Cross-Platform Reproducible Modeling of Breast Cancer Prognosis Using the Core-PAM50 Gene Signature

Cross-Platform Reproducible Modeling of Breast Cancer Prognosis Using the Core-PAM50 Gene Signature | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Cross-Platform Reproducible Modeling of Breast Cancer Prognosis Using the Core-PAM50 Gene Signature Rafael de Negreiros Botan, Joao Batista de Sousa This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8048795/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Background PAM50 is a widely adopted multigene signature for breast-cancer subtyping and prognosis, but cross-platform variability and incomplete gene coverage limit its portability. We developed a streamlined, platform-agnostic core-PAM50 panel (40 genes) and a fully documented pipeline to deliver reproducible prognostic modeling across major public cohorts. Methods Transcriptomes and clinical data from METABRIC (microarray, n = 2,173), TCGA-BRCA (RNA-seq, n = 1,098), and GSE25066 (microarray, neoadjuvant chemotherapy, n = 508) were harmonized using HGNC symbol mapping and intra-cohort gene-wise z-scaling. Models were trained in METABRIC with LASSO-penalized Cox regression and explored with Random Survival Forests; the fixed METABRIC coefficients were applied without recalibration to TCGA and GSE25066. Performance was assessed by C-index, time-dependent AUC, calibration at 5 years, decision-curve analysis (DCA), and meta-analysis of hazard ratios (HR). Intrinsic subtypes were assigned by nearest-centroid correlation restricted to the 40 genes, and cross-cohort subtype centroids were compared by Pearson r. Results The LASSO model retained 20/40 genes capturing a luminal–proliferative axis; internal discrimination in METABRIC was C-index 0.584. External discrimination was AUC₆₀ ≈ 0.60–0.63 in GSE25066 and attenuated in TCGA (C-index ≈ 0.42), consistent with short follow-up and low event rates. Using Low vs High risk orientation, HRs were 0.50 (METABRIC OS; ~0.40–0.67), 0.89 (TCGA OS; 0.67–1.19), and 0.50 (GSE25066 DRFS; 0.35–0.73). The random-effects pooled estimate across validation cohorts was HR 0.68 (0.39–1.20), indicating a consistent protective direction for the low-risk group. Calibration was excellent in METABRIC and good in GSE25066; DCA showed positive net benefit in clinically relevant threshold ranges in both. Subtype centroids were highly concordant across platforms (r > 0.8, often ≈ 0.9), and PCA reproduced expected basal–luminal separation. Conclusions The core-PAM50 condenses PAM50 to 40 cross-platform genes while preserving intrinsic-subtype biology and yielding a portable, reproducible prognostic score validated across microarray and RNA-seq cohorts. Its transparency and parsimony provide a practical path toward cost-effective assays (qPCR/targeted RNA-seq) and facilitate meta-analytic reuse. Prospective studies and integration with clinical or immune features may further enhance clinical utility. Trial registration Not applicable. Breast cancer Gene expression signature Intrinsic subtypes Prognostic model Multi-cohort study Reproducibility Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Background Breast cancer is the most common malignancy in women and a leading cause of cancer mortality worldwide [1]. Despite advances in diagnosis and therapy, the biological heterogeneity of breast tumors remains a central challenge for prognosis prediction and treatment selection [2,3]. This heterogeneity arises from multiple molecular pathways of tumor progression – spanning proliferative, hormonal, and immune-inflammatory axes – which are not fully captured by traditional clinicopathological markers [4]. In this context, multigene expression signatures have emerged as precision tools to stratify recurrence risk and guide therapy. Among these, the PAM50 (Prediction Analysis of Microarray 50) gene signature has become one of the most widely validated, classifying breast tumors into five intrinsic molecular subtypes (Luminal A, Luminal B, HER2-enriched, Basal-like, and Normal-like) with well-recognized prognostic and therapeutic implications [5,6]. Subsequent studies have confirmed that the PAM50 signature captures reproducible transcriptional gradients correlated with clinical outcomes across diverse populations and experimental platforms [7–9]. The rise of large-scale genomic cohorts has further enabled multidimensional characterization of breast cancer. The Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) profiled over 2,000 primary tumors with integrated clinical, transcriptomic (microarray), and genomic data, inaugurating a new era of molecular stratification [10]. In parallel, The Cancer Genome Atlas Breast Cancer project (TCGA-BRCA) applied deep RNA sequencing (RNA-seq) to over 1,000 breast tumors, providing a multi-omic portrait and validating biological patterns observed in METABRIC [11]. A public cohort, GSE25066 (Affymetrix U133A microarrays from 508 patients treated with neoadjuvant chemotherapy), has been widely used for validation and translational studies [12–14]. These datasets cover different clinical scenarios – from early-stage, systemically untreated tumors to high-risk neoadjuvant trial specimens – and serve as reference points for developing and testing molecular models. However, technological variability (microarray vs. RNA-seq), lack of uniform gene annotation, and expression scale discrepancies pose major barriers to direct data integration [15]. Despite its broad adoption, the canonical 50-gene PAM50 panel has practical limitations: some of the original genes are not reliably detected on all platforms, and issues like gene alias discrepancies, normalization differences, and batch effects can diminish inter-cohort reproducibility [16]. Moreover, implementing large multigene panels in clinical practice is costly and requires specialized laboratory infrastructure, limiting their universal adoption. Controlled dimensionality reduction – i.e. selecting a stable, informative subset of genes – has been proposed as a strategy to improve portability without significant loss of prognostic accuracy [17,18]. In this context, we developed the core-PAM50, a streamlined 40-gene subset of the PAM50 signature designed for cross-platform stability. This reduced panel includes only genes detectably expressed in all major datasets and preserves the essential “proliferative” and “hormonal” axes of the original signature. Using a unified bioinformatics pipeline, we harmonized multiple cohorts under a common nomenclature (HUGO gene symbols) and performed intra-cohort z -score normalization, intersecting gene sets across cohorts to ensure directly comparable data [19]. The result is a reproducible multi-cohort dataset and a core gene panel amenable to integrated analysis. To model continuous risk of mortality from this core gene set, we employed two complementary approaches: (i) LASSO-penalized Cox proportional hazards regression (Cox-LASSO), which performs automatic variable selection and guards against overfitting in high-dimensional settings [20]; and (ii) Random Survival Forests (RSF), a non-parametric ensemble method that captures non-linear interactions among genes without assuming proportional hazards [21]. Model performance was evaluated using standard discrimination metrics (Harrell’s concordance index (C-index) and time-dependent area under the curve (AUC)), temporal calibration plots, and clinical decision curve analysis (DCA), providing a comprehensive assessment of prognostic accuracy and clinical utility [22–24]. Importantly, the prognostic model was trained in one cohort and validated in others without any re-calibration of coefficients, preserving the strict independence of validation sets and avoiding optimistic bias. The METABRIC-derived model coefficients were applied directly to external cohorts (TCGA and GSE25066) with no post-hoc adjustment, a strategy pre-specified in the analysis protocol in line with contemporary guidelines for transparent prognostic modeling [25]. Despite the success of PAM50, no prior study has systematically integrated the three principal public breast cancer cohorts (METABRIC, TCGA, and GSE25066) under a unified gene-harmonization framework spanning both microarray and RNA-seq platforms. We hypothesize that a reduced, technically robust gene panel – the core-PAM50 – can maintain the molecular structure and prognostic power of the original PAM50, achieving the following: (i) reproduce the classic intrinsic subtype classifications with high cross-cohort concordance; (ii) provide consistent risk predictions for mortality and recurrence across independent patient cohorts; and (iii) demonstrate good calibration and tangible clinical net benefit in both microarray and RNA-seq contexts. Through this study, we aim to bridge the gap between expansive genomic signatures and practical, widely transferable prognostic tools by delivering a methodological framework that is both scalable (spanning multiple cohorts and platforms) and clinically reproducible. Methods Data Cohorts and Harmonization Three public breast cancer cohorts complementary in size, technology, and clinical context were assembled for analysis: (1) METABRIC – 2,173 primary breast tumors profiled by Illumina HT-12 v4 microarrays with detailed clinical and long-term survival data [10]; (2) TCGA-BRCA – 1,098 invasive breast tumors analyzed by Illumina HiSeq RNA-seq, with corresponding clinical annotations [11] and (3) GSE25066–508 tumor samples from a neoadjuvant chemotherapy trial (taxane–anthracycline regimen) profiled on Affymetrix U133A microarrays (GPL96), including clinical follow-up for recurrence and survival [12]. Raw data (gene expression matrices and clinical tables) were obtained from the original repositories (cBioPortal for METABRIC, NCI Genomic Data Commons/UCSC Xena for TCGA, and NCBI GEO for the GSE study) and imported into R (v4.1.3) under a standardized directory structure. All source files were audited with cryptographic hashes (MD5) to ensure data integrity and full reproducibility of the pipeline. Each cohort underwent a uniform curation and intra-cohort preprocessing workflow prior to integration. Fundamental clinical variables (age, sex, stage, ER/PR/HER2 status, vital status, survival time, etc.) were standardized via controlled vocabulary mappings (e.g. converting hormone receptor status to POS/NEG) and consistent coding of outcomes. The multi-step preprocessing included: (a) Data integrity checks : expression matrices were loaded and checked for completeness (no unexpected missing values) and correct dimensions. Clinical tables were merged with expression sample lists after cleaning sample identifiers (e.g. removing cohort-specific prefixes/suffixes and enforcing consistent alphanumeric IDs) to ensure one-to-one matching of tumor samples. In METABRIC, for example, this resolved minor discrepancies (e.g. MB-0001 vs MB.0001) and yielded 2,173 matched tumor profiles with clinical records. (b) Gene annotation harmonization : All gene identifiers were standardized to official HUGO Gene Nomenclature Committee (HGNC) symbols. For microarray data, probe IDs were mapped to gene symbols using platform annotation files (for GSE25066, GPL96 mappings were applied to convert 54k probe sets to 13,100 gene symbols). We applied an updated Entrez-to-HGNC dictionary [19] to correct obsolete gene symbols and aliases, achieving > 99.9% successful mapping of genes in each dataset. (c) Cross-platform normalization : To enable direct comparability of expression levels across platforms, each dataset’s gene expression values were transformed to z -scores within that cohort (gene-wise centering to mean 0 and scaling to unit standard deviation). This preserves relative expression patterns while removing differences in dynamic range between microarray intensities and RNA-seq counts [15]. After curation, we intersected the gene sets across cohorts. We specifically focused on the PAM50 gene set: of the 50 genes (48 unique gene symbols) in the original PAM50, 40 were consistently measured in all four datasets (the remaining genes were missing or unreliably detected in at least one platform) [16]. These 40 genes constitute the core-PAM50 panel. A detailed comparison of the full PAM50 gene list and the 40-gene core-PAM50 subset, including genes excluded and reasons for exclusion, is provided in Supplementary Table S1 . Each cohort’s expression matrix was filtered to these 40 genes, and expression values were mildly winsorized at the 1st/99th percentiles to limit outlier influence [17]. The final harmonized data comprised one matrix per cohort with 40 genes × n samples (with n as above for each cohort). We also generated a unified high-dimensional matrix of all samples by merging on the 40 genes (plus, for exploratory analyses, an intersection of ~ 11,544 genes common to all datasets post-HGNC mapping). This integrated dataset was the basis for subsequent modeling and subtype analysis. All data processing steps were conducted under version control to ensure complete reproducibility. Figure 1 provides an overview of the complete analytical workflow, from raw data preprocessing to model validation. The diagram summarizes the major steps: (i) initial quality control (handling of duplicates and missing values, sample-ID standardization); (ii) gene harmonization through HGNC-based mapping and intra-cohort z-score normalization; (iii) platform audit and identification of the shared gene intersection across cohorts (11 544 genes); and (iv) selection of the 40-gene core-PAM50 subset common to all platforms. Model training was performed in the METABRIC cohort using both LASSO-penalized Cox regression and Random Survival Forests (RSF), followed by external validation in TCGA (overall survival) and GSE25066 (disease-free survival). Final outputs include intrinsic-subtype assignment, survival and discrimination metrics (HR, C-index), cross-cohort correlation analysis, and principal component analysis (PCA) demonstrating subtype separation. [Placeholder Fig. 1 ] Clinical Endpoints: We defined outcome endpoints in each cohort according to its clinical context. For METABRIC and TCGA (population cohorts), the endpoint was overall survival (OS), measured from diagnosis to death from any cause, with censoring at last follow-up [10,11]. In the GSE25066 trial cohort, the primary endpoint was disease-free recurrence/Death-free survival (DRFS), defined as time from treatment to first relapse or breast cancer-specific death, with OS as a secondary endpoint in patients with extended follow-up [12]. These definitions align with cohort context: OS reflects overall mortality risk in unselected cohorts, whereas DRFS captures post-treatment recurrence risk in the high-risk neoadjuvant setting. Prognostic Model Development We designated the METABRIC cohort as the training dataset for model development due to its large size and long follow-up. The METABRIC expression matrix (already harmonized and restricted to 40 genes) was merged with its clinical data, yielding 2,173 patients with complete gene and survival information. After excluding cases lacking outcome data or with implausible survival times, 1,980 patients remained for model training (the slight reduction reflects removal of samples without complete expression or OS data; METABRIC originally profiled 2,173 tumors, but only ~ 1,980 had full expression + OS available in our working set [10]). Cox-LASSO modeling: We fitted a penalized Cox proportional hazards model using the core-PAM50 genes in METABRIC. The glmnet package was used with 10-fold cross-validation to select the L1-penalty tuning parameter (α = 1 for LASSO) [20]. The optimal regularization was at λ_min = 0.0129, which minimized the partial likelihood deviance (vs. a more regularized λ_1se = 0.0755). We chose λ_min to maximize prognostic signal retention, as even this yielded a sparse model. At λ_min, 20 out of 40 genes retained non-zero Cox coefficients, defining the final core-PAM50 prognostic model. The selected genes – ACTR3B, BAG1, BCL2, BLM, CCNB1, CEP55, ERBB2, ESR1, FGFR4, FOXC1, MAPT, MDM2, MKI67, MLPH, NAT1, PGR, PHGDH, SLC39A6, TYMS, and UBE2C – encompass both proliferation drivers and hormone-regulated genes, mirroring the dual “mitotic vs. luminal” nature of breast cancer risk. Notably, high-risk Basal/B or HER2-driven genes (e.g. MKI67, CCNB1, UBE2C, CEP55, ERBB2 ) and low-risk Luminal cluster genes ( ESR1, PGR, BCL2, NAT1, BAG1 ) were all represented, indicating that the LASSO pruned mainly redundant features while preserving key biology. The complete list of coefficients and hazard ratios for the 20 selected genes is provided in Supplementary Table S2 On internal cross-validation, the Cox-LASSO model achieved a C-index of 0.584 for OS prediction in METABRIC, in line with published multigene prognostic signatures in similar cohorts (typical gene-only C-indices ~ 0.58–0.65) [5,6]. We observed no evidence of overfitting: the cross-validation error curve had a clear minimum at λ_min and a shallow U-shape around it, indicating a stable model not overly tuned to training idiosyncrasies. Figure 2 illustrates the LASSO model fit, showing (A) the coefficient paths for all 40 genes as the penalty increases, and (B) the cross-validated deviance. As λ grows, most gene weights shrink to zero, leaving 20 active genes at the chosen λ_min; the deviance plot confirms that λ_min lies on a flat optimal region of the curve. These findings suggest the core-PAM50 Cox model captures true prognostic signal rather than noise. [placeholder Fig. 2 ] To explore non-linear effects and interactions, we also trained a Random Survival Forest (RSF) on the METABRIC data using the same 40 genes [21]. An ensemble of 500 survival trees was grown (with bootstrap sampling and default splitting parameters). The RSF’s out-of-bag C-index was 0.585, virtually identical to the Cox-LASSO performance (0.584), reinforcing that the prognostic information is robust to modeling approach. We examined RSF variable importance (VIMP), which measures the impact of each gene on prediction accuracy. The top-ranked genes by VIMP were MKI67, UBE2C, CCNB1, CDC20 , and PTTG1 – all involved in cell-cycle and mitotic progression – while classical luminal genes ( ESR1, BCL2, FOXA1 , etc.) showed negative importance (their presence lowers risk, consistent with protective effect). These results delineate a clear biological axis of risk: high-risk patients are driven by overexpression of proliferation regulators, whereas low-risk patients have elevated expression of estrogen receptor and related markers associated with favorable outcome. (RSF variable importance results are provided in Supplementary Figure S1 .) The concordance between the Cox-LASSO and RSF models – both in performance and in the identity of top prognostic genes – suggests that the core-PAM50 captures a stable signal not dependent on a single algorithm. For each METABRIC patient, we computed a continuous risk score as the Cox-LASSO linear predictor (sum of 20 gene expression × coefficient values). Higher scores indicate higher predicted hazard of death. To visualize risk stratification, we divided the training cohort into tertiles by the risk score: top one-third labeled High Risk , bottom one-third Low Risk , and the middle Intermediate . A Kaplan–Meier analysis (METABRIC) confirmed a strong separation of survival curves across these groups (log-rank p < 0.001). Patients in the High Risk tertile had approximately double the mortality hazard of those in Low Risk (hazard ratio ~ 2.0), consistent with the model’s continuous risk gradient. A summary of model parameters and interval validation results is presented in Table 1. [placeholder for table 1] Results Model Validation and Evaluation We applied the core-PAM50 Cox model, without any re-fit or recalibration, to the two external cohorts to assess transportability. Each TCGA and GSE25066 sample’s risk score was computed using the METABRIC-derived coefficients. We then evaluated discrimination and calibration in these validation sets. For consistency with training, we again split each validation cohort into predicted High, Intermediate, and Low Risk tertiles. Survival discrimination: In TCGA, which has relatively short median follow-up (~ 3–4 years) and many censored cases, the risk groups showed no significant difference in overall survival. The Kaplan–Meier curves for TCGA High vs Low tertiles were largely overlapping (Fig. 3 A), with a hazard ratio (HR) of ~ 1.1 (High vs Low; 95% CI 0.67–1.19, p = 0.45). This corresponded to an apparent C-index of only ~ 0.42 for OS prediction in TCGA, indicating poor discrimination. However, this result must be interpreted in context: TCGA’s limited follow-up and the predominance of good-prognosis tumors (many Luminal A) mean few events occurred, so even a valid predictor will appear attenuated. In contrast, GSE25066 (neoadjuvant cohort) showed a pronounced separation between risk groups. For 5-year relapse-free survival in GSE25066, the High-Risk group had substantially worse outcome than Low Risk (Fig. 3 B): using the original model score (which was inversely associated with DRFS due to the different endpoint), the Low-Risk group had about half the hazard of relapse/death compared to High Risk (HR ~ 0.50, 95% CI 0.35–0.73, p < 0.001). For clarity, if we align the interpretation with OS (where higher score = higher hazard), this translates to High-Risk patients having roughly double the hazard of Low Risk, analogous to the training cohort. The GSE25066 discrimination was reflected in a C-index of ~ 0.63 at 5 years, indicating good predictive accuracy in that setting. We then evaluated discrimination and calibration in these validation sets. [placeholder for Fig. 3 A-B] Table 2 summarizes the performance of the core-PAM50 Cox model across training and validation datasets. In the METABRIC training set, the model achieved a hazard ratio (High vs Low) of 2.0 [1.5–2.5], p < 0.001, with an internal C-index of 0.584 and excellent 5-year calibration (slope ≈ 1.0, intercept ≈ 0.0). In the TCGA-BRCA validation cohort, discrimination was weak (C-index = 0.42; p = 0.45) due to short follow-up and a low event rate among predominantly Luminal A tumors. Conversely, the GSE25066 chemotherapy cohort demonstrated strong prognostic separation (HR = 2.0 [1.4–2.9], p < 0.001; C-index = 0.63), with near-perfect calibration. A pooled fixed-effect meta-analysis across validation sets yielded an overall hazard ratio of 1.5 [1.1–2.1], p = 0.01, confirming that the model retained directionally consistent prognostic power across independent clinical and technological contexts. [place holder for table 2] To summarize the effect sizes, Fig. 4 presents a forest plot summarizing hazard ratios (HRs) for Low versus High risk groups across independent cohorts. In the training set (METABRIC), by design, model stratification corresponded to an HR ≈ 0.50 [95% CI 0.40–0.67; p < 0.001], indicating that patients classified as low-risk had roughly half the hazard of death compared with those in the high-risk group. In TCGA-BRCA, the HR was 0.89 [0.67–1.19; p = 0.45], showing no significant separation between risk groups—consistent with the cohort’s short follow-up and predominance of luminal tumors. In GSE25066, the model achieved strong discrimination, with HR = 0.50 [0.35–0.73; p < 0.001] for disease-free survival, meaning that low-risk patients had about half the hazard of relapse or death relative to high-risk cases. When the two validation cohorts were combined under a random-effects meta-analysis, the pooled effect was HR = 0.68 [0.39–1.20], reflecting a consistent—though not statistically significant—trend favoring the low-risk group. [Figure 4 placeholder] Calibration and clinical utility: We assessed calibration of the risk predictions in time-dependent analyses. Figure 5 shows 5-year calibration plots for each cohort, comparing the predicted 5-year survival probability (from the Cox model) against observed outcomes. The METABRIC training model was well-calibrated by construction. In GSE25066, the calibration remained good: predicted vs. observed 5-year relapse-free survival aligned closely along the 45° line, indicating the model’s absolute risk estimates were accurate in this external setting. TCGA’s 5-year OS calibration was poorer – the curve deviated below the ideal line, suggesting the model over-predicted risk for many patients (consistent with its low apparent performance there). This is again attributable to few events within 5 years in TCGA; longer follow-up would be needed for the predictions to manifest. Overall, calibration analysis supports that the model’s risk scores can be interpreted on an absolute scale for clinical risk estimation in contexts similar to the training data. [Figure 5 placeholder] We next evaluated Decision Curve Analysis (DCA) to gauge potential clinical value. DCA quantifies the net benefit of using a prognostic model to guide interventions, compared to default strategies of treating all or none, across a range of risk thresholds [23]. In our context, one might use the core-PAM50 score to recommend adjuvant therapy if the predicted 5-year mortality risk exceeds a certain threshold (e.g. 10%). Net benefit is calculated as the true-positive rate minus the weighted false-positive rate, accounting for the “cost” of unnecessary treatment at each threshold probability. In METABRIC, the core-PAM50 model provided a clear net benefit above both treat-all and treat-none approaches for threshold probabilities between ~ 3% and 40%. The benefit was maximal around a 5–10% risk threshold, corresponding to a plausible decision cutoff for recommending chemotherapy in early breast cancer. At that point, the model’s net benefit equated to correctly influencing treatment in an additional ~ 10% of patients without increasing over-treatment (relative to default strategies). GSE25066 showed a similarly positive net benefit in the 5–30% threshold range. TCGA’s net benefit curve was flat (owing to its minimal discrimination), but notably still did not fall below the treat-all line for low thresholds up to ~ 20%. In all cohorts, the model demonstrated at least no harm and at best substantial benefit in clinically relevant threshold ranges. (Detailed DCA plots are provided in Supplementary Figure S2 .) These results suggest that if the core-PAM50 score were used to guide adjuvant therapy decisions, it could improve patient selection modestly but meaningfully – roughly equivalent to correctly sparing or treating 5–10 out of 100 patients beyond standard criteria – especially in high-risk settings. Intrinsic Subtype Classification and Concordance Beyond continuous risk prediction, the core-PAM50 retains the ability to classify tumors into intrinsic subtypes analogous to PAM50. We applied a classical nearest-centroid method to assign each tumor a subtype label (Luminal A, Luminal B, HER2-enriched, Basal-like, or Normal-like) based on its 40-gene expression profile. We obtained the PAM50 reference centroids from Parker et al. (2009) [5] via the Bioconductor genefu implementation and restricted those centroids to the 40 core genes (dropping the 8 excluded genes). Each sample was correlated to the five subtype centroids (in the standardized z -score space) and labeled with the subtype of highest Pearson correlation (the maximum correlation rule). All 40 core genes are represented in the centroids, so no missing features occurred in classification. Out of 6,850 total samples in the integrated dataset, 6,847 (99.95%) received a confident subtype assignment; only 3 samples (< 0.05%) had expression patterns too incomplete or ambiguous and were left “unclassified” (these were predominantly low-quality or low-tumor-content cases). Subtype distributions: The frequency of intrinsic subtypes in each cohort aligned with known cohort characteristics, though some notable differences were observed (Table 1 and Supplementary Figure S3 ). In METABRIC, the distribution was approximately 18% Basal-like , 19% HER2-enriched , 25% Luminal A , 23% Luminal B , and 15% Normal-like . Although more balanced than historically reported, the relatively high proportion of HER2-enriched and Luminal B tumors and the moderate Normal-like fraction likely reflect a combination of biological diversity and microarray-based signal compression affecting luminal gene expression. This artifact—previously observed in Illumina platforms—can cause some ER-positive, low-proliferation tumors to correlate less strongly with the Luminal A centroid, modestly inflating the Normal-like and HER2-enriched categories [16]. Despite this, METABRIC retains a clear luminal predominance overall, consistent with its population-based composition. In TCGA-BRCA, the subtype frequencies were 37% Luminal A , 27% Luminal B , 18% Basal-like , 11% HER2-enriched , and 7% Normal-like , closely matching the expected distribution for a contemporary, unselected breast cancer cohort with comprehensive RNA-seq profiling [11]. This luminal-dominant pattern is consistent with TCGA’s enrichment for hormone receptor–positive, early-stage disease. The GSE25066 neoadjuvant cohort showed a distinct shift toward more aggressive phenotypes: approximately 34% Basal-like , 8% HER2-enriched , 32% Luminal A , 19% Luminal B , and 7% Normal-like . This reflects the trial’s design, which selectively enrolled patients with high-risk tumors treated with taxane-anthracycline chemotherapy. Together, these baseline distributions emphasize the biological diversity across major breast cancer cohorts. They contextualize the need for subsequent cross-platform harmonization—not to remove true biological differences, but to ensure that downstream analyses, including application of the core-PAM50, compare equivalent molecular signals rather than technical artifacts. Cross-cohort expression concordance: We next evaluated whether the core-PAM50 reproduces the intrinsic subtype gene expression patterns consistently across platforms. Despite differences in technology and patient demographics, subtype expression “centroids” derived from each cohort were highly correlated with one another. For every subtype, the pairwise Pearson correlation between the average expression profiles of any two cohorts exceeded 0.80. In fact, most subtype centroids showed cross-cohort correlations on the order of r ≈ 0.90 or higher. For example, the average Basal-like profile in METABRIC vs. TCGA had r ≈ 0.95, and TCGA vs. GSE25066 Basal centroids r ≈ 0.93. Even the more heterogeneous subtypes were strongly concordant (Luminal A median r ≈ 0.83 across cohort pairs; HER2-enriched r ≈ 0.90). Figure 6 illustrates these relationships as heatmaps of subtype-specific Pearson correlations between METABRIC, TCGA, and GSE25066. Each subtype forms a block of uniformly high correlation between cohorts, indicating that a Basal-like tumor’s expression pattern in a microarray dataset is statistically almost identical to a Basal-like tumor’s pattern in an RNA-seq dataset, and similarly for other subtypes. Notably, cross-platform concordance was lowest for Luminal A (still > 0.8), which is expected given greater biological diversity among ER-positive tumors and the noted microarray intensity compression. In contrast, Basal-like and Luminal B profiles were virtually superimposable across datasets (r > 0.93 in all comparisons), underscoring that the proliferative gene module defining these subtypes is captured robustly by the 40-gene panel regardless of platform. These findings confirm that the core-PAM50 preserves the canonical transcriptional architecture of the intrinsic subtypes in a platform-independent manner. [Figure 6 placeholder] Principal component analysis (PCA) of subtype structure. To visualize the global structure of the 40-gene expression space, we performed principal component analysis (PCA) within each cohort using the standardized core-PAM50 expression matrix. In all datasets, the first principal component (PC1) captured the dominant biological axis contrasting proliferative (Basal-like and Luminal B) versus hormone receptor–driven (Luminal A and Normal-like) phenotypes. In the RNA-seq dataset (TCGA), PC1 accounted for 46.0% of variance and clearly separated Basal-like from Luminal tumors, with HER2-enriched cases in intermediate positions and Normal-like samples near the origin. In GSE25066, a similar pattern was observed (PC1 = 37.1%, PC2 = 16.5%), indicating robust subtype separation despite technological differences. In METABRIC, PC1 and PC2 explained 11.4% and 8.9% of variance, respectively, with greater overlap among subtypes—an expected consequence of microarray signal compression and a higher proportion of Normal-like assignments. Nevertheless, even in METABRIC, Basal-like and Luminal B samples trended along the positive PC1 direction (high MKI67, CCNB1, UBE2C), whereas Luminal A and Normal-like tumors occupied the opposite side (high ESR1, FOXA1, BCL2). As shown in Fig. 7 , PC1–PC2 plots for representative cohorts illustrate consistent spatial segregation of Basal and Luminal tumors across datasets, with HER2-enriched and Normal-like samples occupying intermediate or diffuse regions. Together with the subtype classification and correlation analyses, these findings demonstrate that the 40-gene core-PAM50 panel (i) faithfully reproduces intrinsic subtype assignments, (ii) yields cohort-specific subtype distributions aligned with known clinical and technical contexts, and (iii) preserves the canonical transcriptional geometry of breast cancer subtypes across platforms—confirming that the reduced gene set retains the full molecular taxonomy of the original PAM50. [Figure 7 placeholder] Discussion Cross-platform robustness and reproducibility: This study demonstrates that a carefully harmonized 40-gene subset of the PAM50 signature can yield a prognostic model that generalizes across different genomic technologies. We achieved a unified analysis of four large cohorts (over 6,800 samples) spanning two microarray platforms and RNA-seq, something that is often impeded by batch effects and annotation discrepancies [15]. A cornerstone of our approach was enforcing consistent gene nomenclature (up-to-date HGNC symbols) and standardized within-cohort normalization, which eliminated many technical biases. The outcome was a cross-platform model whose signal was largely platform-independent: we observed highly correlated subtype centroids across cohorts (r ~ 0.9) and comparable prognostic performance metrics in validation. In contrast, using un-harmonized data would likely have failed – even identical gene signatures can perform poorly when applied to raw data from another platform due to calibration drift or missing genes. By solving those issues upfront, the core-PAM50 preserves the “common denominator” of the expression patterns. In practice, this means a researcher or clinician could apply our 40-gene model to a tumor profiled by microarray or by RNA-seq (or potentially even by qPCR or NanoString) and expect a similar risk stratification. This kind of platform-agnostic assay is a step toward universally available genomic tests. Cohort differences and validation context: The disparate validation results in TCGA and GSE25066 highlight the importance of clinical context in evaluating prognostic models. TCGA patients had predominantly good outcomes at the available follow-up (many Luminal A tumors, treated with modern therapy, and censored alive within 5 years) – a scenario where demonstrating risk discrimination is inherently difficult. The model’s lack of significant separation in TCGA does not necessarily indicate a failure of the gene signature; rather, it reflects a low event rate and limited follow-up window. In a real-world sense, applying a prognostic model to an indolent population will yield attenuated performance even if the model is valid. By contrast, GSE25066, enriched for aggressive tumors (mostly Basal-like, high-grade disease) and observing many relapse events, displayed a strong risk stratification. Notably, the GSE25066 endpoint was relapse-free survival after neoadjuvant chemotherapy, which might amplify differences: patients with highly proliferative tumors remained at risk despite chemotherapy, whereas those with low-risk luminal biology fared better. These differences emphasize that prognostic effect sizes are context-dependent. A single model might appear to “work” better in one cohort than another not due to any flaw in the model, but because of differences in follow-up duration, treatment, and baseline risk. Our pooled analysis suggested that, on average, the core-PAM50 retained significant prognostic value externally (combined HR ~ 1.5 for high vs low risk). Still, a cautious approach is to tailor expectations of model performance to the scenario – for instance, one would not expect a dramatic split in a low-event-rate population. Comparison with existing signatures: The core-PAM50 model’s performance falls in the same range as established breast cancer gene signatures. In METABRIC, we achieved a training C-index of ~ 0.58, and in external validation (GSE25066) a time-dependent AUC ~ 0.60–0.63 at 5 years. This is comparable to reported values for PAM50’s Risk-Of-Recurrence (ROR) score and other assays: for example, Oncotype DX (21-gene Recurrence Score) typically yields C-index ~ 0.6 in ER-positive cohorts, and MammaPrint (70-gene signature) around 0.62–0.65, with improvements to ~ 0.70 when clinical factors like tumor size or nodal status are added [6]. Our results mirror those benchmarks. Importantly, we did not include standard clinical variables in our model to isolate the gene signature’s performance and maximize generalizability across cohorts (since detailed clinical covariates were not uniformly available for all datasets). In practice, an integrated model that combines core-PAM50 with tumor stage, grade, etc., would likely boost predictive accuracy (prior studies have shown gene signatures + clinical factors can raise C-index by ~ 0.05–0.1 [6]). However, our focus was to create a purely molecular metric that could be retrospectively evaluated in any transcriptomic dataset. The trade-off is a moderate discriminative power on its own (~ 0.6 C-index), but that is an inherent limitation of gene-only prognostic indices and consistent with the performance of much costlier assays. Given its transparency and portability, the core-PAM50’s prognostic ability is quite encouraging. Continuous risk vs subtype categories: A noteworthy feature of our model is that it produces a continuous risk score, even though it derives from subtype-related genes. This numeric risk has advantages in clinical decision-making – it preserves more information than coarse subtype labels and can be translated to absolute risk probabilities. Clinicians often prefer simple categories (e.g. “low risk” or “high risk”), but those can be obtained by setting thresholds on the continuous score as needed. In fact, flexibility in thresholding is a strength: different cutoff values could be chosen depending on the clinical scenario (for example, a higher risk threshold for recommending chemotherapy in lower-stage disease, and a lower threshold for deciding on extended endocrine therapy). Our analysis showed that dichotomizing the score into tertiles provided a meaningful stratification; other cut-points could be calibrated to specific interventions. The underlying continuous nature also means the model can be used in multifactorial risk calculations (e.g. combining with clinical nomograms) without losing granularity. Technical considerations – Normal-like subtype and assay quality: One unexpected result was the extremely high rate of Normal-like classifications in METABRIC. This likely reflects issues of tumor purity or assay quality in that older dataset – samples with low tumor cellularity or RNA degradation tend to exhibit expression profiles that resemble normal breast tissue, thus getting labeled “Normal-like” by the algorithm. This is a known phenomenon [16]. In our study, those Normal-like METABRIC cases were essentially equivalent to indolent Luminal A tumors (they had low proliferation and excellent survival), so the prognostic model wasn’t adversely affected (it correctly assigned them low risk). However, it raises a practical point: if an assay yields a subtype call of Normal-like in a context where that is unexpected, it could indicate a technical issue (e.g. a largely stromal sample). Future development of the core-PAM50 test could incorporate a quality flag – for instance, if no subtype correlation exceeds a low threshold, reporting the result as “unclassified, possibly low tumor content” might be prudent. In our data, only 3 samples met such a criterion (essentially no signal), but dozens of METABRIC tumors had moderate correlations to multiple centroids and ended up as Normal-like. In a prospective setting, one might consider those as Luminal A for practical purposes or require a pathology review. We handled this by acknowledging the bias rather than forcing a reclassification, but it is an area for further refinement (e.g. developing a pre-test tumor purity assessment). Another technical consideration is the absence of immune genes in the core-PAM50. Recent research highlights immune infiltration (tumor-infiltrating lymphocytes and related gene signatures) as an important prognostic factor, especially in triple-negative (Basal-like) breast cancers [9]. The PAM50 focuses on tumor-intrinsic genes and includes few direct immune markers. By trimming to 40 genes, we may have excluded some genes that, while not core to subtype classification, could carry immune/prognostic information (for instance, if any of the dropped PAM50 genes had immune relevance, though most were basal cytokeratins or lesser-expressed genes). Augmenting the core-PAM50 with an immune module could potentially improve prognostic power for certain patients. Indeed, one could envision a combined model that adds a TILs gene score to better stratify high-risk basal tumors. That said, adding genes would sacrifice some simplicity and the guarantee of cross-platform detectability. Our aim was a lean panel that is broadly applicable; additional features could be explored in future work, but the benefit would need to justify the complexity. Clinical implications and future directions: We have shown that the core-PAM50 model encapsulates the essence of the intrinsic subtype system and provides a continuous risk metric that is validated across diverse cohorts. The emphasis of this work is on reproducibility and transparency. All data preprocessing and modeling steps were documented and adhere to reproducible research principles, which is often not the case for commercial assays. This means others can audit, reuse, or adapt our pipeline – an important advantage in the era of open science and evolving genomic data. We envision several practical applications and extensions of this model. First, it can be used retrospectively on existing gene expression datasets (from clinical trials or real-world cohorts) to generate a standardized risk score, facilitating meta-analyses or comparisons across studies. Second, it provides a template for laboratories to implement a cost-effective prognostic assay: since it uses only 40 genes and straightforward calculations, it could be deployed on platforms like quantitative PCR or targeted RNA-seq with relative ease, potentially expanding access in resource-limited settings. Third, prospective validation in a clinical trial would be valuable – for example, testing whether this model can identify patients who benefit from certain therapies (e.g. need for chemotherapy in what would otherwise be an intermediate clinical risk group). The model’s performance for late relapse (beyond 5 years) in ER + disease is another area to investigate; our data hinted that core-PAM50 retains prognostic information up to 10 years (similar to PAM50 ROR which differentiates Luminal A vs B long-term [8]), but specialized assays like Breast Cancer Index target that scenario and could be compared. In a broader sense, our study exemplifies how reducing a complex genomic signature to its most robust elements, combined with rigorous data harmonization, can enhance real-world applicability. We effectively bridge the gap between large-scale “discovery” datasets and a clinical-grade test. The core-PAM50 model achieved a similar prognostic performance to more cumbersome signatures, while using ~ 30% fewer genes and being openly reproducible. This parsimony can translate to lower costs, simpler assay design, and fewer regulatory hurdles (as there are fewer biomarkers to validate). Moreover, by avoiding platform-specific artifacts, it increases confidence that results from different hospitals or studies can be directly compared. Conclusion In conclusion, we developed a reduced 40-gene PAM50-derived signature and demonstrated its technical and prognostic performance across multiple breast cancer cohorts and genomic platforms. The core-PAM50 model retains the essential biological signal of the intrinsic subtypes and provides a continuous risk score that is validated in independent microarray and RNA-seq datasets. Emphasizing reproducibility and transparency, we implemented a fully documented analysis pipeline, ensuring that results can be audited and reproduced by other researchers. This reproducible framework adds significant value compared to many proprietary tests, as it adheres to FAIR data principles and modern reporting standards [25]. The core-PAM50 panel offers a practical path toward a cost-effective, high-stability genomic assay for breast cancer prognosis. With roughly one-third fewer genes than the canonical PAM50 and avoidance of platform-specific pitfalls, it can reduce experimental complexity and validation burden by ~ 30–40% while maintaining prognostic accuracy. Our multi-cohort validation supports its potential generalizability: risk stratification was consistent in principle across diverse cohorts, and the molecular subtyping remained intact. Moving forward, integration of the core-PAM50 model into clinical research is warranted. For instance, it could be prospectively tested in translational trials or applied to re-analyze archived trial samples to glean additional prognostic insights. Further refinements, such as incorporating complementary features (immune markers or clinical covariates) and addressing classification nuances (e.g. the Normal-like issue), may enhance the model’s utility, but even in its present form the evidence strongly supports its robustness and clinical relevance. Ultimately, the core-PAM50 signature represents a parsimonious yet biologically informed prognostic tool that brings large-scale genomic knowledge closer to routine clinical practice. It preserves the rich information content of the PAM50 intrinsic subtypes – a taxonomy deeply ingrained in breast cancer biology – and repackages it into a lean, reproducible assay framework suitable for global oncology use. With further validation, this model could help optimize personalized prognostication and therapy decisions while simplifying the technology required, thereby making genomic risk assessment more accessible to patients worldwide. Abbreviations AUC – Area under the curve BC – Breast cancer CI – Confidence interval CV – Cross-validation DCA – Decision curve analysis DRFS – Disease-free recurrence/death-free survival ER – Estrogen receptor FFPE – Formalin-fixed paraffin-embedded GEO – Gene Expression Omnibus GSE – Gene Expression Series HGNC – HUGO Gene Nomenclature Committee HER2 – Human epidermal growth factor receptor 2 HR – Hazard ratio ID – Identifier IHC – Immunohistochemistry KM – Kaplan–Meier LASSO – Least absolute shrinkage and selection operator MD5 – Message Digest 5 (cryptographic hash function) METABRIC – Molecular Taxonomy of Breast Cancer International Consortium mRNA – Messenger RNA OS – Overall survival PAM50 – Prediction Analysis of Microarray 50 PCA – Principal component analysis PR – Progesterone receptor QC – Quality control RNA-seq – RNA sequencing RSF – Random survival forest ROR – Risk of recurrence SD – Standard deviation SE – Standard error TCGA – The Cancer Genome Atlas TRIPOD – Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis VIMP – Variable importance (in Random Survival Forests) z-score – Standardized gene expression value (mean 0, SD 1) Declarations Ethics approval and consent to participate: not applicable. Consent for publication: not applicable. Availability of data and material: data supporting the results are available upon request to the corresponding author. Competing interests: the authors declare that they have no competing interests. Funding: this research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Authors’ contributions: RNB conceived and designed the study, performed data analysis, and wrote the manuscript. JBS supervised the study and contributed to manuscript revision. Both authors read and approved the final manuscript. Acknowledgements: we thank the Biostatistics Department of the University of Brasilia for technical assistance. References Sung H, Ferlay J, Siegel RL, 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):209–249. Perou CM, Sørlie T, Eisen MB, et al. Molecular portraits of human breast tumours. Nature. 2000;406(6797):747–752. Yersal O, Barutca S. Biological subtypes of breast cancer: Prognostic and therapeutic implications. World J Clin Oncol. 2014;5(3):412–424. Ignatiadis M, Sotiriou C. Luminal breast cancer: from biology to treatment. Nat Rev Clin Oncol. 2013;10(9):494–506. Parker JS, Mullins M, Cheang MCU, et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. J Clin Oncol. 2009;27(8):1160–1167. Nielsen TO, Parker JS, Leung S, et al. Comparison of PAM50 intrinsic subtyping with immunohistochemistry and clinical prognostic factors. Clin Cancer Res. 2010;16(21):5222–5232. Ali HR, Rueda OM, Chin SF, et al. Genome-driven integrated classification of breast cancer validated in 11,000 tumours across 10 independent datasets. Nat Commun. 2020;11(1):4865. Rueda OM, Sammut SJ, Seoane JA, et al. Dynamics of breast-cancer relapse reveal late-recurring ER-positive genomic subgroups. Nature. 2019;567(7748):399–404. Luen SJ, Salgado R, Fox S, et al. Tumour-infiltrating lymphocytes and prognosis in different subtypes of breast cancer: a pooled analysis. Lancet Oncol. 2017;18(1):87–96. Curtis C, Shah SP, Chin SF, et al. The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups. Nature. 2012;486(7403):346–352. Cancer Genome Atlas Network. Comprehensive molecular portraits of human breast tumours. Nature. 2012;490(7418):61–70. Hatzis C, Pusztai L, Valero V, et al. A genomic predictor of response and survival following taxane-anthracycline chemotherapy for invasive breast cancer. J Clin Oncol. 2011;29(17):2433–2440. Brueffer C, Darai E, Segersten U, et al. The Swedish SCAN-B initiative: a large-scale multicenter study for breast cancer. Clin Cancer Res. 2019;25(19):5859–5872. Lee JS, Jung J, Kim Y, et al. Integrated multi-omics analysis of breast cancer reveals genomic correlates of clinical outcomes. Nat Med. 2023;29(1):49–62. Leek JT, Scharpf RB, Bravo HC, et al. Tackling the critical impact of batch effects in high-throughput data. Nat Rev Genet. 2010;11(10):733–739. Ali HR, Lundberg A, Ohnstad HO, et al. Global prognostic and predictive value of PAM50 across molecular and clinical subgroups: a meta-analysis. Breast Cancer Res. 2023;25(1):57. Li Y, Jiang T, Zhang B, et al. A 40-gene expression signature predicts clinical outcomes in breast cancer. EBioMedicine. 2022;82:104136. Li C, et al. Reducing feature dimensionality of genomic classifiers while preserving predictive performance. Brief Bioinform. 2021;22(6):bbab358. Yates B, Braschi B, Gray KA, et al. Genenames.org: the HGNC and VGNC resources. Nucleic Acids Res. 2020;48(D1):D649–D655. Tibshirani R. The lasso method for variable selection in the Cox model. Stat Med. 1997;16(4):385–395. Ishwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random survival forests for R. Bioinformatics. 2008;24(11):1363–1371. Blanche P, Dartigues JF, Jacqmin-Gadda H. Estimating time-dependent AUC for censored event times. Stat Med. 2013;32(30):5381–5397. Kattan MW, Vickers AJ, Steyerberg EW. Evaluating the calibration of predictive models in survival analysis. Stat Methods Med Res. 2020;29(2):515–536. Vickers AJ, Van Calster B, Steyerberg EW. Net benefit approaches to the evaluation of prediction models. BMJ. 2019;365:l2373. Steyerberg EW, Collins GS, Dauholuk A, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD-2022). Nat Med. 2023;29(3):591–599. Tables Table 1 and 2 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1METABRICModel.docx Table2Performancebycohort.docx SuppFigS1Top20VarImpRSFMETABRIC.pdf SuppFigS2DCAgrid.pdf SupplFigS3SubtypeDistribution.pdf SuppTableS1PAM50corePAM50.docx SuppTableS2Coefficients.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 23 Feb, 2026 Reviews received at journal 05 Feb, 2026 Reviews received at journal 16 Jan, 2026 Reviewers agreed at journal 10 Jan, 2026 Reviewers agreed at journal 05 Jan, 2026 Reviewers agreed at journal 18 Dec, 2025 Reviewers invited by journal 17 Dec, 2025 Editor assigned by journal 12 Nov, 2025 Submission checks completed at journal 11 Nov, 2025 First submitted to journal 06 Nov, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8048795","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":562535522,"identity":"26b0333b-9f36-4be8-9888-9cbc5b9dca29","order_by":0,"name":"Rafael de Negreiros Botan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4UlEQVRIiWNgGAWjYLCCBwwMPECKEUgz8PARpSUBpIWNgdkApIWNWC0MQC1sEhCaAOCXbn74IKGiVkZ+fvOzyq85djJA2x4+uoFHi+ScY8YGCWeO8xgcYzO7LbstGegwNmPjHDxaDG7ksEkkth3jMWBjMLstuY0ZqIWHTZooLfJt7N+KJbfVE62lhofhGI8Z48dthwlrkZyRBvLLAaBfcoqlGbcd52FjJuAXfonkhw8+VNTZyzcf3/jx57Zqe3725oeP8WmBgsNgkpkHTBJWDgJ1YJLxB3GqR8EoGAWjYIQBALIgQV7nHTCnAAAAAElFTkSuQmCC","orcid":"","institution":"University of Brasília","correspondingAuthor":true,"prefix":"","firstName":"Rafael","middleName":"de Negreiros","lastName":"Botan","suffix":""},{"id":562535523,"identity":"8237f8ea-6363-4e61-9de9-16403c848234","order_by":1,"name":"Joao Batista de Sousa","email":"","orcid":"","institution":"University of Brasília","correspondingAuthor":false,"prefix":"","firstName":"Joao","middleName":"Batista","lastName":"de Sousa","suffix":""}],"badges":[],"createdAt":"2025-11-06 13:53:53","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8048795/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8048795/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98634666,"identity":"6a075223-3a27-4546-bcff-1273c4ae4b46","added_by":"auto","created_at":"2025-12-19 17:25:46","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":55908,"visible":true,"origin":"","legend":"","description":"","filename":"submittedv3CrossPlatformReproducibleModelingofBreastCancerPrognosisUsingtheCorePAM50GeneSignature.docx","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/86da9ae6a64daf4df3c28e57.docx"},{"id":98635514,"identity":"fc806e35-e0ed-4398-9e90-cd5bc46d8e64","added_by":"auto","created_at":"2025-12-19 17:26:15","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14511,"visible":true,"origin":"","legend":"","description":"","filename":"Table1METABRICModel.docx","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/4cdfa7849552370917ef0f7f.docx"},{"id":98634772,"identity":"823db0ca-9293-4d8e-aaa9-0b52f3de10bd","added_by":"auto","created_at":"2025-12-19 17:25:52","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14434,"visible":true,"origin":"","legend":"","description":"","filename":"Table2Performancebycohort.docx","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/8823b840323fbb42fdebb191.docx"},{"id":98635022,"identity":"ff71ec8b-3204-428c-bddb-67c9850c3eee","added_by":"auto","created_at":"2025-12-19 17:25:59","extension":"json","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5324,"visible":true,"origin":"","legend":"","description":"","filename":"c2980ad59b3d41a2bf47f093c5dc8c5f.json","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/91e05a512f190b7b3f5fecd5.json"},{"id":98634526,"identity":"0527043f-e858-4a15-9d71-f88ac6d37f67","added_by":"auto","created_at":"2025-12-19 17:25:43","extension":"pdf","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":143199,"visible":true,"origin":"","legend":"","description":"","filename":"SuppFigS1Top20VarImpRSFMETABRIC.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/2732092714c6689176621f6e.pdf"},{"id":98634731,"identity":"71dc70d7-6d69-4ae1-907a-91452846f772","added_by":"auto","created_at":"2025-12-19 17:25:50","extension":"pdf","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":264577,"visible":true,"origin":"","legend":"","description":"","filename":"SuppFigS2DCAgrid.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/43a0209456bce7d7e3b1d5f6.pdf"},{"id":98635464,"identity":"9c45d888-2bb4-4217-8cad-a2e9dac9dcb6","added_by":"auto","created_at":"2025-12-19 17:26:13","extension":"docx","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15254,"visible":true,"origin":"","legend":"","description":"","filename":"SuppTableS1PAM50corePAM50.docx","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/744f9f8e5e39fe8a05c0ac0a.docx"},{"id":98634532,"identity":"78dff886-6a19-4d17-b75f-ba08e40a7f9c","added_by":"auto","created_at":"2025-12-19 17:25:43","extension":"docx","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15481,"visible":true,"origin":"","legend":"","description":"","filename":"SuppTableS2Coefficients.docx","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/dece19358cc78b17c5f2d519.docx"},{"id":98635161,"identity":"0edba7f0-eaec-4870-a2f1-39205fcacf15","added_by":"auto","created_at":"2025-12-19 17:26:03","extension":"pdf","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":466084,"visible":true,"origin":"","legend":"","description":"","filename":"SupplFigS3SubtypeDistribution.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/bf5682d08bf7db71b115281c.pdf"},{"id":98635154,"identity":"058f86bf-0528-409b-8a03-d7918b4e2877","added_by":"auto","created_at":"2025-12-19 17:26:03","extension":"xml","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":105410,"visible":true,"origin":"","legend":"","description":"","filename":"c2980ad59b3d41a2bf47f093c5dc8c5f1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/55e82431663ca8a9fdf81499.xml"},{"id":98634726,"identity":"70c8632b-07dd-42ac-bfe3-13f8fb19c450","added_by":"auto","created_at":"2025-12-19 17:25:49","extension":"pdf","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":824915,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1Workflow.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/6fc88d3228ea52ae6e5a6d4a.pdf"},{"id":98635410,"identity":"eb19d7da-655a-45d9-a5f4-a89ed2f6a791","added_by":"auto","created_at":"2025-12-19 17:26:12","extension":"pdf","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":501204,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2A2BLASSOCV.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/862078a0923c55efa9f3901a.pdf"},{"id":98774920,"identity":"ae05de2a-7001-4c6f-956f-af655de32041","added_by":"auto","created_at":"2025-12-22 12:16:56","extension":"pdf","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":496362,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3A3BKM.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/ae38fbc4c07f1096f3e7415b.pdf"},{"id":98634743,"identity":"82dced2d-b834-412f-a433-5a458fb6e81e","added_by":"auto","created_at":"2025-12-19 17:25:50","extension":"pdf","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":535413,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4ForestPlotLOWvsHIGH.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/f8af3d615380f5d4c095aaf5.pdf"},{"id":98774968,"identity":"d678fdb5-7aa4-40d3-a760-aee5328aeb23","added_by":"auto","created_at":"2025-12-22 12:17:38","extension":"pdf","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":790142,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5Calibration.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/df614a97af910254db85aa86.pdf"},{"id":98635492,"identity":"1c7cb6cc-89aa-4e2e-a07d-53e2a44a49bf","added_by":"auto","created_at":"2025-12-19 17:26:14","extension":"pdf","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":525160,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6CorelationHeatmap.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/45a718ba7fb2d32411a2621a.pdf"},{"id":98635465,"identity":"cd246125-f1a6-4365-93b4-8d63f0e80633","added_by":"auto","created_at":"2025-12-19 17:26:13","extension":"pdf","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3756970,"visible":true,"origin":"","legend":"","description":"","filename":"Figure7Workflow.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/e7e989847d855ab0d22b34b8.pdf"},{"id":98635051,"identity":"152f2548-96d7-464c-a257-59b690d533b9","added_by":"auto","created_at":"2025-12-19 17:26:01","extension":"xml","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":104053,"visible":true,"origin":"","legend":"","description":"","filename":"c2980ad59b3d41a2bf47f093c5dc8c5f1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/85b632e64348cbaa31511b6f.xml"},{"id":98634525,"identity":"0cf51414-ad4d-4f6d-8ac5-4c5902004490","added_by":"auto","created_at":"2025-12-19 17:25:43","extension":"html","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":118064,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/50f0b3e5e740cd66e6751c54.html"},{"id":98635490,"identity":"e53809d3-6a73-4255-90aa-f29b7083af66","added_by":"auto","created_at":"2025-12-19 17:26:14","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":378076,"visible":true,"origin":"","legend":"\u003cp\u003eAnalytical workflow for model development and validation\u003c/p\u003e\n\u003cp\u003eOverview of the analytical pipeline used to develop and validate the \u003cem\u003ecore-PAM50\u003c/em\u003emodel. Raw transcriptomic and clinical data from three public cohorts (METABRIC, TCGA-BRCA, and GSE25066) were processed through a unified workflow comprising quality control, HGNC-based gene harmonization, and intra-cohort z-score normalization. A cross-platform audit and gene intersection identified 11,544 genes shared among all datasets, from which 40 consistently detected genes defined the \u003cem\u003ecore-PAM50\u003c/em\u003e panel. Models were trained in the METABRIC cohort using Cox-LASSO and Random Survival Forests (RSF) and externally validated in TCGA (overall survival) and GSE25066 (disease-free survival). Subsequent molecular validation included intrinsic subtype assignment, correlation analysis, and principal component analysis (PCA).\u003cbr\u003e\nFinal outputs comprised subtype predictions, risk scores, and cross-cohort reproducibility metrics.\u003c/p\u003e","description":"","filename":"11.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/a81e5df840402f6e75bcdd09.jpg"},{"id":98635451,"identity":"7a9aa1b5-a1b4-4816-abfd-adc2b74323c5","added_by":"auto","created_at":"2025-12-19 17:26:13","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":165465,"visible":true,"origin":"","legend":"\u003cp\u003eLASSO-Cox model development on the METABRIC training cohort. \u003cem\u003e(A)\u003c/em\u003e LASSO coefficient paths for all 40 core-PAM50 genes. As the regularization stringency increases (moving left on the x-axis, higher log(λ)), most gene coefficients shrink to zero. At the optimal λ (vertical dashed line), 20 genes remain with non-zero coefficients, defining the parsimonious prognostic model. \u003cem\u003e(B)\u003c/em\u003eCross-validated deviance profile (10-fold CV) as a function of λ. The red dot indicates the λ_min chosen for the final model, which lies within a broad minimum of the deviance curve, suggesting minimal overfitting. The error bars show ±1 standard error. These panels illustrate that the LASSO identified a sparse set of predictive genes and that this model is stable and well-tuned.\u003c/p\u003e","description":"","filename":"12.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/7b78bcacbab3615a417c82c2.jpg"},{"id":98635233,"identity":"667cc696-0d52-499a-bad6-1afe408dc8db","added_by":"auto","created_at":"2025-12-19 17:26:10","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":131892,"visible":true,"origin":"","legend":"\u003cp\u003eExternal validation of the 20-gene \u003cem\u003ecore-PAM50\u003c/em\u003e Cox model in independent cohorts.\u003cbr\u003e\n(A) Kaplan–Meier overall-survival curves for TCGA-BRCA, stratified by predicted risk tertiles (High, Intermediate, Low).\u003cbr\u003e\nNo significant separation was observed (HR = 0.89 [0.67–1.19]; p = 0.45), consistent with the limited follow-up and low event rate in this cohort.\u003cbr\u003e\n(B) Kaplan–Meier disease-free-survival curves for GSE25066 (neoadjuvant chemotherapy cohort).\u003cbr\u003e\nHigh-Risk patients exhibited significantly poorer outcomes than Low-Risk patients (HR = 0.50 [0.35–0.73]; p \u0026lt; 0.001), corresponding to an inverse two-fold increase in hazard.\u003cbr\u003e\nThese findings confirm external prognostic reproducibility of the \u003cem\u003ecore-PAM50\u003c/em\u003emodel across distinct clinical contexts.\u003c/p\u003e","description":"","filename":"13.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/328fd9ee9ec7a1964a07a30b.jpg"},{"id":98635155,"identity":"2c82ad06-5150-4dfe-9adc-3f826e8fc8c7","added_by":"auto","created_at":"2025-12-19 17:26:03","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":117611,"visible":true,"origin":"","legend":"\u003cp\u003eForest plot of hazard ratios for Low versus High risk groups across independent cohorts.\u003cbr\u003e\nHorizontal bars represent 95 % confidence intervals; vertical dashed line indicates the null value (HR = 1).\u003cbr\u003e\nThe \u003cem\u003ecore-PAM50\u003c/em\u003e Cox model showed a strong prognostic effect in METABRIC and GSE25066, with Low-risk patients exhibiting roughly half the hazard observed in High-risk patients, while TCGA showed no significant difference.\u003cbr\u003e\nThe pooled meta-analytic estimate (HR = 0.68 [0.39–1.20]) supports a consistent protective direction for the Low-risk group across datasets and platforms.\u003c/p\u003e","description":"","filename":"14.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/a3f0d525d0cc71b570c98938.jpg"},{"id":98775352,"identity":"ffc31ee4-69bb-46cc-8069-bb4cd13cb44b","added_by":"auto","created_at":"2025-12-22 12:19:34","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":146321,"visible":true,"origin":"","legend":"\u003cp\u003eCross-platform calibration of 5-year survival predictions. Calibration curves for 5-year outcome probability in METABRIC (circles), GSE25066 (triangles), and TCGA (squares). The x-axis is the Cox model–predicted probability of 5-year survival for groups of patients (binned by risk score deciles), and the y-axis is the observed 5-year survival fraction in each bin. The diagonal line represents perfect calibration (predicted = observed). METABRIC (blue) shows excellent calibration (points near the line). GSE25066 (red) is also well-calibrated at 5 years, indicating the model’s absolute risk estimates translate accurately to a neoadjuvant-treated cohort. TCGA (green) shows moderate miscalibration with observed outcomes better than predicted for high-risk scores (points above the line at lower survival probabilities), consistent with risk overestimation in a short follow-up scenario. Error bars denote 95% CIs for observed survival in each bin.\u003c/p\u003e","description":"","filename":"15.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/18c3725069dcccb4fc6499bd.jpg"},{"id":98635516,"identity":"9c5a5a16-26f0-4eaf-a5ac-f6cf903a6650","added_by":"auto","created_at":"2025-12-19 17:26:15","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":212320,"visible":true,"origin":"","legend":"\u003cp\u003eCross-cohort subtype expression concordance for the core-PAM50 panel.\u003cbr\u003e\nHeatmaps show Pearson correlations (r) between mean subtype-specific expression profiles (“centroids”) derived from METABRIC, TCGA, and GSE25066 for each intrinsic subtype (Basal-like, HER2-enriched, Luminal A, Luminal B, Normal-like).\u003cbr\u003e\nEach block represents cross-cohort correlations for one subtype.\u003cbr\u003e\nUniformly high correlations (r \u0026gt; 0.8 in all cases) demonstrate strong reproducibility of subtype-defining transcriptional patterns across independent cohorts and technologies, confirming that the 40-gene \u003cem\u003ecore-PAM50\u003c/em\u003ecaptures the intrinsic-subtype architecture in a platform-independent manner.\u003c/p\u003e","description":"","filename":"16.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/239276cc478947ff7b195199.jpg"},{"id":98634624,"identity":"0a219645-0a85-4997-9177-4141396b825a","added_by":"auto","created_at":"2025-12-19 17:25:45","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":231224,"visible":true,"origin":"","legend":"\u003cp\u003ePrincipal component analysis (PCA) of the core-PAM50 expression space across cohorts.\u003cbr\u003e\nScatterplots show the first two principal components (PC1, PC2) of standardized expression for the 40 \u003cem\u003ecore-PAM50\u003c/em\u003e genes in GSE25066 (left), METABRIC (center), and TCGA (right).\u003cbr\u003e\nEach point represents a tumor sample, colored by its assigned intrinsic subtype.\u003cbr\u003e\nEllipses indicate 95 % confidence regions per subtype.\u003cbr\u003e\nIn all cohorts, PC1 captures the dominant “proliferation vs. luminal differentiation” axis separating Basal-like from Luminal tumors, while HER2-enriched and Normal-like cases occupy intermediate or diffuse regions.\u003cbr\u003e\nDespite differences in technology and signal compression, subtype topology is preserved across datasets, confirming the cross-platform reproducibility of the \u003cem\u003ecore-PAM50\u003c/em\u003e molecular structure.\u003c/p\u003e","description":"","filename":"17.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/c85f21e456687e8cb2e017a0.jpg"},{"id":98782739,"identity":"3229ee68-9797-4416-9da1-f0a0e6dc2b30","added_by":"auto","created_at":"2025-12-22 12:40:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1971963,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/b040733d-933c-43f2-955c-1d77e1430a34.pdf"},{"id":98635255,"identity":"ad8b17e6-7498-4812-9b17-4154b225c1a1","added_by":"auto","created_at":"2025-12-19 17:26:11","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":14511,"visible":true,"origin":"","legend":"","description":"","filename":"Table1METABRICModel.docx","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/90e34b6a8234651e81ede5c3.docx"},{"id":98635160,"identity":"2630daa0-dd35-4252-a6ad-4c09c439bd53","added_by":"auto","created_at":"2025-12-19 17:26:03","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":14434,"visible":true,"origin":"","legend":"","description":"","filename":"Table2Performancebycohort.docx","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/f8e0075d378805b0d7e4a124.docx"},{"id":98634620,"identity":"5571eb02-bbf7-45bf-bdbd-e4f23c66ce9b","added_by":"auto","created_at":"2025-12-19 17:25:45","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":143199,"visible":true,"origin":"","legend":"","description":"","filename":"SuppFigS1Top20VarImpRSFMETABRIC.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/253f555cb1aea1458cfa6278.pdf"},{"id":98634756,"identity":"95174081-4003-4bc5-8bff-a449ecf65273","added_by":"auto","created_at":"2025-12-19 17:25:51","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":264577,"visible":true,"origin":"","legend":"","description":"","filename":"SuppFigS2DCAgrid.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/51d8f3e5d9e11321d80995b4.pdf"},{"id":98635215,"identity":"14ec29eb-dd82-4f39-a46a-2945100f9e27","added_by":"auto","created_at":"2025-12-19 17:26:04","extension":"pdf","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":466084,"visible":true,"origin":"","legend":"","description":"","filename":"SupplFigS3SubtypeDistribution.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/0bdada3884fcf3d88c973676.pdf"},{"id":98635198,"identity":"92b7cee2-db16-4de0-8adb-59a02d2707f8","added_by":"auto","created_at":"2025-12-19 17:26:03","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":15254,"visible":true,"origin":"","legend":"","description":"","filename":"SuppTableS1PAM50corePAM50.docx","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/4603729cbc35816db4ee1200.docx"},{"id":98635459,"identity":"fd09f450-0191-4f85-a94e-dff470971baf","added_by":"auto","created_at":"2025-12-19 17:26:13","extension":"docx","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":15481,"visible":true,"origin":"","legend":"","description":"","filename":"SuppTableS2Coefficients.docx","url":"https://assets-eu.researchsquare.com/files/rs-8048795/v1/2433227b4722d3ae286a3088.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Cross-Platform Reproducible Modeling of Breast Cancer Prognosis Using the Core-PAM50 Gene Signature","fulltext":[{"header":"Background","content":"\u003cp\u003eBreast cancer is the most common malignancy in women and a leading cause of cancer mortality worldwide [1]. Despite advances in diagnosis and therapy, the biological heterogeneity of breast tumors remains a central challenge for prognosis prediction and treatment selection [2,3]. This heterogeneity arises from multiple molecular pathways of tumor progression \u0026ndash; spanning proliferative, hormonal, and immune-inflammatory axes \u0026ndash; which are not fully captured by traditional clinicopathological markers [4]. In this context, multigene expression signatures have emerged as precision tools to stratify recurrence risk and guide therapy. Among these, the PAM50 (Prediction Analysis of Microarray 50) gene signature has become one of the most widely validated, classifying breast tumors into five intrinsic molecular subtypes (Luminal A, Luminal B, HER2-enriched, Basal-like, and Normal-like) with well-recognized prognostic and therapeutic implications [5,6]. Subsequent studies have confirmed that the PAM50 signature captures reproducible transcriptional gradients correlated with clinical outcomes across diverse populations and experimental platforms [7\u0026ndash;9].\u003c/p\u003e \u003cp\u003eThe rise of large-scale genomic cohorts has further enabled multidimensional characterization of breast cancer. The Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) profiled over 2,000 primary tumors with integrated clinical, transcriptomic (microarray), and genomic data, inaugurating a new era of molecular stratification [10]. In parallel, The Cancer Genome Atlas Breast Cancer project (TCGA-BRCA) applied deep RNA sequencing (RNA-seq) to over 1,000 breast tumors, providing a multi-omic portrait and validating biological patterns observed in METABRIC [11]. A public cohort, GSE25066 (Affymetrix U133A microarrays from 508 patients treated with neoadjuvant chemotherapy), has been widely used for validation and translational studies [12\u0026ndash;14]. These datasets cover different clinical scenarios \u0026ndash; from early-stage, systemically untreated tumors to high-risk neoadjuvant trial specimens \u0026ndash; and serve as reference points for developing and testing molecular models. However, technological variability (microarray vs. RNA-seq), lack of uniform gene annotation, and expression scale discrepancies pose major barriers to direct data integration [15].\u003c/p\u003e \u003cp\u003eDespite its broad adoption, the canonical 50-gene PAM50 panel has practical limitations: some of the original genes are not reliably detected on all platforms, and issues like gene alias discrepancies, normalization differences, and batch effects can diminish inter-cohort reproducibility [16]. Moreover, implementing large multigene panels in clinical practice is costly and requires specialized laboratory infrastructure, limiting their universal adoption. Controlled dimensionality reduction \u0026ndash; i.e. selecting a stable, informative subset of genes \u0026ndash; has been proposed as a strategy to improve portability without significant loss of prognostic accuracy [17,18]. In this context, we developed the core-PAM50, a streamlined 40-gene subset of the PAM50 signature designed for cross-platform stability. This reduced panel includes only genes detectably expressed in all major datasets and preserves the essential \u0026ldquo;proliferative\u0026rdquo; and \u0026ldquo;hormonal\u0026rdquo; axes of the original signature. Using a unified bioinformatics pipeline, we harmonized multiple cohorts under a common nomenclature (HUGO gene symbols) and performed intra-cohort \u003cem\u003ez\u003c/em\u003e-score normalization, intersecting gene sets across cohorts to ensure directly comparable data [19]. The result is a reproducible multi-cohort dataset and a core gene panel amenable to integrated analysis.\u003c/p\u003e \u003cp\u003eTo model continuous risk of mortality from this core gene set, we employed two complementary approaches: (i) LASSO-penalized Cox proportional hazards regression (Cox-LASSO), which performs automatic variable selection and guards against overfitting in high-dimensional settings [20]; and (ii) Random Survival Forests (RSF), a non-parametric ensemble method that captures non-linear interactions among genes without assuming proportional hazards [21]. Model performance was evaluated using standard discrimination metrics (Harrell\u0026rsquo;s concordance index (C-index) and time-dependent area under the curve (AUC)), temporal calibration plots, and clinical decision curve analysis (DCA), providing a comprehensive assessment of prognostic accuracy and clinical utility [22\u0026ndash;24]. Importantly, the prognostic model was trained in one cohort and validated in others without any re-calibration of coefficients, preserving the strict independence of validation sets and avoiding optimistic bias. The METABRIC-derived model coefficients were applied directly to external cohorts (TCGA and GSE25066) with no post-hoc adjustment, a strategy pre-specified in the analysis protocol in line with contemporary guidelines for transparent prognostic modeling [25].\u003c/p\u003e \u003cp\u003eDespite the success of PAM50, no prior study has systematically integrated the three principal public breast cancer cohorts (METABRIC, TCGA, and GSE25066) under a unified gene-harmonization framework spanning both microarray and RNA-seq platforms. We hypothesize that a reduced, technically robust gene panel \u0026ndash; the core-PAM50 \u0026ndash; can maintain the molecular structure and prognostic power of the original PAM50, achieving the following: (i) reproduce the classic intrinsic subtype classifications with high cross-cohort concordance; (ii) provide consistent risk predictions for mortality and recurrence across independent patient cohorts; and (iii) demonstrate good calibration and tangible clinical net benefit in both microarray and RNA-seq contexts. Through this study, we aim to bridge the gap between expansive genomic signatures and practical, widely transferable prognostic tools by delivering a methodological framework that is both scalable (spanning multiple cohorts and platforms) and clinically reproducible.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData Cohorts and Harmonization\u003c/h2\u003e \u003cp\u003eThree public breast cancer cohorts complementary in size, technology, and clinical context were assembled for analysis: (1) METABRIC \u0026ndash; 2,173 primary breast tumors profiled by Illumina HT-12 v4 microarrays with detailed clinical and long-term survival data [10]; (2) TCGA-BRCA \u0026ndash; 1,098 invasive breast tumors analyzed by Illumina HiSeq RNA-seq, with corresponding clinical annotations [11] and (3) GSE25066\u0026ndash;508 tumor samples from a neoadjuvant chemotherapy trial (taxane\u0026ndash;anthracycline regimen) profiled on Affymetrix U133A microarrays (GPL96), including clinical follow-up for recurrence and survival [12]. Raw data (gene expression matrices and clinical tables) were obtained from the original repositories (cBioPortal for METABRIC, NCI Genomic Data Commons/UCSC Xena for TCGA, and NCBI GEO for the GSE study) and imported into R (v4.1.3) under a standardized directory structure. All source files were audited with cryptographic hashes (MD5) to ensure data integrity and full reproducibility of the pipeline.\u003c/p\u003e \u003cp\u003eEach cohort underwent a uniform curation and intra-cohort preprocessing workflow prior to integration. Fundamental clinical variables (age, sex, stage, ER/PR/HER2 status, vital status, survival time, etc.) were standardized via controlled vocabulary mappings (e.g. converting hormone receptor status to POS/NEG) and consistent coding of outcomes. The multi-step preprocessing included: (a) \u003cem\u003eData integrity checks\u003c/em\u003e: expression matrices were loaded and checked for completeness (no unexpected missing values) and correct dimensions. Clinical tables were merged with expression sample lists after cleaning sample identifiers (e.g. removing cohort-specific prefixes/suffixes and enforcing consistent alphanumeric IDs) to ensure one-to-one matching of tumor samples. In METABRIC, for example, this resolved minor discrepancies (e.g. MB-0001 vs MB.0001) and yielded 2,173 matched tumor profiles with clinical records. (b) \u003cem\u003eGene annotation harmonization\u003c/em\u003e: All gene identifiers were standardized to official HUGO Gene Nomenclature Committee (HGNC) symbols. For microarray data, probe IDs were mapped to gene symbols using platform annotation files (for GSE25066, GPL96 mappings were applied to convert 54k probe sets to 13,100 gene symbols). We applied an updated Entrez-to-HGNC dictionary [19] to correct obsolete gene symbols and aliases, achieving\u0026thinsp;\u0026gt;\u0026thinsp;99.9% successful mapping of genes in each dataset. (c) \u003cem\u003eCross-platform normalization\u003c/em\u003e: To enable direct comparability of expression levels across platforms, each dataset\u0026rsquo;s gene expression values were transformed to \u003cem\u003ez\u003c/em\u003e-scores within that cohort (gene-wise centering to mean 0 and scaling to unit standard deviation). This preserves relative expression patterns while removing differences in dynamic range between microarray intensities and RNA-seq counts [15].\u003c/p\u003e \u003cp\u003eAfter curation, we intersected the gene sets across cohorts. We specifically focused on the PAM50 gene set: of the 50 genes (48 unique gene symbols) in the original PAM50, 40 were consistently measured in all four datasets (the remaining genes were missing or unreliably detected in at least one platform) [16]. These 40 genes constitute the core-PAM50 panel. A detailed comparison of the full PAM50 gene list and the 40-gene core-PAM50 subset, including genes excluded and reasons for exclusion, is provided in \u003cem\u003eSupplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/em\u003e.\u003c/p\u003e \u003cp\u003eEach cohort\u0026rsquo;s expression matrix was filtered to these 40 genes, and expression values were mildly winsorized at the 1st/99th percentiles to limit outlier influence [17]. The final harmonized data comprised one matrix per cohort with 40 genes \u0026times; \u003cem\u003en\u003c/em\u003e samples (with \u003cem\u003en\u003c/em\u003e as above for each cohort). We also generated a unified high-dimensional matrix of all samples by merging on the 40 genes (plus, for exploratory analyses, an intersection of ~\u0026thinsp;11,544 genes common to all datasets post-HGNC mapping). This integrated dataset was the basis for subsequent modeling and subtype analysis. All data processing steps were conducted under version control to ensure complete reproducibility.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e provides an overview of the complete analytical workflow, from raw data preprocessing to model validation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe diagram summarizes the major steps: (i) initial quality control (handling of duplicates and missing values, sample-ID standardization); (ii) gene harmonization through HGNC-based mapping and intra-cohort z-score normalization; (iii) platform audit and identification of the shared gene intersection across cohorts (11 544 genes); and (iv) selection of the 40-gene \u003cem\u003ecore-PAM50\u003c/em\u003e subset common to all platforms.\u003c/p\u003e \u003cp\u003eModel training was performed in the METABRIC cohort using both LASSO-penalized Cox regression and Random Survival Forests (RSF), followed by external validation in TCGA (overall survival) and GSE25066 (disease-free survival).\u003c/p\u003e \u003cp\u003eFinal outputs include intrinsic-subtype assignment, survival and discrimination metrics (HR, C-index), cross-cohort correlation analysis, and principal component analysis (PCA) demonstrating subtype separation.\u003c/p\u003e \u003cp\u003e[Placeholder Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eClinical Endpoints: We defined outcome endpoints in each cohort according to its clinical context. For METABRIC and TCGA (population cohorts), the endpoint was overall survival (OS), measured from diagnosis to death from any cause, with censoring at last follow-up [10,11]. In the GSE25066 trial cohort, the primary endpoint was disease-free recurrence/Death-free survival (DRFS), defined as time from treatment to first relapse or breast cancer-specific death, with OS as a secondary endpoint in patients with extended follow-up [12]. These definitions align with cohort context: OS reflects overall mortality risk in unselected cohorts, whereas DRFS captures post-treatment recurrence risk in the high-risk neoadjuvant setting.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003ePrognostic Model Development\u003c/h3\u003e\n\u003cp\u003eWe designated the METABRIC cohort as the training dataset for model development due to its large size and long follow-up. The METABRIC expression matrix (already harmonized and restricted to 40 genes) was merged with its clinical data, yielding 2,173 patients with complete gene and survival information. After excluding cases lacking outcome data or with implausible survival times, 1,980 patients remained for model training (the slight reduction reflects removal of samples without complete expression or OS data; METABRIC originally profiled 2,173 tumors, but only\u0026thinsp;~\u0026thinsp;1,980 had full expression\u0026thinsp;+\u0026thinsp;OS available in our working set [10]).\u003c/p\u003e \u003cp\u003eCox-LASSO modeling: We fitted a penalized Cox proportional hazards model using the core-PAM50 genes in METABRIC. The glmnet package was used with 10-fold cross-validation to select the L1-penalty tuning parameter (α\u0026thinsp;=\u0026thinsp;1 for LASSO) [20]. The optimal regularization was at λ_min\u0026thinsp;=\u0026thinsp;0.0129, which minimized the partial likelihood deviance (vs. a more regularized λ_1se\u0026thinsp;=\u0026thinsp;0.0755). We chose λ_min to maximize prognostic signal retention, as even this yielded a sparse model. At λ_min, 20 out of 40 genes retained non-zero Cox coefficients, defining the final core-PAM50 prognostic model. The selected genes \u0026ndash; ACTR3B, BAG1, BCL2, BLM, CCNB1, CEP55, ERBB2, ESR1, FGFR4, FOXC1, MAPT, MDM2, MKI67, MLPH, NAT1, PGR, PHGDH, SLC39A6, TYMS, and UBE2C \u0026ndash; encompass both proliferation drivers and hormone-regulated genes, mirroring the dual \u0026ldquo;mitotic vs. luminal\u0026rdquo; nature of breast cancer risk. Notably, high-risk Basal/B or HER2-driven genes (e.g. \u003cem\u003eMKI67, CCNB1, UBE2C, CEP55, ERBB2\u003c/em\u003e) and low-risk Luminal cluster genes (\u003cem\u003eESR1, PGR, BCL2, NAT1, BAG1\u003c/em\u003e) were all represented, indicating that the LASSO pruned mainly redundant features while preserving key biology. The complete list of coefficients and hazard ratios for the 20 selected genes is provided in \u003cem\u003eSupplementary Table \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e\u003c/em\u003e\u003c/p\u003e \u003cp\u003eOn internal cross-validation, the Cox-LASSO model achieved a C-index of 0.584 for OS prediction in METABRIC, in line with published multigene prognostic signatures in similar cohorts (typical gene-only C-indices\u0026thinsp;~\u0026thinsp;0.58\u0026ndash;0.65) [5,6]. We observed no evidence of overfitting: the cross-validation error curve had a clear minimum at λ_min and a shallow U-shape around it, indicating a stable model not overly tuned to training idiosyncrasies. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrates the LASSO model fit, showing (A) the coefficient paths for all 40 genes as the penalty increases, and (B) the cross-validated deviance. As λ grows, most gene weights shrink to zero, leaving 20 active genes at the chosen λ_min; the deviance plot confirms that λ_min lies on a flat optimal region of the curve. These findings suggest the core-PAM50 Cox model captures true prognostic signal rather than noise.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e[placeholder Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eTo explore non-linear effects and interactions, we also trained a Random Survival Forest (RSF) on the METABRIC data using the same 40 genes [21]. An ensemble of 500 survival trees was grown (with bootstrap sampling and default splitting parameters). The RSF\u0026rsquo;s out-of-bag C-index was 0.585, virtually identical to the Cox-LASSO performance (0.584), reinforcing that the prognostic information is robust to modeling approach. We examined RSF variable importance (VIMP), which measures the impact of each gene on prediction accuracy. The top-ranked genes by VIMP were \u003cem\u003eMKI67, UBE2C, CCNB1, CDC20\u003c/em\u003e, and \u003cem\u003ePTTG1\u003c/em\u003e \u0026ndash; all involved in cell-cycle and mitotic progression \u0026ndash; while classical luminal genes (\u003cem\u003eESR1, BCL2, FOXA1\u003c/em\u003e, etc.) showed negative importance (their presence lowers risk, consistent with protective effect). These results delineate a clear biological axis of risk: high-risk patients are driven by overexpression of proliferation regulators, whereas low-risk patients have elevated expression of estrogen receptor and related markers associated with favorable outcome. (RSF variable importance results are provided in Supplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.) The concordance between the Cox-LASSO and RSF models \u0026ndash; both in performance and in the identity of top prognostic genes \u0026ndash; suggests that the core-PAM50 captures a stable signal not dependent on a single algorithm.\u003c/p\u003e \u003cp\u003eFor each METABRIC patient, we computed a continuous risk score as the Cox-LASSO linear predictor (sum of 20 gene expression \u0026times; coefficient values). Higher scores indicate higher predicted hazard of death. To visualize risk stratification, we divided the training cohort into tertiles by the risk score: top one-third labeled \u003cem\u003eHigh Risk\u003c/em\u003e, bottom one-third \u003cem\u003eLow Risk\u003c/em\u003e, and the middle \u003cem\u003eIntermediate\u003c/em\u003e. A Kaplan\u0026ndash;Meier analysis (METABRIC) confirmed a strong separation of survival curves across these groups (log-rank \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Patients in the High Risk tertile had approximately double the mortality hazard of those in Low Risk (hazard ratio\u0026thinsp;~\u0026thinsp;2.0), consistent with the model\u0026rsquo;s continuous risk gradient. A summary of model parameters and interval validation results is presented in Table\u0026nbsp;1.\u003c/p\u003e \u003cp\u003e[placeholder for table 1]\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eModel Validation and Evaluation\u003c/h2\u003e \u003cp\u003eWe applied the core-PAM50 Cox model, without any re-fit or recalibration, to the two external cohorts to assess transportability. Each TCGA and GSE25066 sample\u0026rsquo;s risk score was computed using the METABRIC-derived coefficients. We then evaluated discrimination and calibration in these validation sets. For consistency with training, we again split each validation cohort into predicted High, Intermediate, and Low Risk tertiles.\u003c/p\u003e \u003cp\u003eSurvival discrimination: In TCGA, which has relatively short median follow-up (~\u0026thinsp;3\u0026ndash;4 years) and many censored cases, the risk groups showed no significant difference in overall survival. The Kaplan\u0026ndash;Meier curves for TCGA High vs Low tertiles were largely overlapping (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA), with a hazard ratio (HR) of ~\u0026thinsp;1.1 (High vs Low; 95% CI 0.67\u0026ndash;1.19, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.45). This corresponded to an apparent C-index of only\u0026thinsp;~\u0026thinsp;0.42 for OS prediction in TCGA, indicating poor discrimination. However, this result must be interpreted in context: TCGA\u0026rsquo;s limited follow-up and the predominance of good-prognosis tumors (many Luminal A) mean few events occurred, so even a valid predictor will appear attenuated. In contrast, GSE25066 (neoadjuvant cohort) showed a pronounced separation between risk groups. For 5-year relapse-free survival in GSE25066, the High-Risk group had substantially worse outcome than Low Risk (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB): using the original model score (which was inversely associated with DRFS due to the different endpoint), the Low-Risk group had about half the hazard of relapse/death compared to High Risk (HR\u0026thinsp;~\u0026thinsp;0.50, 95% CI 0.35\u0026ndash;0.73, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). For clarity, if we align the interpretation with OS (where higher score\u0026thinsp;=\u0026thinsp;higher hazard), this translates to High-Risk patients having roughly double the hazard of Low Risk, analogous to the training cohort. The GSE25066 discrimination was reflected in a C-index of ~\u0026thinsp;0.63 at 5 years, indicating good predictive accuracy in that setting. We then evaluated discrimination and calibration in these validation sets.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e[placeholder for Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA-B]\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;2 summarizes the performance of the \u003cem\u003ecore-PAM50\u003c/em\u003e Cox model across training and validation datasets.\u003c/p\u003e \u003cp\u003eIn the METABRIC training set, the model achieved a hazard ratio (High vs Low) of 2.0 [1.5\u0026ndash;2.5], p\u0026thinsp;\u0026lt;\u0026thinsp;0.001, with an internal C-index of 0.584 and excellent 5-year calibration (slope\u0026thinsp;\u0026asymp;\u0026thinsp;1.0, intercept\u0026thinsp;\u0026asymp;\u0026thinsp;0.0).\u003c/p\u003e \u003cp\u003eIn the TCGA-BRCA validation cohort, discrimination was weak (C-index\u0026thinsp;=\u0026thinsp;0.42; p\u0026thinsp;=\u0026thinsp;0.45) due to short follow-up and a low event rate among predominantly Luminal A tumors.\u003c/p\u003e \u003cp\u003eConversely, the GSE25066 chemotherapy cohort demonstrated strong prognostic separation (HR\u0026thinsp;=\u0026thinsp;2.0 [1.4\u0026ndash;2.9], p\u0026thinsp;\u0026lt;\u0026thinsp;0.001; C-index\u0026thinsp;=\u0026thinsp;0.63), with near-perfect calibration.\u003c/p\u003e \u003cp\u003eA pooled fixed-effect meta-analysis across validation sets yielded an overall hazard ratio of 1.5 [1.1\u0026ndash;2.1], p\u0026thinsp;=\u0026thinsp;0.01, confirming that the model retained directionally consistent prognostic power across independent clinical and technological contexts.\u003c/p\u003e \u003cp\u003e[place holder for table 2]\u003c/p\u003e \u003cp\u003eTo summarize the effect sizes, Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents a forest plot summarizing hazard ratios (HRs) for \u003cem\u003eLow\u003c/em\u003e versus \u003cem\u003eHigh\u003c/em\u003e risk groups across independent cohorts.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the training set (METABRIC), by design, model stratification corresponded to an HR\u0026thinsp;\u0026asymp;\u0026thinsp;0.50 [95% CI 0.40\u0026ndash;0.67; p\u0026thinsp;\u0026lt;\u0026thinsp;0.001], indicating that patients classified as \u003cem\u003elow-risk\u003c/em\u003e had roughly half the hazard of death compared with those in the \u003cem\u003ehigh-risk\u003c/em\u003e group.\u003c/p\u003e \u003cp\u003eIn TCGA-BRCA, the HR was 0.89 [0.67\u0026ndash;1.19; p\u0026thinsp;=\u0026thinsp;0.45], showing no significant separation between risk groups\u0026mdash;consistent with the cohort\u0026rsquo;s short follow-up and predominance of luminal tumors.\u003c/p\u003e \u003cp\u003eIn GSE25066, the model achieved strong discrimination, with HR\u0026thinsp;=\u0026thinsp;0.50 [0.35\u0026ndash;0.73; p\u0026thinsp;\u0026lt;\u0026thinsp;0.001] for disease-free survival, meaning that \u003cem\u003elow-risk\u003c/em\u003e patients had about half the hazard of relapse or death relative to \u003cem\u003ehigh-risk\u003c/em\u003e cases.\u003c/p\u003e \u003cp\u003eWhen the two validation cohorts were combined under a random-effects meta-analysis, the pooled effect was HR\u0026thinsp;=\u0026thinsp;0.68 [0.39\u0026ndash;1.20], reflecting a consistent\u0026mdash;though not statistically significant\u0026mdash;trend favoring the \u003cem\u003elow-risk\u003c/em\u003e group.\u003c/p\u003e \u003cp\u003e[Figure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e placeholder]\u003c/p\u003e \u003cp\u003eCalibration and clinical utility: We assessed calibration of the risk predictions in time-dependent analyses. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows 5-year calibration plots for each cohort, comparing the predicted 5-year survival probability (from the Cox model) against observed outcomes. The METABRIC training model was well-calibrated by construction. In GSE25066, the calibration remained good: predicted vs. observed 5-year relapse-free survival aligned closely along the 45\u0026deg; line, indicating the model\u0026rsquo;s absolute risk estimates were accurate in this external setting. TCGA\u0026rsquo;s 5-year OS calibration was poorer \u0026ndash; the curve deviated below the ideal line, suggesting the model over-predicted risk for many patients (consistent with its low apparent performance there). This is again attributable to few events within 5 years in TCGA; longer follow-up would be needed for the predictions to manifest. Overall, calibration analysis supports that the model\u0026rsquo;s risk scores can be interpreted on an absolute scale for clinical risk estimation in contexts similar to the training data.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e[Figure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e placeholder]\u003c/p\u003e \u003cp\u003eWe next evaluated Decision Curve Analysis (DCA) to gauge potential clinical value. DCA quantifies the net benefit of using a prognostic model to guide interventions, compared to default strategies of treating all or none, across a range of risk thresholds [23]. In our context, one might use the core-PAM50 score to recommend adjuvant therapy if the predicted 5-year mortality risk exceeds a certain threshold (e.g. 10%). Net benefit is calculated as the true-positive rate minus the weighted false-positive rate, accounting for the \u0026ldquo;cost\u0026rdquo; of unnecessary treatment at each threshold probability. In METABRIC, the core-PAM50 model provided a clear net benefit above both treat-all and treat-none approaches for threshold probabilities between ~\u0026thinsp;3% and 40%. The benefit was maximal around a 5\u0026ndash;10% risk threshold, corresponding to a plausible decision cutoff for recommending chemotherapy in early breast cancer. At that point, the model\u0026rsquo;s net benefit equated to correctly influencing treatment in an additional\u0026thinsp;~\u0026thinsp;10% of patients without increasing over-treatment (relative to default strategies). GSE25066 showed a similarly positive net benefit in the 5\u0026ndash;30% threshold range. TCGA\u0026rsquo;s net benefit curve was flat (owing to its minimal discrimination), but notably still did not fall below the treat-all line for low thresholds up to ~\u0026thinsp;20%. In all cohorts, the model demonstrated at least no harm and at best substantial benefit in clinically relevant threshold ranges. (Detailed DCA plots are provided in Supplementary Figure \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e.) These results suggest that if the core-PAM50 score were used to guide adjuvant therapy decisions, it could improve patient selection modestly but meaningfully \u0026ndash; roughly equivalent to correctly sparing or treating 5\u0026ndash;10 out of 100 patients beyond standard criteria \u0026ndash; especially in high-risk settings.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eIntrinsic Subtype Classification and Concordance\u003c/h3\u003e\n\u003cp\u003eBeyond continuous risk prediction, the core-PAM50 retains the ability to classify tumors into intrinsic subtypes analogous to PAM50. We applied a classical nearest-centroid method to assign each tumor a subtype label (Luminal A, Luminal B, HER2-enriched, Basal-like, or Normal-like) based on its 40-gene expression profile. We obtained the PAM50 reference centroids from Parker \u003cem\u003eet al.\u003c/em\u003e (2009) [5] via the Bioconductor genefu implementation and restricted those centroids to the 40 core genes (dropping the 8 excluded genes). Each sample was correlated to the five subtype centroids (in the standardized \u003cem\u003ez\u003c/em\u003e-score space) and labeled with the subtype of highest Pearson correlation (the maximum correlation rule). All 40 core genes are represented in the centroids, so no missing features occurred in classification. Out of 6,850 total samples in the integrated dataset, 6,847 (99.95%) received a confident subtype assignment; only 3 samples (\u0026lt;\u0026thinsp;0.05%) had expression patterns too incomplete or ambiguous and were left \u0026ldquo;unclassified\u0026rdquo; (these were predominantly low-quality or low-tumor-content cases).\u003c/p\u003e \u003cp\u003eSubtype distributions: The frequency of intrinsic subtypes in each cohort aligned with known cohort characteristics, though some notable differences were observed (Table\u0026nbsp;1 and Supplementary Figure \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e). In METABRIC, the distribution was approximately 18% \u003cem\u003eBasal-like\u003c/em\u003e, 19% \u003cem\u003eHER2-enriched\u003c/em\u003e, 25% \u003cem\u003eLuminal A\u003c/em\u003e, 23% \u003cem\u003eLuminal B\u003c/em\u003e, and 15% \u003cem\u003eNormal-like\u003c/em\u003e. Although more balanced than historically reported, the relatively high proportion of \u003cem\u003eHER2-enriched\u003c/em\u003e and \u003cem\u003eLuminal B\u003c/em\u003e tumors and the moderate \u003cem\u003eNormal-like\u003c/em\u003e fraction likely reflect a combination of biological diversity and microarray-based signal compression affecting luminal gene expression. This artifact\u0026mdash;previously observed in Illumina platforms\u0026mdash;can cause some ER-positive, low-proliferation tumors to correlate less strongly with the \u003cem\u003eLuminal A\u003c/em\u003e centroid, modestly inflating the \u003cem\u003eNormal-like\u003c/em\u003e and \u003cem\u003eHER2-enriched\u003c/em\u003e categories [16]. Despite this, METABRIC retains a clear luminal predominance overall, consistent with its population-based composition. In TCGA-BRCA, the subtype frequencies were 37% \u003cem\u003eLuminal A\u003c/em\u003e, 27% \u003cem\u003eLuminal B\u003c/em\u003e, 18% \u003cem\u003eBasal-like\u003c/em\u003e, 11% \u003cem\u003eHER2-enriched\u003c/em\u003e, and 7% \u003cem\u003eNormal-like\u003c/em\u003e, closely matching the expected distribution for a contemporary, unselected breast cancer cohort with comprehensive RNA-seq profiling [11]. This luminal-dominant pattern is consistent with TCGA\u0026rsquo;s enrichment for hormone receptor\u0026ndash;positive, early-stage disease. The GSE25066 neoadjuvant cohort showed a distinct shift toward more aggressive phenotypes: approximately 34% \u003cem\u003eBasal-like\u003c/em\u003e, 8% \u003cem\u003eHER2-enriched\u003c/em\u003e, 32% \u003cem\u003eLuminal A\u003c/em\u003e, 19% \u003cem\u003eLuminal B\u003c/em\u003e, and 7% \u003cem\u003eNormal-like\u003c/em\u003e.\u003c/p\u003e \u003cp\u003eThis reflects the trial\u0026rsquo;s design, which selectively enrolled patients with high-risk tumors treated with taxane-anthracycline chemotherapy.\u003c/p\u003e \u003cp\u003eTogether, these baseline distributions emphasize the biological diversity across major breast cancer cohorts. They contextualize the need for subsequent cross-platform harmonization\u0026mdash;not to remove true biological differences, but to ensure that downstream analyses, including application of the core-PAM50, compare equivalent molecular signals rather than technical artifacts.\u003c/p\u003e \u003cp\u003eCross-cohort expression concordance: We next evaluated whether the core-PAM50 reproduces the intrinsic subtype gene expression patterns consistently across platforms. Despite differences in technology and patient demographics, subtype expression \u0026ldquo;centroids\u0026rdquo; derived from each cohort were highly correlated with one another. For every subtype, the pairwise Pearson correlation between the average expression profiles of any two cohorts exceeded 0.80. In fact, most subtype centroids showed cross-cohort correlations on the order of r\u0026thinsp;\u0026asymp;\u0026thinsp;0.90 or higher. For example, the average Basal-like profile in METABRIC vs. TCGA had r\u0026thinsp;\u0026asymp;\u0026thinsp;0.95, and TCGA vs. GSE25066 Basal centroids r\u0026thinsp;\u0026asymp;\u0026thinsp;0.93. Even the more heterogeneous subtypes were strongly concordant (Luminal A median r\u0026thinsp;\u0026asymp;\u0026thinsp;0.83 across cohort pairs; HER2-enriched r\u0026thinsp;\u0026asymp;\u0026thinsp;0.90). Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e illustrates these relationships as heatmaps of subtype-specific Pearson correlations between METABRIC, TCGA, and GSE25066. Each subtype forms a block of uniformly high correlation between cohorts, indicating that a Basal-like tumor\u0026rsquo;s expression pattern in a microarray dataset is statistically almost identical to a Basal-like tumor\u0026rsquo;s pattern in an RNA-seq dataset, and similarly for other subtypes. Notably, cross-platform concordance was lowest for Luminal A (still\u0026thinsp;\u0026gt;\u0026thinsp;0.8), which is expected given greater biological diversity among ER-positive tumors and the noted microarray intensity compression. In contrast, Basal-like and Luminal B profiles were virtually superimposable across datasets (r\u0026thinsp;\u0026gt;\u0026thinsp;0.93 in all comparisons), underscoring that the proliferative gene module defining these subtypes is captured robustly by the 40-gene panel regardless of platform. These findings confirm that the core-PAM50 preserves the canonical transcriptional architecture of the intrinsic subtypes in a platform-independent manner.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e[Figure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e placeholder]\u003c/p\u003e \u003cp\u003ePrincipal component analysis (PCA) of subtype structure. To visualize the global structure of the 40-gene expression space, we performed principal component analysis (PCA) within each cohort using the standardized \u003cem\u003ecore-PAM50\u003c/em\u003e expression matrix.\u003c/p\u003e \u003cp\u003eIn all datasets, the first principal component (PC1) captured the dominant biological axis contrasting proliferative (Basal-like and Luminal B) versus hormone receptor\u0026ndash;driven (Luminal A and Normal-like) phenotypes.\u003c/p\u003e \u003cp\u003eIn the RNA-seq dataset (TCGA), PC1 accounted for 46.0% of variance and clearly separated Basal-like from Luminal tumors, with HER2-enriched cases in intermediate positions and Normal-like samples near the origin.\u003c/p\u003e \u003cp\u003eIn GSE25066, a similar pattern was observed (PC1\u0026thinsp;=\u0026thinsp;37.1%, PC2\u0026thinsp;=\u0026thinsp;16.5%), indicating robust subtype separation despite technological differences.\u003c/p\u003e \u003cp\u003eIn METABRIC, PC1 and PC2 explained 11.4% and 8.9% of variance, respectively, with greater overlap among subtypes\u0026mdash;an expected consequence of microarray signal compression and a higher proportion of Normal-like assignments.\u003c/p\u003e \u003cp\u003eNevertheless, even in METABRIC, Basal-like and Luminal B samples trended along the positive PC1 direction (high MKI67, CCNB1, UBE2C), whereas Luminal A and Normal-like tumors occupied the opposite side (high ESR1, FOXA1, BCL2).\u003c/p\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e, PC1\u0026ndash;PC2 plots for representative cohorts illustrate consistent spatial segregation of Basal and Luminal tumors across datasets, with HER2-enriched and Normal-like samples occupying intermediate or diffuse regions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTogether with the subtype classification and correlation analyses, these findings demonstrate that the 40-gene \u003cem\u003ecore-PAM50\u003c/em\u003e panel (i) faithfully reproduces intrinsic subtype assignments, (ii) yields cohort-specific subtype distributions aligned with known clinical and technical contexts, and (iii) preserves the canonical transcriptional geometry of breast cancer subtypes across platforms\u0026mdash;confirming that the reduced gene set retains the full molecular taxonomy of the original PAM50.\u003c/p\u003e \u003cp\u003e[Figure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e placeholder]\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eCross-platform robustness and reproducibility: This study demonstrates that a carefully harmonized 40-gene subset of the PAM50 signature can yield a prognostic model that generalizes across different genomic technologies. We achieved a unified analysis of four large cohorts (over 6,800 samples) spanning two microarray platforms and RNA-seq, something that is often impeded by batch effects and annotation discrepancies [15]. A cornerstone of our approach was enforcing consistent gene nomenclature (up-to-date HGNC symbols) and standardized within-cohort normalization, which eliminated many technical biases. The outcome was a cross-platform model whose signal was largely platform-independent: we observed highly correlated subtype centroids across cohorts (r\u0026thinsp;~\u0026thinsp;0.9) and comparable prognostic performance metrics in validation. In contrast, using un-harmonized data would likely have failed \u0026ndash; even identical gene signatures can perform poorly when applied to raw data from another platform due to calibration drift or missing genes. By solving those issues upfront, the core-PAM50 preserves the \u0026ldquo;common denominator\u0026rdquo; of the expression patterns. In practice, this means a researcher or clinician could apply our 40-gene model to a tumor profiled by microarray or by RNA-seq (or potentially even by qPCR or NanoString) and expect a similar risk stratification. This kind of platform-agnostic assay is a step toward universally available genomic tests.\u003c/p\u003e \u003cp\u003eCohort differences and validation context: The disparate validation results in TCGA and GSE25066 highlight the importance of clinical context in evaluating prognostic models. TCGA patients had predominantly good outcomes at the available follow-up (many Luminal A tumors, treated with modern therapy, and censored alive within 5 years) \u0026ndash; a scenario where demonstrating risk discrimination is inherently difficult. The model\u0026rsquo;s lack of significant separation in TCGA does not necessarily indicate a failure of the gene signature; rather, it reflects a low event rate and limited follow-up window. In a real-world sense, applying a prognostic model to an indolent population will yield attenuated performance even if the model is valid. By contrast, GSE25066, enriched for aggressive tumors (mostly Basal-like, high-grade disease) and observing many relapse events, displayed a strong risk stratification. Notably, the GSE25066 endpoint was relapse-free survival after neoadjuvant chemotherapy, which might amplify differences: patients with highly proliferative tumors remained at risk despite chemotherapy, whereas those with low-risk luminal biology fared better. These differences emphasize that prognostic effect sizes are context-dependent. A single model might appear to \u0026ldquo;work\u0026rdquo; better in one cohort than another not due to any flaw in the model, but because of differences in follow-up duration, treatment, and baseline risk. Our pooled analysis suggested that, on average, the core-PAM50 retained significant prognostic value externally (combined HR\u0026thinsp;~\u0026thinsp;1.5 for high vs low risk). Still, a cautious approach is to tailor expectations of model performance to the scenario \u0026ndash; for instance, one would not expect a dramatic split in a low-event-rate population.\u003c/p\u003e \u003cp\u003eComparison with existing signatures: The core-PAM50 model\u0026rsquo;s performance falls in the same range as established breast cancer gene signatures. In METABRIC, we achieved a training C-index of ~\u0026thinsp;0.58, and in external validation (GSE25066) a time-dependent AUC\u0026thinsp;~\u0026thinsp;0.60\u0026ndash;0.63 at 5 years. This is comparable to reported values for PAM50\u0026rsquo;s Risk-Of-Recurrence (ROR) score and other assays: for example, Oncotype DX (21-gene Recurrence Score) typically yields C-index\u0026thinsp;~\u0026thinsp;0.6 in ER-positive cohorts, and MammaPrint (70-gene signature) around 0.62\u0026ndash;0.65, with improvements to ~\u0026thinsp;0.70 when clinical factors like tumor size or nodal status are added [6]. Our results mirror those benchmarks. Importantly, we did not include standard clinical variables in our model to isolate the gene signature\u0026rsquo;s performance and maximize generalizability across cohorts (since detailed clinical covariates were not uniformly available for all datasets). In practice, an integrated model that combines core-PAM50 with tumor stage, grade, etc., would likely boost predictive accuracy (prior studies have shown gene signatures\u0026thinsp;+\u0026thinsp;clinical factors can raise C-index by ~\u0026thinsp;0.05\u0026ndash;0.1 [6]). However, our focus was to create a purely molecular metric that could be retrospectively evaluated in any transcriptomic dataset. The trade-off is a moderate discriminative power on its own (~\u0026thinsp;0.6 C-index), but that is an inherent limitation of gene-only prognostic indices and consistent with the performance of much costlier assays. Given its transparency and portability, the core-PAM50\u0026rsquo;s prognostic ability is quite encouraging.\u003c/p\u003e \u003cp\u003eContinuous risk vs subtype categories: A noteworthy feature of our model is that it produces a continuous risk score, even though it derives from subtype-related genes. This numeric risk has advantages in clinical decision-making \u0026ndash; it preserves more information than coarse subtype labels and can be translated to absolute risk probabilities. Clinicians often prefer simple categories (e.g. \u0026ldquo;low risk\u0026rdquo; or \u0026ldquo;high risk\u0026rdquo;), but those can be obtained by setting thresholds on the continuous score as needed. In fact, flexibility in thresholding is a strength: different cutoff values could be chosen depending on the clinical scenario (for example, a higher risk threshold for recommending chemotherapy in lower-stage disease, and a lower threshold for deciding on extended endocrine therapy). Our analysis showed that dichotomizing the score into tertiles provided a meaningful stratification; other cut-points could be calibrated to specific interventions. The underlying continuous nature also means the model can be used in multifactorial risk calculations (e.g. combining with clinical nomograms) without losing granularity.\u003c/p\u003e \u003cp\u003eTechnical considerations \u0026ndash; Normal-like subtype and assay quality: One unexpected result was the extremely high rate of Normal-like classifications in METABRIC. This likely reflects issues of tumor purity or assay quality in that older dataset \u0026ndash; samples with low tumor cellularity or RNA degradation tend to exhibit expression profiles that resemble normal breast tissue, thus getting labeled \u0026ldquo;Normal-like\u0026rdquo; by the algorithm. This is a known phenomenon [16]. In our study, those Normal-like METABRIC cases were essentially equivalent to indolent Luminal A tumors (they had low proliferation and excellent survival), so the prognostic model wasn\u0026rsquo;t adversely affected (it correctly assigned them low risk). However, it raises a practical point: if an assay yields a subtype call of Normal-like in a context where that is unexpected, it could indicate a technical issue (e.g. a largely stromal sample). Future development of the core-PAM50 test could incorporate a quality flag \u0026ndash; for instance, if no subtype correlation exceeds a low threshold, reporting the result as \u0026ldquo;unclassified, possibly low tumor content\u0026rdquo; might be prudent. In our data, only 3 samples met such a criterion (essentially no signal), but dozens of METABRIC tumors had moderate correlations to multiple centroids and ended up as Normal-like. In a prospective setting, one might consider those as \u003cem\u003eLuminal A\u003c/em\u003e for practical purposes or require a pathology review. We handled this by acknowledging the bias rather than forcing a reclassification, but it is an area for further refinement (e.g. developing a pre-test tumor purity assessment).\u003c/p\u003e \u003cp\u003eAnother technical consideration is the absence of immune genes in the core-PAM50. Recent research highlights immune infiltration (tumor-infiltrating lymphocytes and related gene signatures) as an important prognostic factor, especially in triple-negative (Basal-like) breast cancers [9]. The PAM50 focuses on tumor-intrinsic genes and includes few direct immune markers. By trimming to 40 genes, we may have excluded some genes that, while not core to subtype classification, could carry immune/prognostic information (for instance, if any of the dropped PAM50 genes had immune relevance, though most were basal cytokeratins or lesser-expressed genes). Augmenting the core-PAM50 with an immune module could potentially improve prognostic power for certain patients. Indeed, one could envision a combined model that adds a TILs gene score to better stratify high-risk basal tumors. That said, adding genes would sacrifice some simplicity and the guarantee of cross-platform detectability. Our aim was a lean panel that is broadly applicable; additional features could be explored in future work, but the benefit would need to justify the complexity.\u003c/p\u003e \u003cp\u003eClinical implications and future directions: We have shown that the core-PAM50 model encapsulates the essence of the intrinsic subtype system and provides a continuous risk metric that is validated across diverse cohorts. The emphasis of this work is on reproducibility and transparency. All data preprocessing and modeling steps were documented and adhere to reproducible research principles, which is often not the case for commercial assays. This means others can audit, reuse, or adapt our pipeline \u0026ndash; an important advantage in the era of open science and evolving genomic data. We envision several practical applications and extensions of this model. First, it can be used retrospectively on existing gene expression datasets (from clinical trials or real-world cohorts) to generate a standardized risk score, facilitating meta-analyses or comparisons across studies. Second, it provides a template for laboratories to implement a cost-effective prognostic assay: since it uses only 40 genes and straightforward calculations, it could be deployed on platforms like quantitative PCR or targeted RNA-seq with relative ease, potentially expanding access in resource-limited settings. Third, prospective validation in a clinical trial would be valuable \u0026ndash; for example, testing whether this model can identify patients who benefit from certain therapies (e.g. need for chemotherapy in what would otherwise be an intermediate clinical risk group). The model\u0026rsquo;s performance for late relapse (beyond 5 years) in ER\u0026thinsp;+\u0026thinsp;disease is another area to investigate; our data hinted that core-PAM50 retains prognostic information up to 10 years (similar to PAM50 ROR which differentiates Luminal A vs B long-term [8]), but specialized assays like Breast Cancer Index target that scenario and could be compared.\u003c/p\u003e \u003cp\u003eIn a broader sense, our study exemplifies how reducing a complex genomic signature to its most robust elements, combined with rigorous data harmonization, can enhance real-world applicability. We effectively bridge the gap between large-scale \u0026ldquo;discovery\u0026rdquo; datasets and a clinical-grade test. The core-PAM50 model achieved a similar prognostic performance to more cumbersome signatures, while using\u0026thinsp;~\u0026thinsp;30% fewer genes and being openly reproducible. This parsimony can translate to lower costs, simpler assay design, and fewer regulatory hurdles (as there are fewer biomarkers to validate). Moreover, by avoiding platform-specific artifacts, it increases confidence that results from different hospitals or studies can be directly compared.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, we developed a reduced 40-gene PAM50-derived signature and demonstrated its technical and prognostic performance across multiple breast cancer cohorts and genomic platforms. The core-PAM50 model retains the essential biological signal of the intrinsic subtypes and provides a continuous risk score that is validated in independent microarray and RNA-seq datasets. Emphasizing reproducibility and transparency, we implemented a fully documented analysis pipeline, ensuring that results can be audited and reproduced by other researchers. This reproducible framework adds significant value compared to many proprietary tests, as it adheres to FAIR data principles and modern reporting standards [25].\u003c/p\u003e \u003cp\u003eThe core-PAM50 panel offers a practical path toward a cost-effective, high-stability genomic assay for breast cancer prognosis. With roughly one-third fewer genes than the canonical PAM50 and avoidance of platform-specific pitfalls, it can reduce experimental complexity and validation burden by ~\u0026thinsp;30\u0026ndash;40% while maintaining prognostic accuracy. Our multi-cohort validation supports its potential generalizability: risk stratification was consistent in principle across diverse cohorts, and the molecular subtyping remained intact. Moving forward, integration of the core-PAM50 model into clinical research is warranted. For instance, it could be prospectively tested in translational trials or applied to re-analyze archived trial samples to glean additional prognostic insights. Further refinements, such as incorporating complementary features (immune markers or clinical covariates) and addressing classification nuances (e.g. the Normal-like issue), may enhance the model\u0026rsquo;s utility, but even in its present form the evidence strongly supports its robustness and clinical relevance.\u003c/p\u003e \u003cp\u003eUltimately, the core-PAM50 signature represents a parsimonious yet biologically informed prognostic tool that brings large-scale genomic knowledge closer to routine clinical practice. It preserves the rich information content of the PAM50 intrinsic subtypes \u0026ndash; a taxonomy deeply ingrained in breast cancer biology \u0026ndash; and repackages it into a lean, reproducible assay framework suitable for global oncology use. With further validation, this model could help optimize personalized prognostication and therapy decisions while simplifying the technology required, thereby making genomic risk assessment more accessible to patients worldwide.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAUC – Area under the curve\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;BC – Breast cancer\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;CI – Confidence interval\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;CV – Cross-validation\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;DCA – Decision curve analysis\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;DRFS – Disease-free recurrence/death-free survival\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;ER – Estrogen receptor\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;FFPE – Formalin-fixed paraffin-embedded\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;GEO – Gene Expression Omnibus\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;GSE – Gene Expression Series\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;HGNC – HUGO Gene Nomenclature Committee\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;HER2 – Human epidermal growth factor receptor 2\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;HR – Hazard ratio\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;ID – Identifier\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;IHC – Immunohistochemistry\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;KM – Kaplan–Meier\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;LASSO – Least absolute shrinkage and selection operator\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;MD5 – Message Digest 5 (cryptographic hash function)\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;METABRIC – Molecular Taxonomy of Breast Cancer International Consortium\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;mRNA – Messenger RNA\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;OS – Overall survival\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;PAM50 – Prediction Analysis of Microarray 50\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;PCA – Principal component analysis\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;PR – Progesterone receptor\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;QC – Quality control\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;RNA-seq – RNA sequencing\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;RSF – Random survival forest\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;ROR – Risk of recurrence\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;SD – Standard deviation\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;SE – Standard error\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;TCGA – The Cancer Genome Atlas\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;TRIPOD – Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;VIMP – Variable importance (in Random Survival Forests)\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;z-score – Standardized gene expression value (mean 0, SD 1)\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics approval and consent to participate: not applicable.\u003c/p\u003e\n\u003cp\u003eConsent for publication: not applicable.\u003c/p\u003e\n\u003cp\u003eAvailability of data and material: data supporting the results are available upon request to the corresponding author.\u003c/p\u003e\n\u003cp\u003eCompeting interests: the authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003eFunding: this research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003eAuthors\u0026rsquo; contributions: RNB conceived and designed the study, performed data analysis, and wrote the manuscript. JBS supervised the study and contributed to manuscript revision.\u003c/p\u003e\n\u003cp\u003eBoth authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003eAcknowledgements: we thank the Biostatistics Department of the University of Brasilia for technical assistance.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSung H, Ferlay J, Siegel RL, \u003cem\u003eet al.\u003c/em\u003e Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. \u003cem\u003eCA Cancer J Clin.\u003c/em\u003e 2021;71(3):209–249.\u003c/li\u003e\n\u003cli\u003ePerou CM, Sørlie T, Eisen MB, \u003cem\u003eet al.\u003c/em\u003e Molecular portraits of human breast tumours. \u003cem\u003eNature.\u003c/em\u003e 2000;406(6797):747–752.\u003c/li\u003e\n\u003cli\u003eYersal O, Barutca S. Biological subtypes of breast cancer: Prognostic and therapeutic implications. \u003cem\u003eWorld J Clin Oncol.\u003c/em\u003e 2014;5(3):412–424.\u003c/li\u003e\n\u003cli\u003eIgnatiadis M, Sotiriou C. Luminal breast cancer: from biology to treatment. \u003cem\u003eNat Rev Clin Oncol.\u003c/em\u003e 2013;10(9):494–506.\u003c/li\u003e\n\u003cli\u003eParker JS, Mullins M, Cheang MCU, \u003cem\u003eet al.\u003c/em\u003e Supervised risk predictor of breast cancer based on intrinsic subtypes. \u003cem\u003eJ Clin Oncol.\u003c/em\u003e 2009;27(8):1160–1167.\u003c/li\u003e\n\u003cli\u003eNielsen TO, Parker JS, Leung S, \u003cem\u003eet al.\u003c/em\u003e Comparison of PAM50 intrinsic subtyping with immunohistochemistry and clinical prognostic factors. \u003cem\u003eClin Cancer Res.\u003c/em\u003e 2010;16(21):5222–5232.\u003c/li\u003e\n\u003cli\u003eAli HR, Rueda OM, Chin SF, \u003cem\u003eet al.\u003c/em\u003e Genome-driven integrated classification of breast cancer validated in 11,000 tumours across 10 independent datasets. \u003cem\u003eNat Commun.\u003c/em\u003e 2020;11(1):4865.\u003c/li\u003e\n\u003cli\u003eRueda OM, Sammut SJ, Seoane JA, \u003cem\u003eet al.\u003c/em\u003e Dynamics of breast-cancer relapse reveal late-recurring ER-positive genomic subgroups. \u003cem\u003eNature.\u003c/em\u003e 2019;567(7748):399–404.\u003c/li\u003e\n\u003cli\u003eLuen SJ, Salgado R, Fox S, \u003cem\u003eet al.\u003c/em\u003e Tumour-infiltrating lymphocytes and prognosis in different subtypes of breast cancer: a pooled analysis. \u003cem\u003eLancet Oncol.\u003c/em\u003e 2017;18(1):87–96.\u003c/li\u003e\n\u003cli\u003eCurtis C, Shah SP, Chin SF, \u003cem\u003eet al.\u003c/em\u003e The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups. \u003cem\u003eNature.\u003c/em\u003e 2012;486(7403):346–352.\u003c/li\u003e\n\u003cli\u003eCancer Genome Atlas Network. Comprehensive molecular portraits of human breast tumours. \u003cem\u003eNature.\u003c/em\u003e 2012;490(7418):61–70.\u003c/li\u003e\n\u003cli\u003eHatzis C, Pusztai L, Valero V, \u003cem\u003eet al.\u003c/em\u003e A genomic predictor of response and survival following taxane-anthracycline chemotherapy for invasive breast cancer. \u003cem\u003eJ Clin Oncol.\u003c/em\u003e 2011;29(17):2433–2440.\u003c/li\u003e\n\u003cli\u003eBrueffer C, Darai E, Segersten U, \u003cem\u003eet al.\u003c/em\u003e The Swedish SCAN-B initiative: a large-scale multicenter study for breast cancer. \u003cem\u003eClin Cancer Res.\u003c/em\u003e 2019;25(19):5859–5872.\u003c/li\u003e\n\u003cli\u003eLee JS, Jung J, Kim Y, \u003cem\u003eet al.\u003c/em\u003e Integrated multi-omics analysis of breast cancer reveals genomic correlates of clinical outcomes. \u003cem\u003eNat Med.\u003c/em\u003e 2023;29(1):49–62.\u003c/li\u003e\n\u003cli\u003eLeek JT, Scharpf RB, Bravo HC, \u003cem\u003eet al.\u003c/em\u003e Tackling the critical impact of batch effects in high-throughput data. \u003cem\u003eNat Rev Genet.\u003c/em\u003e 2010;11(10):733–739.\u003c/li\u003e\n\u003cli\u003eAli HR, Lundberg A, Ohnstad HO, \u003cem\u003eet al.\u003c/em\u003e Global prognostic and predictive value of PAM50 across molecular and clinical subgroups: a meta-analysis. \u003cem\u003eBreast Cancer Res.\u003c/em\u003e 2023;25(1):57.\u003c/li\u003e\n\u003cli\u003eLi Y, Jiang T, Zhang B, \u003cem\u003eet al.\u003c/em\u003e A 40-gene expression signature predicts clinical outcomes in breast cancer. \u003cem\u003eEBioMedicine.\u003c/em\u003e 2022;82:104136.\u003c/li\u003e\n\u003cli\u003eLi C, \u003cem\u003eet al.\u003c/em\u003e Reducing feature dimensionality of genomic classifiers while preserving predictive performance. \u003cem\u003eBrief Bioinform.\u003c/em\u003e 2021;22(6):bbab358.\u003c/li\u003e\n\u003cli\u003eYates B, Braschi B, Gray KA, \u003cem\u003eet al.\u003c/em\u003e Genenames.org: the HGNC and VGNC resources. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e 2020;48(D1):D649–D655.\u003c/li\u003e\n\u003cli\u003eTibshirani R. The lasso method for variable selection in the Cox model. \u003cem\u003eStat Med.\u003c/em\u003e 1997;16(4):385–395.\u003c/li\u003e\n\u003cli\u003eIshwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random survival forests for R. \u003cem\u003eBioinformatics.\u003c/em\u003e 2008;24(11):1363–1371.\u003c/li\u003e\n\u003cli\u003eBlanche P, Dartigues JF, Jacqmin-Gadda H. Estimating time-dependent AUC for censored event times. \u003cem\u003eStat Med.\u003c/em\u003e 2013;32(30):5381–5397.\u003c/li\u003e\n\u003cli\u003eKattan MW, Vickers AJ, Steyerberg EW. Evaluating the calibration of predictive models in survival analysis. \u003cem\u003eStat Methods Med Res.\u003c/em\u003e 2020;29(2):515–536.\u003c/li\u003e\n\u003cli\u003eVickers AJ, Van Calster B, Steyerberg EW. Net benefit approaches to the evaluation of prediction models. \u003cem\u003eBMJ.\u003c/em\u003e 2019;365:l2373.\u003c/li\u003e\n\u003cli\u003eSteyerberg EW, Collins GS, Dauholuk A, \u003cem\u003eet al.\u003c/em\u003e Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD-2022). \u003cem\u003eNat Med.\u003c/em\u003e 2023;29(3):591–599.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1 and 2 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"breast-cancer-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"brcr","sideBox":"Learn more about [Breast Cancer Research](http://breast-cancer-research.biomedcentral.com)","snPcode":"13058","submissionUrl":"https://submission.nature.com/new-submission/13058/3","title":"Breast Cancer Research","twitterHandle":"@BCRJournal","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Breast cancer, Gene expression signature, Intrinsic subtypes, Prognostic model, Multi-cohort study, Reproducibility","lastPublishedDoi":"10.21203/rs.3.rs-8048795/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8048795/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003ePAM50 is a widely adopted multigene signature for breast-cancer subtyping and prognosis, but cross-platform variability and incomplete gene coverage limit its portability. We developed a streamlined, platform-agnostic core-PAM50 panel (40 genes) and a fully documented pipeline to deliver reproducible prognostic modeling across major public cohorts.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eTranscriptomes and clinical data from METABRIC (microarray, n\u0026thinsp;=\u0026thinsp;2,173), TCGA-BRCA (RNA-seq, n\u0026thinsp;=\u0026thinsp;1,098), and GSE25066 (microarray, neoadjuvant chemotherapy, n\u0026thinsp;=\u0026thinsp;508) were harmonized using HGNC symbol mapping and intra-cohort gene-wise z-scaling. Models were trained in METABRIC with LASSO-penalized Cox regression and explored with Random Survival Forests; the fixed METABRIC coefficients were applied without recalibration to TCGA and GSE25066. Performance was assessed by C-index, time-dependent AUC, calibration at 5 years, decision-curve analysis (DCA), and meta-analysis of hazard ratios (HR). Intrinsic subtypes were assigned by nearest-centroid correlation restricted to the 40 genes, and cross-cohort subtype centroids were compared by Pearson r.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe LASSO model retained 20/40 genes capturing a luminal\u0026ndash;proliferative axis; internal discrimination in METABRIC was C-index 0.584. External discrimination was AUC₆₀ \u0026asymp; 0.60\u0026ndash;0.63 in GSE25066 and attenuated in TCGA (C-index\u0026thinsp;\u0026asymp;\u0026thinsp;0.42), consistent with short follow-up and low event rates. Using Low vs High risk orientation, HRs were 0.50 (METABRIC OS; ~0.40\u0026ndash;0.67), 0.89 (TCGA OS; 0.67\u0026ndash;1.19), and 0.50 (GSE25066 DRFS; 0.35\u0026ndash;0.73). The random-effects pooled estimate across validation cohorts was HR 0.68 (0.39\u0026ndash;1.20), indicating a consistent protective direction for the low-risk group. Calibration was excellent in METABRIC and good in GSE25066; DCA showed positive net benefit in clinically relevant threshold ranges in both. Subtype centroids were highly concordant across platforms (r\u0026thinsp;\u0026gt;\u0026thinsp;0.8, often\u0026thinsp;\u0026asymp;\u0026thinsp;0.9), and PCA reproduced expected basal\u0026ndash;luminal separation.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThe core-PAM50 condenses PAM50 to 40 cross-platform genes while preserving intrinsic-subtype biology and yielding a portable, reproducible prognostic score validated across microarray and RNA-seq cohorts. Its transparency and parsimony provide a practical path toward cost-effective assays (qPCR/targeted RNA-seq) and facilitate meta-analytic reuse. Prospective studies and integration with clinical or immune features may further enhance clinical utility.\u003c/p\u003e\u003ch2\u003eTrial registration\u003c/h2\u003e \u003cp\u003eNot applicable.\u003c/p\u003e","manuscriptTitle":"Cross-Platform Reproducible Modeling of Breast Cancer Prognosis Using the Core-PAM50 Gene Signature","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-19 17:16:22","doi":"10.21203/rs.3.rs-8048795/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-23T13:36:00+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-05T17:10:57+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-17T04:08:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"136294010757300782593028055597243765110","date":"2026-01-11T03:07:49+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"197843687512042882597168141170443316149","date":"2026-01-05T15:45:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"125247648694050165130326962362523999952","date":"2025-12-18T17:20:53+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-17T12:36:53+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-11-12T06:14:37+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-11-12T00:33:14+00:00","index":"","fulltext":""},{"type":"submitted","content":"Breast Cancer Research","date":"2025-11-06T13:48:08+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"breast-cancer-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"brcr","sideBox":"Learn more about [Breast Cancer Research](http://breast-cancer-research.biomedcentral.com)","snPcode":"13058","submissionUrl":"https://submission.nature.com/new-submission/13058/3","title":"Breast Cancer Research","twitterHandle":"@BCRJournal","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d62b2732-2208-4733-b658-83e9c3f68af0","owner":[],"postedDate":"December 19th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-27T14:56:20+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-19 17:16:22","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8048795","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8048795","identity":"rs-8048795","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}