datasets supporting the conclusions of this article are available in open repositories as described in the original studies. The TMArQ IHC-based cell counts supporting the conclusions of this article are included within the article as Additional file 2 (Supplementary Table 2).

CODE AVAILABILITY

All analyses were performed using open-source software such as the R statistical language and Python. The TMArQ pipeline is publicly available at https://github.com/StaafLab/TMArQ.

FUNDING

Financial support for this study was provided by the Swedish Cancer Society (CAN 2021/1407, and a 2018 Senior Investigator Award [JS]), the Mrs Berta Kamprad Foundation (FBKS-2020-5 + FBKS-21-27(328)), the Swedish Research Council (2021-01800), and Swedish governmental funding (ALF, grant 2022/0021), as well as a clinical research internship from Region Skåne (MA). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the paper.

AUTHORS CONTRIBUTIONS

Conception and design: Suze Roostee, Johan Staaf, Mattias Aine

Collection and assembly of data: Suze Roostee, Johan Staaf, Mattias Aine

Provision of study material or patients: Johan Staaf

Data analysis and interpretation: Suze Roostee, Mats Jönsson, Daniel Ehinger, Johan Staaf, Mattias Aine

Financial support: Johan Staaf, Mattias Aine

Administrative support: Johan Staaf, Suze Roostee, Mattias Aine

Manuscript writing: Suze Roostee, Johan Staaf, and Mattias Aine with input from all authors

Final approval of manuscript: All authors

Agree to be accountable for all aspects of the work: All authors

ACKNOWLEDGEMENT

The authors would like to acknowledge patients and clinicians participating in the SCAN-B study, personnel at the central SCAN-B laboratory at the Division of Oncology and Pathology, Lund University, the Swedish national breast cancer quality registry (NKBC), Regional Cancer Center South, RBC Syd, and the South Sweden Breast Cancer Group (SSBCG).

Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2021;71(3):209-49.
Foulkes WD, Smith IE, Reis-Filho JS. Triple-negative breast cancer. N Engl J Med. 2010;363(20):1938-48.
Staaf J, Glodzik D, Bosch A, Vallon-Christersson J, Reutersward C, Hakkinen J, et al. Whole-genome sequencing of triple-negative breast cancers in a population-based clinical study. Nat Med. 2019;25(10):1526-33.
Davies H, Glodzik D, Morganella S, Yates LR, Staaf J, Zou X, et al. HRDetect is a predictor of BRCA1 and BRCA2 deficiency based on mutational signatures. Nat Med. 2017;23(4):517-25.
de Jong VMT, Wang Y, Ter Hoeve ND, Opdam M, Stathonikos N, Jozwiak K, et al. Prognostic Value of Stromal Tumor-Infiltrating Lymphocytes in Young, Node-Negative, Triple-Negative Breast Cancer Patients Who Did Not Receive (neo)Adjuvant Systemic Therapy. J Clin Oncol. 2022;40(21):2361-74.
Adams S, Gray RJ, Demaria S, Goldstein L, Perez EA, Shulman LN, et al. Prognostic value of tumor-infiltrating lymphocytes in triple-negative breast cancers from two phase III randomized adjuvant breast cancer trials: ECOG 2197 and ECOG 1199. J Clin Oncol. 2014;32(27):2959-66.
Park JH, Jonas SF, Bataillon G, Criscitiello C, Salgado R, Loi S, et al. Prognostic value of tumor-infiltrating lymphocytes in patients with early-stage triple-negative breast cancers (TNBC) who did not receive adjuvant chemotherapy. Ann Oncol. 2019;30(12):1941-9.
Petitprez F, Sun CM, Lacroix L, Sautes-Fridman C, de Reynies A, Fridman WH. Quantitative Analyses of the Tumor Microenvironment Composition and Orientation in the Era of Precision Medicine. Front Oncol. 2018;8:390.
Debien V, De Caluwe A, Wang X, Piccart-Gebhart M, Tuohy VK, Romano E, et al. Immunotherapy in breast cancer: an overview of current strategies and perspectives. NPJ Breast Cancer. 2023;9(1):7.
Hammerl D, Martens JWM, Timmermans M, Smid M, Trapman-Jansen AM, Foekens R, et al. Spatial immunophenotypes predict response to anti-PD1 treatment and capture distinct paths of T cell evasion in triple negative breast cancer. Nat Commun. 2021;12(1):5668.
Aine M, Boyaci C, Hartman J, Hakkinen J, Mitra S, Campos AB, et al. Molecular analyses of triple-negative breast cancer in the young and elderly. Breast Cancer Res. 2021;23(1):20.
Tsang JY, Tse GM. Update on triple-negative breast cancers - highlighting subtyping update and treatment implication. Histopathology. 2023;82(1):17-35.
Lehmann BD, Jovanovic B, Chen X, Estrada MV, Johnson KN, Shyr Y, et al. Refinement of Triple-Negative Breast Cancer Molecular Subtypes: Implications for Neoadjuvant Chemotherapy Selection. PLoS One. 2016;11(6):e0157368.
Hu B, Sajid M, Lv R, Liu L, Sun C. A review of spatial profiling technologies for characterizing the tumor microenvironment in immuno-oncology. Front Immunol. 2022;13:996721.
Brueffer C, Vallon-Christersson J, Grabau D, Ehinger A, Hakkinen J, Hegardt C, et al. Clinical Value of RNA Sequencing-Based Classifiers for Prediction of the Five Conventional Breast Cancer Biomarkers: A Report From the Population-Based Multicenter Sweden Cancerome Analysis Network-Breast Initiative. JCO Precis Oncol. 2018;2.
Chebil G, Bendahl PO, Ferno M, South Sweden Breast Cancer G, North Sweden Breast Cancer G. Estrogen and progesterone receptor assay in paraffin-embedded breast cancer--reproducibility of assessment. Acta Oncol. 2003;42(1):43-7.
Bai Y, Cole K, Martinez-Morilla S, Ahmed FS, Zugazagoitia J, Staaf J, et al. An Open-Source, Automated Tumor-Infiltrating Lymphocyte Algorithm for Prognosis in Triple-Negative Breast Cancer. Clin Cancer Res. 2021;27(20):5557-65.
Rojas F, Hernandez S, Lazcano R, Laberiano-Fernandez C, Parra ER. Multiplex Immunofluorescence and the Digital Image Analysis Workflow for Evaluation of the Tumor Immune Environment in Translational Research. Front Oncol. 2022;12:889886.
Shamai G, Livne A, Polonia A, Sabo E, Cretu A, Bar-Sela G, et al. Deep learning-based image analysis predicts PD-L1 status from H&E-stained histopathology images in breast cancer. Nat Commun. 2022;13(1):6753.
Caicedo JC, Cooper S, Heigwer F, Warchal S, Qiu P, Molnar C, et al. Data-analysis strategies for image-based cell profiling. Nat Methods. 2017;14(9):849-63.
Saal LH, Vallon-Christersson J, Hakkinen J, Hegardt C, Grabau D, Winter C, et al. The Sweden Cancerome Analysis Network - Breast (SCAN-B) Initiative: a large-scale multicenter infrastructure towards implementation of breast cancer genomic analyses in the clinical routine. Genome Med. 2015;7(1):20.
Ryden L, Loman N, Larsson C, Hegardt C, Vallon-Christersson J, Malmberg M, et al. Minimizing inequality in access to precision medicine in breast cancer by real-time population-based molecular analysis in the SCAN-B initiative. Br J Surg. 2018;105(2):e158-e68.
Staaf J, Hakkinen J, Hegardt C, Saal LH, Kimbung S, Hedenfalk I, et al. RNA sequencing-based single sample predictors of molecular subtype and risk of recurrence for clinical assessment of early-stage breast cancer. NPJ Breast Cancer. 2022;8(1):94.
Newman AM, Steen CB, Liu CL, Gentles AJ, Chaudhuri AA, Scherer F, et al. Determining cell type abundance and expression from bulk tissues with digital cytometry. Nat Biotechnol. 2019;37(7):773-82.
van der Walt S, Schonberger JL, Nunez-Iglesias J, Boulogne F, Warner JD, Yager N, et al. scikit-image: image processing in Python. PeerJ. 2014;2:e453.
Ruifrok AC, Johnston DA. Quantification of histochemical staining by color deconvolution. Anal Quant Cytol Histol. 2001;23(4):291-9.
Schmidt U, Weigert M, Broaddus C, Myers G, editors. Cell detection with star-convex polygons. Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11; 2018: Springer.
Naylor P, Lae M, Reyal F, Walter T. Segmentation of Nuclei in Histopathology Images by Deep Regression of the Distance Map. IEEE Trans Med Imaging. 2019;38(2):448-59.
Harano K, Wang Y, Lim B, Seitz RS, Morris SW, Bailey DB, et al. Rates of immune cell infiltration in patients with triple-negative breast cancer by molecular subtype. PLoS One. 2018;13(10):e0204513.
Thompson KJ, Leon-Ferre RA, Sinnwell JP, Zahrieh DM, Suman VJ, Metzger FO, et al. Luminal androgen receptor breast cancer subtype and investigation of the microenvironment and neoadjuvant chemotherapy response. NAR Cancer. 2022;4(2):zcac018.
Kraiss S, Spiess S, Reihsaus E, Montenarh M. Correlation of metabolic stability and altered quaternary structure of oncoprotein p53 with cell transformation. Exp Cell Res. 1991;192(1):157-64.
Finlay CA, Hinds PW, Tan TH, Eliyahu D, Oren M, Levine AJ. Activating mutations for transformation by p53 produce a gene product that forms an hsc70-p53 complex with an altered half-life. Mol Cell Biol. 1988;8(2):531-9.
Laptenko O, Tong DR, Manfredi J, Prives C. The Tail That Wags the Dog: How the Disordered C-Terminal Domain Controls the Transcriptional Activities of the p53 Tumor-Suppressor Protein. Trends Biochem Sci. 2016;41(12):1022-34.
Desmedt C, Haibe-Kains B, Wirapati P, Buyse M, Larsimont D, Bontempi G, et al. Biological processes associated with breast cancer clinical outcome depend on the molecular subtypes. Clin Cancer Res. 2008;14(16):5158-65.
Salgado R, Denkert C, Demaria S, Sirtaine N, Klauschen F, Pruneri G, et al. The evaluation of tumor-infiltrating lymphocytes (TILs) in breast cancer: recommendations by an International TILs Working Group 2014. Ann Oncol. 2015;26(2):259-71.
Nederlof I, De Bortoli D, Bareche Y, Nguyen B, de Maaker M, Hooijer GKJ, et al. Comprehensive evaluation of methods to assess overall and cell-specific immune infiltrates in breast cancer. Breast Cancer Res. 2019;21(1):151.
Dagogo-Jack I, Shaw AT. Tumour heterogeneity and resistance to cancer therapies. Nat Rev Clin Oncol. 2018;15(2):81-94.
Buisseret L, Desmedt C, Garaud S, Fornili M, Wang X, Van den Eyden G, et al. Reliability of tumor-infiltrating lymphocyte and tertiary lymphoid structure assessment in human breast cancer. Mod Pathol. 2017;30(9):1204-12.
Inge LJ, Dennis E. Development and applications of computer image analysis algorithms for scoring of PD-L1 immunohistochemistry. Immunooncol Technol. 2020;6:2-8.
Schmid P, Rugo HS, Adams S, Schneeweiss A, Barrios CH, Iwata H, et al. Atezolizumab plus nab-paclitaxel as first-line treatment for unresectable, locally advanced or metastatic triple-negative breast cancer (IMpassion130): updated efficacy results from a randomised, double-blind, placebo-controlled, phase 3 trial. Lancet Oncol. 2020;21(1):44-59.
Miles D, Gligorov J, Andre F, Cameron D, Schneeweiss A, Barrios C, et al. Primary results from IMpassion131, a double-blind, placebo-controlled, randomised phase III trial of first-line paclitaxel with or without atezolizumab for unresectable locally advanced/metastatic triple-negative breast cancer. Ann Oncol. 2021;32(8):994-1004.
Locy H, Verhulst S, Cools W, Waelput W, Brock S, Cras L, et al. Assessing Tumor-Infiltrating Lymphocytes in Breast Cancer: A Proposal for Combining Immunohistochemistry and Gene Expression Analysis to Refine Scoring. Front Immunol. 2022;13:794175.
Danaher P, Warren S, Dennis L, D'Amico L, White A, Disis ML, et al. Gene expression markers of Tumor Infiltrating Leukocytes. J Immunother Cancer. 2017;5:18.
Sturm G, Finotello F, Petitprez F, Zhang JD, Baumbach J, Fridman WH, et al. Comprehensive evaluation of transcriptome-based cell-type quantification methods for immuno-oncology. Bioinformatics. 2019;35(14):i436-i45.
Aran D, Hu Z, Butte AJ. xCell: digitally portraying the tissue cellular heterogeneity landscape. Genome Biol. 2017;18(1):220.
Lehmann BD, Bauer JA, Chen X, Sanders ME, Chakravarthy AB, Shyr Y, et al. Identification of human triple-negative breast cancer subtypes and preclinical models for selection of targeted therapies. J Clin Invest. 2011;121(7):2750-67.
Echavarria I, Lopez-Tarruella S, Picornell A, Garcia-Saenz JA, Jerez Y, Hoadley K, et al. Pathological Response in a Triple-Negative Breast Cancer Cohort Treated with Neoadjuvant Carboplatin and Docetaxel According to Lehmann's Refined Classification. Clin Cancer Res. 2018;24(8):1845-52.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Automated image quantification of immunohistochemistry-stained immune cells in triple-negative breast cancer

Status:

Version 1

Abstract

Background:

Methods:

Results:

Conclusions:

Figures

BACKGROUND

METHODS

Patient cohort, clinical data, and ethics approval

TMA construction, IHC staining and TIL estimates

RNA-sequencing and whole genome sequencing data

Quantification of immune antibodies in IHC

Integration of IHC quantifications with additional data

RESULTS

Comparison of automated TMArQ pipeline cell counts to available pathology scores

Comparison of TMArQ automated pipeline results to matched bulk tissue RNA-sequencing data

Antibody to antibody comparisons and joint antibody expression patterns

Automated TMArQ cell counts are associated with differences in patient outcome in TNBC

DISCUSSION

CONCLUSIONS

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1