Content Detail

• Content META

  Title	Machine learning-based spatial characterization of tumor-immune microenvironment in the EORTC 10994/BIG 1-00 early breast cancer trial
  ID	7540583
  External ID	7608997782
  Author	Unknown
  Published Date	March 7, 2025, 11:28 a.m.
  Summary	Study design and patient cohort description The present analysis is a translational substudy of the multicenter, international, randomized phase III EORTC 10994/BIG 1-00 clinical trial (Clinicaltrial
  Tags	None
  URL	https://www.nature.com/articles/s41523-025-00730-1
  Sentiment	Negative: 0.02334    Neutral: 0.94497    Positive: 0.03170
  Relevance	0.00000

• Content Details

  Content ID	4845829
  Key	7ec91eef-550e-41a4-80d0-2a9c9bfda1e5
  Content	Study design and patient cohort description The present analysis is a translational substudy of the multicenter, international, randomized phase III EORTC 10994/BIG 1-00 clinical trial (Clinicaltrials.gov identifier NCT00017095). The study enrolled women (<71 years) with histologically-proven locally advanced, inflammatory, or large operable primary invasive BC who were candidates for neoadjuvant chemotherapy. Patients were randomly assigned in a 1:1 ratio to receive six cycles of taxane-based (docetaxel for three cycles followed by three cycles of epirubicin + docetaxel, T-ET; 928 patients) or non-taxane-based (5- fluorouracil, epirubicin, cyclophosphamide, FEC100 or tailored FEC; 928 patients) chemotherapy, prior to surgery. Post-surgery therapy included hormonal treatment and radiotherapy, according to the protocol guidelines. None of the patients received neoadjuvant trastuzumab or radiotherapy before surgery10. Patients who had signed informed consent, with available tissue microarray (TMA) and with IHC-based subtype defined per TMA were included in this translational study. The analyses are reported according to the Reporting Recommendations for Tumor Marker Prognostic Studies (REMARK) guidelines42 (Supplementary Table 9). The EORTC 10994/BIG 1-00 randomized phase III trial was registered at ClinicalTrials.gov (NCT00017095) and approved by the Ethics Committees in all participating centers (Supplementary Table 10). The proposed translational study has been reviewed and approved by the EORTC Translational Research Advisory Committee and Headquarters and the tissue material has been handled according to the signed material transfer agreement by the two organizations (Karolinska Institutet and EORTC). Only the tissue from the patients who have previously provided/signed informed consent for additional biologic research on their samples was included and analyzed. All patients have signed informed consent -prior registration- for inclusion in the trial and the assessment of p53 mutational status. The present study has been performed according to the Declaration of Helsinki, principles of Good Clinical Practice and was also approved by the Swedish Medical Product Agency and the Regional Ethical Committee in Stockholm (Dnr 01-387, Dnr 2005/472-32, Dnr 2005/738-32, Dnr 2008/897-32, Dnr 2013/954-32, Dnr 2018/540-32, Dnr 2021-05174, Dnr 2022-00657-02). Tissue sample preparation and histopathology assessment Formalin-fixed paraffin-embedded (FFPE) tumor tissue blocks from the initial diagnostic biopsies (before initiation of neoadjuvant chemotherapy) have been collected from all patients. Hematoxylin & eosin (H&E)-stained slides were reviewed by a certified pathologist (G.M.) and representative tumor-rich areas were selected for TMA construction, performed at Institut Bergonié, Bordeaux, France. For each patient, up to three cores (0.6 mm in diameter each) were used for the TMA, as previously described43. TP53 mutational status analysis (assessed via a yeast functional assay and further validated through Sanger and next-generation sequencing) has been previously performed on baseline frozen tumor biopsies from all patients included in the study, with most (80%) having evaluable TP53 mutational status10,43,44. Tumor IHC-based phenotyping and biomarker interpretation were performed on TMA as follows: Estrogen receptor (ER) and progesterone receptor (PR) were defined as positive in case of ≥1% tumor cell expression; high Ki67 expression was defined as ≥14% of positive cells; human Epidermal Growth Factor Receptor 2 (HER2) was evaluated according to the ASCO/CAP 2013 recommendations and was considered positive if immunohistochemistry (IHC) 3+ or if IHC 2+ and ≥6 HER2 gene copies using in situ hybridization43. Tumors were classified into three subtypes as such: hormone receptor-positive/HER2-negative (HR+/HER2-negative), HER2-positive and triple negative (TN). Histological grade, type as well as pathological response (after neoadjuvant chemotherapy) were locally evaluated at each participating center. Digital evaluation of tumor-infiltrating lymphocytes One FFPE section (4 μm thickness) was obtained from each TMA tissue block and stained with hematoxylin and eosin (H&E). H&E slides were digitized using the Nanozoomer 2.0-HT (Hamamatsu Photonics K.K.) platform at 20× magnification. TILs were subsequently enumerated using the digitally-assisted, image-based automated scoring CNN11 algorithm13,45. This previously trained and optimized classifier -which can be used with the open-source QuPath software- was used for detecting lymphocytes, tumor cells, stromal cells (e.g., fibroblasts) and “other” cells and for defining annotation and accumulative areas of each cell type (measured in mm2). Quality control of the tissue morphology (artifacts, necrosis) and cell segmentation was performed and subsequently five different TILs variables were derived from the different annotations as follows: i) eTILs (%) = 100 * (TILs/sum of tumor cells and TILs); ii) etTILs (%) = 100 * (TILs/All detected cells (i.e., tumor cells + fibroblasts + other cells)); iii) esTILs (%) = 100 * (TILs/Stromal cells (i.e TILs + fibroblasts + other cells)); iv) eaTILs (mm2) = TILs/tumor region areas analyzed; v) easTILs (%) = 100 * (sum of TILs area (mm2)/stroma area (mm2)), mirroring the definition provided by the International TILs Working Group guidelines)1. Intratumoral TILs were included in the respective calculations of all the aforementioned metrics. Multiplex immunofluorescence staining, image processing, and analysis Staining We performed multiplex immunofluorescence staining on FFPE TMA tumor tissue sections (thickness: 4 μm), using the OpalTM 7-color Solid Tumor Immunology Kit (Akoya Biosciences, Marlborough, MA, USA), according to the manufacturer’s instructions. The automated Leica Bond RXm system (Leica Biosystems, Buffalo Grove, IL, USA) was used for sequential staining with a panel of lymphocytic and macrophage markers using antibodies against CD4, CD8, PD-L1, PD-1, FoxP3, CD68. For the detection of epithelial tissue, a combination of antibodies against cytokeratin and E-cadherin was used, as previously described36,46,47. The list of antibodies, reagents, and experimental conditions used for the multiplex immunofluorescence staining are listed in Supplementary Table 11. Imaging, image analysis, and quality control The stained TMA were imaged using the Vectra® PolarisTM Automated Quantitative Pathology Imaging System (Akoya Biosciences, Marlborough, MA, USA) in multispectral mode at a resolution of 0,496 μm/pixel, which resulted in a total of 2350 multispectral images of size 0.93 mm × 0.7 mm. Each of the images was manually reviewed by three investigators (I.Z., A.M., C.B.) and curated to exclude artifacts, staining defects, and accumulation of immune cells in necrotic areas and intraglandular structures. Any discrepancies were resolved by a pathologist (A.Mez.). The number of images used for further analysis was 1414. The vendor-provided machine learning algorithm, implemented in the inForm® image analysis software, was trained to split tissue into three categories: tumor compartment, stromal compartment, or blank areas. The training was performed on a selection of representative cores by providing a set of samples that was manually annotated. Cell segmentation was performed using DAPI nuclear staining. The perinuclear region at 7 pixels from the nuclear border was considered as the cytoplasm area. Cell phenotyping The cell phenotyping function of the inForm® image analysis software was used to manually define a representative subset of cells positive to expression of each of the markers and a subset of cells negative to all markers (also considering the rare immune phenotypes). The marker expression was evaluated in the cytoplasm of the segmented cells as the number of photons, normalized to exposure time. Exceptions were made for i) FoxP3 (expressed in nuclei and thus its signal was analyzed in the nuclear region of segmented cells) and ii) CD4 marker, which demonstrated more diffuse staining (probably due to epitope instability) often covering also nuclei regions, and thus, the average marker expression level in total cell region (nucleus + cytoplasm) was used48. Intensity cutoffs for the markers were determined in the R programming environment (version 3.6.0). The marker-specific cutoffs were defined by analyzing the distributions of the positive intensities for the marker in the manually-annotated cells and by controlling for the background levels visually and in the negatively annotated cells. The empirical cumulative distribution function was used for visual control. The minimal cutoff was selected as minimal recorded signal level (among the manually annotated cells) if the empirical cumulative distribution demonstrated steep monotonous tendency. In case of individual low-level outliers, they were considered as negative. In case of prominent “kneepoint” in the graph the expression of the marker was back-controlled by reviewing the annotated images and controlling for the marker expression. Additionally, for each of the markers the images were screened for the areas of non-specific staining which were considered as background. Due to the high fraction of CK-positive cells in the entire cohort, a different approach was used to define its cutoff. The complete dataset, containing 1,790,165 cells, was analyzed to evaluate the distribution of CK expression, measured as marker mean expression in cell cytoplasm. The distribution had two peaks, reflecting non-CK cells (low expression) and CK-positive cells (high expression) and Otsu algorithm49 was used to find minima between two groups. Cell classification Using the established cutoff levels, every cell was characterized as positive or negative for each marker in the panel. These data were used to classify the cell and define its immune subtype as illustrated in Supplementary Table 12. Multiple TMA cores from the same tumor/patient were merged and cell abundance was normalized to the total tissue area. Cell subclasses were quantified and normalized to tissue area thus resulting in final metric called “cell density” (units per mm2), which was calculated in total tissue area, stroma, and tumor regions for each cell subclass. For illustration purposes, the cell density values were log2- transformed. The multispectral imaging/analysis workflow is depicted in Supplementary Fig. 7. Spatial image analysis and topographical interactions The previously described Spatial Image Analysis of Tissues (SPIAT) toolkit50 was utilized to facilitate the detection of cellular interactions among immune cells (including regulatory T-cells, cytotoxic T-cells, macrophages, and T-helpers and the respective expression of PD-L1/PD-1 checkpoints as reference cells) with tumor cells (target cells) and spatial cell localization. The applied metrics included the normalized mixing score and the entropy gradient which both evaluated the co-localization of immune and tumor cells. The analysis was performed on single reconstructed images from multiple patient-specific TMA stitched with a proper distance between them to avoid potential overlaps. Normalized mixing score (NMS) and the “aggregated entropy” were calculated as a function of different radii around the reference cells (immune cells here), varying from 50 up (close-distance interactions) to 600 μm to cover the whole TMA area with a step of 50μm. The NMS uses a normalization scale factor -accounting for cell number- and receives a numerical value according to the following formula, described in SPIAT: NMS = number of interactions between reference and target cells × (total number of reference cells – 1)/2 × number of interactions between reference cell types x total number of target cells. The higher the score, the higher the degree of co-localization between immune and tumor cells. Given that cell number and types in TMAs are fixed, the definition of entropy entails mainly the variability in cell number rather than degree of randomness. Immune cells were selected as reference, target cells were identified for each selected radius around the reference cells and entropy scores were subsequently calculated for each radius (“aggregated entropy” as described in the SPIAT toolkit)50. Higher entropy scores correspond to high cell colocalization between reference and target cells and vice versa for low entropy scores (unbalanced cell types – low colocalization). Next, patient-specific gradient entropy curves were generated from the “aggregated entropy” values along the different radii (for illustration purposes all curves were graphically displayed following a 10μm radius step) and classified either as “Attraction-like” or “Repulsion-like” patterns according to the slope of the curve. If the slope was negative close to the lowest radius, the pattern of interaction was defined as ”Attraction”, indicating higher aggregated entropy values close to reference cells, thus high colocalization. If the slope was positive close to the lowest radius, the pattern of interaction is defined as ”Repulsion”, indicating higher aggregated entropy values far from reference cells, thus low colocalization. In case of no difference on entropy values across the radii, there was a zero slope50. The aforementioned analyses were performed in the R programming environment. Outcomes definition and statistical analysis Pathologic complete response (pCR) was defined as in the original publication of the primary efficacy analysis of the trial as the absence of residual invasive cancer (or very few scattered cancer cells left), with or without residual DCIS and with negative axillary lymph nodes (ypT0/is ypN0). Non-pCR patients included those who presented with tumor progression on neoadjuvant chemotherapy or who did not undergo surgery43. Patients with missing information were excluded from the analysis. Progression-free survival (PFS), which was the primary endpoint of the trial10, was defined as the time from randomization to progression on neoadjuvant therapy, locoregional relapse (invasive cancer), first distant metastasis, invasive contralateral breast cancer or death from any cause, whichever occurred first. Second primary invasive non-breast cancers, DCIS or LCIS (ipsilateral or contralateral) were not considered as events. Continuous outcomes were tested using the Wilcoxon rank-sum or Kruskal–Wallis test while binary outcomes with Pearson’s chi-squared or Fisher’s exact test. The Spearman’s rank correlation coefficient was used for correlations between continuous variables. The different immune cell subpopulations and digital metrics derived by dTILs and multiplex immunofluorescence were correlated to pCR using univariate and multivariable logistic regression models. Median follow-up was calculated using the reverse Kaplan-Meier method. Survival (PFS) was calculated using the Kaplan-Meier method and differences in PFS were calculated with the log-rank test. Time to failure was modeled using proportional hazard regression. Logistic regression and Cox regression models are presented as odds ratios and hazard ratios, respectively, with the respective 95% confidence intervals and Wald p-values. Multivariable models were adjusted for tumor size, nodal status and treatment, and stratified for subtype since subtype violated proportionality of hazards. All p-values were two-sided with 5% as the level of significance. Adjustment for multiplicity was not performed. All analyses were performed the Stata software (v. 17, StataCorp, College Station, TX, USA).
  Content Reference URL	None

• Publisher

  Publisher Name	Nature (#2153)
  Publisher URL	https://www.nature.com/
  Description	None
  Trust Flow Score	61
  Citation Flow Score	58

• Data Source

  Data Source Name	Quantexa News (#5)
  Data Source Type	news
  Data Source URL	https://app.aylien.com/dashboard
  Description	Welcome to Quantexa News API — the most powerful way to aggregate, filter, and integrate global news data into your apps and models.
  Weight	10