Pretreatment Circulating Vascular Biomarkers Predict Cancer Therapy–Related Cardiac Dysfunction During HER2+ Breast Cancer Treatment Authors: Dakota Gustafson, Priya Mistry, Crizza Ching, Inbar Nardi-Agmon, Christopher Yu, Chun-Po Steve Fan, Christian Houbois, Eitan Amir, Thomas Marwick, Husam Abdel-Qadir, Chris McIntosh, Paaladinesh Thavendiranathan, Jason E Fish Published: 2025 ABSTRACT AbstractBackgroundBlood biomarkers to predict cancer therapy–related cardiac dysfunction (CTRCD) risk remain limited.ObjectivesThe aim of this study was to identify circulating biomarkers associated with CTRCD risk in HER2+ breast cancer patients.MethodsIn the discovery cohort, women with early-stage HER2+ breast cancer receiving anthracycline and trastuzumab therapy underwent serial evaluation with cardiac magnetic resonance imaging (CMR), echocardiography, clinical assessments, and blood biobanking every 3 months. Multiomics profiling of 3 circulating cardiac damage biomarkers and 35 markers of inflammation, angiogenesis and endothelial activation and profiling of >2,000 plasma microRNAs were performed before and early during treatment (3 and 6 months). CTRCD was defined by left ventricular ejection fraction measured on CMR, and sensitivity analyses used echocardiography. Pretreatment protein biomarkers were measured in a validation cohort.ResultsAmong 136 women, 37 (27.2%) developed CMR-defined CTRCD within 15 months. The endothelial activation markers angiopoietin-2, endothelin-1, and endoglin were elevated before and during treatment, while sE-selectin was elevated during treatment in patients who developed CTRCD. Inflammatory biomarkers (myeloperoxidase, interferon-gamma–induced protein-10, and interferon-α) were significantly higher before treatment in patients who developed CTRCD. No differences were observed in cardiac injury biomarkers (troponin I, B-type natriuretic peptide, and growth differentiation factor-15). Pretreatment plasma microRNAs revealed distinct CTRCD-associated signatures. Integrating pretreatment clinical variables, CMR parameters, and biomarkers into a single random forest model, angiopoietin-2, myeloperoxidase, and endoglin were identified as the strongest predictors of CTRCD, findings that were subsequently validated in a cohort of 38 HER2+ breast cancer patients.ConclusionsPretreatment endothelial-centric and inflammatory biomarkers outperformed both clinical and CMR measures in predicting CTRCD during chemotherapy.Key Words: angiopoeitin-2, biomarkers; breast cancer; cancer therapy–related cardiac dysfunction; cardio-oncology; endothelium; inflammation; machine learning; microRNA; myeloperoxidase; vascular Although advances in therapy have improved survival in early-stage breast cancer, treatment remains associated with an increased risk for heart failure.1,2 This is particularly true for women with HER2+ breast cancer receiving sequential anthracycline and trastuzumab therapy, who are at risk for cancer therapy-related cardiac dysfunction (CTRCD) and heart failure.2 Although the use of this regimen is declining, it remains common in patients with a high tumor burden.3 As such, the early identification of patients at risk for CTRCD, before or early during therapy, remains an unmet clinical need.4 Several protein blood biomarkers, including cardiac troponins, natriuretic peptides, and markers of cardiac stress or inflammation (eg, growth differentiation factor 15 [GDF-15]), have been explored for CTRCD risk stratification; however, evidence is mixed.5, 6, 7, 8, 9 To date, there are limited data on biomarkers related to the vascular endothelium for risk prediction, despite a growing appreciation of the deleterious effects of antineoplastics on the vasculature.10, 11, 12, 13 These agents may directly damage the coronary microcirculation or promote inflammation and coagulation, both of which can compromise endothelial health.13, 14, 15 Analyses of select endothelium-related markers such as nicotinamide adenine dinucleotide phosphate oxidase 4,16 arginine–nitric oxide metabolites,17 and activation of coagulation or inflammatory markers6,8,13,18,19 have highlighted their involvement in CTRCD and their potential role in prognostication. Intriguingly, some of these endothelial activation, coagulation, and inflammatory markers appear to be elevated pretreatment,13 suggesting that there may be a discernible pretreatment risk profile. Transcriptomic profiling of circulating plasma microRNAs (miRNAs), because of their stability and high dimensionality, offers further promise for early prognostication.20,21 Recently, multidimensional approaches, including machine learning models to integrate clinical data with broad omics technologies, have efficiently delineated complex patient phenotypes and their associations with clinical outcomes.22,23 Accordingly, we leveraged a highly phenotyped discovery cohort of women with early-stage HER2+ breast cancer who underwent rigorous cardiovascular surveillance24,25 to address 2 key objectives: 1) to identify the association between pre–cancer therapy or early treatment changes in biomarkers (ie, measures of cardiac, inflammatory, and endothelial activation and damage and whole-transcriptome plasma miRNA sequencing) with the development of CTRCD; and 2) to evaluate the relative prognostic performance of patient-specific clinical factors, pretreatment cardiac imaging parameters, and circulating biomarkers for CTRCD risk prediction using machine learning approaches. Key findings were independently validated in a separate cohort of women with early-stage HER2+ breast cancer treated with anthracyclines and trastuzumab. Methods Synopsis of study design, participant demographics, and biospecimen collection The discovery cohort represents a secondary analysis of a prospectively recruited longitudinal cohort study (EMBRACE-MRI [Evaluation of Myocardial Changes During Breast Adenocarcinoma Therapy to Detect Cardiotoxicity Earlier With MRI]; NCT02306538) enrolling 136 consecutive patients with early-stage (stages I-III) HER2+ breast cancer receiving sequential therapy with anthracyclines and trastuzumab with or without adjuvant radiotherapy, treated in Toronto, Canada, between November 2013 and July 2019 (inclusion and exclusion criteria are listed in Supplemental Table I).24,25 All participants were ≥18 years of age and provided written informed consent in accordance with protocols approved by the research ethics board of the University Health Network (13-6543). Patients were assessed before chemotherapy (baseline; time point 1 [TP1]) and at 2 to 3 months (postanthracycline, pretrastuzumab; TP2), 6 months (3 months into trastuzumab; TP3), 9 months (6 months into trastuzumab, postradiation; TP4), 12 months (9 months into trastuzumab; TP5), and 15 months (post-trastuzumab; TP6) during cancer therapy (Figure 1A). Echocardiography with 2-dimensional (2D) and 3-dimensional (3D) images for assessment of left ventricular function was performed at each visit. Cardiac magnetic resonance imaging (CMR) was performed at each visit except at TP5. Baseline clinical history was obtained at enrollment, and comprehensive clinical assessments were conducted at each visit (Supplemental Table II). Blood samples were collected by venipuncture for standard of care clinical laboratory tests and for biobanking at each TP. Additional details on cardiac imaging protocols, imaging postprocessing, and biospecimen collection and handling are provided in the Supplemental Appendix. Circulating Cardiac Biomarkers Do Not Distinguish Patients Who Developed CTRCD (A) Study schema showing timing of cardiac imaging and biospecimen collection in the EMBRACE-MRI (Evaluation of Myocardial Changes During Breast Adenocarcinoma Therapy to Detect Cardiotoxicity Earlier With MRI) discovery cohort (n = 136). (B) Longitudinal changes in left ventricular ejection fraction (LVEF), as assessed using cardiac magnetic resonance imaging (CMR). Thirty-seven patients developed cancer therapy–related cardiac dysfunction (CTRCD) during the study period on the basis of CMR-defined criteria established by the Cardiac Review and Evaluation Committee, while 99 patients remained free of dysfunction. Values are presented as geometric means with 95% CIs. (C) Circulating cardiac biomarkers were quantified at baseline and early-treatment time points (TP2 and TP3). High-sensitivity cardiac troponin I (hs-cTnI) and B-type natriuretic peptide (BNP) were measured in the full cohort (n = 37 CTRCD, n = 99 no-CTRCD), while growth differentiation factor-15 (GDF-15) was measured in a subset of patients (n = 27 per group). Violin plots display the median (Q1-Q3). Asterisks below the x-axis indicate significant intracohort changes from baseline, while those above the data points denote significant differences between the CTRCD and no-CTRCD groups at each corresponding TP. Statistical comparisons were performed using 2-way analysis of variance with Šidák’s post hoc test (LVEF) or the Kruskal-Wallis test with Dunn’s multiple-comparisons test (hs-cTnI, BNP, and GDF-15). ∗P < 0.05, ∗∗P < 0.01, and ∗∗∗P < 0.001. The validation cohort comprised 38 patients with early-stage (stages I-III) HER2+ breast cancer receiving therapy with anthracyclines and trastuzumab with or without adjuvant radiotherapy, in Toronto (Supplemental Table III). The patients were from the following studies: EMBRACE-2 (14 patients; NCT02306538), SPARE-HF (Statins for the Primary Prevention of Heart Failure in Patients Receiving Anthracycline Pilot Study) (17 patients; NCT03186404),26 and the Toronto arm of SUCCOUR (Strain Surveillance During Chemotherapy for Improving Cardiovascular Outcomes) (7 patients; ACTRN12620001094965).27 Within each study, we identified patients who had high-quality, biobanked biospecimens suitable for our analyses; for the SUCCOUR cohort, we further restricted inclusion to those who did not undergo strain-based intervention during follow-up, so as not to mask the detection of CTRCD events. The follow-up and echocardiographic imaging protocols in these studies were identical to those of the EMBRACE-MRI study, with 2D and 3D echocardiography and collection of blood performed at each clinical visit. CTRCD categorization Adjudication of CTRCD in the discovery cohort was based on CMR using the criteria established by the Cardiac Review and Evaluation Committee (CREC), defined as 1) a ≥5% absolute reduction in left ventricular ejection fraction (LVEF) from baseline to an LVEF <55% with signs or symptoms of heart failure (n = 2); 2) a ≥10% absolute reduction in LVEF from baseline to <55% without accompanying signs or symptoms of heart failure (n = 33); or 3) a decline in LVEF of >10% in patients with baseline LVEFs <55% (n = 2).28 For sensitivity analyses, we defined CTRCD using these LVEF criteria with 2D and 3D echocardiography (n = 32) and also using the European Society of Cardiology moderate CTRCD criteria on the basis of LVEF (≥10% decline in LVEF to <50% using 2D or 3D echocardiography [n = 15] and CMR [n = 16] separately). For the validation cohort, only 2D and 3D echocardiographic data were available across studies, and CTRCD was defined by CREC (n = 10) and European Society of Cardiology (n = 9) criteria, separately. Additional details are provided in the Supplemental Appendix. Laboratory assessment High-sensitivity cardiac troponin I and B-type natriuretic peptide were measured at baseline and at every follow-up visit. The limits of detection for the high-sensitivity cardiac troponin I and B-type natriuretic peptide assays were 2 and 10 pg/mL, respectively, using the Architect i2000SR immunoassay analyzer (Abbott Core Laboratory); these measures were performed in a centralized core laboratory. Cytokine and chemokine profiling and endothelial protein biomarker assay Cytokine levels in plasma samples were measured using the Luminex 30-Plex Magnetic Bead Panel Immunology Cytokine/Chemokine Assay (EMD Millipore) on a Luminex 100/200 System according to the manufacturer’s instructions. Additional details are provided in the Supplemental Appendix. In parallel, myeloperoxidase (MPO) and GDF-15 were colorimetrically measured with Quantikine enzyme-linked immunosorbent assays (catalog numbers DMYE00B and DGD150, respectively, R&D Systems), according to the manufacturer’s instructions. Circulating levels of angiopoietin-2 (ANGPT2; lower limit of quantification 9.91 pg/mL), soluble CD62 antigen-like family member E (sE-selectin; lower limit of quantification 4.22 pg/mL), endothelin-1 (ET-1; lower limit of quantification 0.250 pg/mL), and soluble CD105/endoglin (ENG; lower limit of quantification 21.8 pg/mL) were quantified in platelet-reduced plasma samples as before21 using the Simple Plex Ella (ProteinSimple) multiplex platform according to the manufacturer’s instructions. All Simple Plex values are reported as the mean of three readings. HTG EdgeSeq miRNA Whole Transcriptome Assay from plasma MiRNAs were quantified in 30 μL of platelet-reduced plasma using the HTG EdgeSeq miRNA Whole Transcriptome Assay (HTG Molecular). Details are provided in the Supplemental Appendix. Blinding procedures for molecular analyses Blinding was performed at the stage of assay initiation where possible (ie, loading of samples and sequencing). Data visualization and statistical analysis Descriptive analysis Clinical characteristics were analyzed using summary statistics. Continuous variables are presented as median (Q1-Q3) or mean ± SD, and dichotomous or polytomous variables are described using frequencies and percentages. Between-group differences were evaluated using 2-sample Student’s t-tests or Mann-Whitney U tests for continuous variables and chi-square or Fisher exact tests for categorical variables (as appropriate). Biomarker analysis Normality was assessed using the D’Agostino-Pearson test. The unpaired Student’s t-test was used for 2-group comparisons, and 2-way analysis of variance analysis with Šidák’s post hoc test was used when multiple groups across several TPs were compared. If distributions were non-normal, we used the Mann-Whitney U test for the analysis of 2 groups and the Kruskal-Wallis test with Dunn’s post hoc test for multiple-group comparisons. Correction for multiple testing was applied to the multiplex enzyme-linked immunosorbent assay data using the Benjamini-Hochberg false discovery rate post hoc method. No corrections for covariates were performed. Associations of 2 continuous variables were analyzed using Spearman correlation matrices and linear regression analysis. Data were analyzed using Prism version 10.0.0 for MacOS (GraphPad Software). P values <0.05 were considered to indicate statistical significance. MiRNA analysis Differential expression was determined using DeSeq2,29 which corrects for multiple comparisons using the Benjamini-Hochberg false discovery rate post hoc method (eg, Q value). MiRNA target pathway analysis was conducted using the Kyoto Encyclopedia of Genes and Genomes database (DIANA-miRPath)30,31 and tested for enrichment using a hypergeometric test with adjustment for multiple comparisons (Benjamini-Hochberg false discovery rate). Risk assessment using machine learning We performed leave-one-out cross-validation of the discovery cohort data to build models to predict CTRCD and assess performance. Categorical features were hot encoded, with missing variables recorded as additional categorical variables, having −1.0 for numerical features. For each experiment, a random decision forest model was fit to the training data set and evaluated on the independent left-out sample. Model performance was assessed using the area under the receiver-operating characteristic (AUROC) curve calculated over the entire testing set. Average AUROC curves and 95% CIs were calculated using the percentile method with bootstrapping over the entire testing set (repeated 1,000 times). The specificity of each model was assessed at the model threshold that achieved ≥90% sensitivity. Feature importance was estimated using Shapley additive explanations values.32 Available clinical, CMR, and biomarker (protein and miRNA) data that were measured across all patients at baseline in the discovery cohort were included in the model (see Supplemental Table IV for a list of the features included and their definitions). MiRNA data were normalized using the median-ratio method.33 To reduce the dimensionality of the miRNA data, the mean expression of the entire data set was calculated, and the top 10% of expressed miRNAs were selected for further analysis (reducing the total number to 208; Supplemental Table V). Data missingness is indicated in Supplemental Table VI. Analysis of the performance of the model was assessed using protein biomarker data from a validation cohort. Data in the validation cohort were processed using the same procedures as the discovery cohort with regard to data missingness and normalization. Model outputs per patient in the validation cohort were averaged across the 136 models (from leave-one-out cross-validation), thus retaining the same model set. Results Patient demographics and identification of patients with CTRCD in the discovery cohort We included 136 women (mean age 51.1 ± 9.2 years) undergoing therapy for HER2+ breast cancer (Supplemental Tables I and II) in the discovery cohort. The lowest LVEF for the whole cohort was reached 3 months into trastuzumab therapy (ie, TP3) (Figure 1B); 37 patients (27.2%) developed CMR-defined CTRCD by CREC definitions. Baseline demographics and laboratory values were similar between those who did and did not develop CTRCD (Supplemental Table II), except that patients who went on to develop CTRCD exhibited a higher prevalence of hyperlipidemia, elevated resting heart rate, and modestly reduced feature-tracking strain (Supplemental Table II). Lack of distinguishing cardiac biomarkers in patients who developed CTRCD At baseline, median high-sensitivity cardiac troponin I concentrations were at the lower limit of assay detection in both patients who developed CTRCD and those who did not. Although high-sensitivity cardiac troponin I and GDF-15 levels increased across the entire cohort during treatment (TP2 and TP3), none of the cardiac biomarkers demonstrated statistically significant differences between the CTRCD and no-CTRCD groups at any TP (Figure 1C). Patients who developed CTRCD had elevated circulating levels of several inflammatory and endothelial-centric markers before and early during treatment An exploratory analysis was conducted using a panel of 35 biomarkers related to immune activation, angiogenesis, and endothelial activation. Because of the extensive number of analytes and assay costs, a randomly selected subset of patients was analyzed at baseline, TP2, and TP3. Several circulating immune and angiogenic markers, including granulocyte colony-stimulating factor, interleukin-10, interleukin-15, interferon γ–induced protein-10, monocyte chemoattractant protein-1, macrophage inflammatory protein-1β, tumor necrosis factor-α, and vascular endothelial growth factor A, were significantly elevated in the overall cohort during treatment (TP2 and/or TP3) (Figure 2A, Supplemental Figure I, Supplemental Table VII). Notably, at baseline, patients who subsequently developed CTRCD exhibited significantly higher levels of interferon-α, interleukin-10, interferon γ–induced protein-10, and MPO and lower levels of interleukin-1β compared with those who did not develop CTRCD (Figure 2A, Supplemental Figure I). Among these, interferon γ–induced protein-10 remained elevated throughout treatment (TP2 and TP3) in the CTRCD group. Of the endothelial activation markers, ANGPT2, sE-selectin, ENG, and ET-1 were all robustly increased during treatment in the overall cohort (ie, TP2 and TP3) (Figure 2B). Among these markers, 2 were significantly elevated at baseline between the CTRCD and no-CTRCD groups (ie, ANGPT2 and ENG) (Figure 2B), and these markers remained elevated at early treatment TPs. In contrast, sE-selectin and ET-1 were elevated during therapy (TP2/TP3 and TP3, respectively) in patients who developed CTRCD (Figure 2B). All endothelial-centric markers (ANGPT2, ENG, ET-1, and sE-selectin), cardiac markers (cardiac troponin I, B-type natriuretic peptide, and GDF-15), along with MPO, were subsequently measured in all pretreatment baseline samples from the discovery cohort, which revealed a significant elevation in the levels of ANGPT2, ENG, ET-1, and MPO in patients who later developed CTRCD (Figure 3A). ANGPT2 levels were correlated with MPO and ENG in baseline samples (Supplemental Figures IIA and IIB). Additionally, ANGPT2, MPO, and ENG levels at baseline were correlated with the decline in CMR LVEF during treatment (Figure 3B). Circulating Inflammatory and Endothelial Activation Biomarkers Distinguish Patients Who Developed CTRCD (A) The circulating inflammatory biomarkers myeloperoxidase (MPO), interferon-α, interferon γ–induced protein-10 (IP-10), and interleukin-1β (IL-1β) were measured at baseline and during early treatment (TP2 and TP3) in the discovery cohort. MPO was assessed in a subset of patients (n = 27 CTRCD, n = 27 no-CTRCD), while interferon-α, IP-10, and IL-1β were measured in a broader group (n = 25 CTRCD, n = 54 no-CTRCD). Additional inflammatory markers are presented in Supplemental Figure I, with multiple testing correction results detailed in Supplemental Table VII. (B) The endothelial activation markers angiopoietin-2 (ANGPT2), soluble CD62 antigen-like family member E (sE-selectin), endothelin-1 (ET-1), and soluble CD105/endoglin (ENG) were evaluated at the same TPs in a larger subset (n = 37 CTRCD, n = 70 no-CTRCD). Violin plots represent the median (Q1-Q3). Asterisks below the x-axis denote significant intracohort changes from baseline, while asterisks above the plots indicate significant differences between the CTRCD and no-CTRCD groups at each TP. Statistical comparisons were performed using 2-way analysis of variance with Šidák’s post hoc test (all data except MPO) or the Kruskal-Wallis test with Dunn’s multiple-comparisons test (MPO). ∗P < 0.05, ∗∗P < 0.01, and ∗∗∗P < 0.001. Abbreviations as in Figure 1. Analysis of Pretreatment Levels of Inflammatory and Endothelial Activation Biomarkers Distinguishes Those at Risk for CTRCD (A) Baseline concentrations of ANGPT2, sE-selectin, ENG, ET-1, MPO, GDF-15, hs-cTnI, and BNP were measured in all patients in the discovery cohort (n = 37 CTRCD, n = 99 no-CTRCD). The violin plot indicates median (Q1-Q3). ∗P < 0.05, ∗∗P < 0.01, and ∗∗∗P < 0.001 (unpaired Mann-Whitney U test). (B) Baseline concentrations of ANGPT2, MPO, and ENG plotted as a function of the change in CMR LVEF during cancer treatment. Simple linear regression with Spearman correlation coefficients is shown. Abbreviations as in Figures 1 and 2. Sensitivity analyses using alternative CTRCD definitions, including the CREC criteria and the European Society of Cardiology’s moderate CTRCD threshold on the basis of LVEF from CMR or echocardiography, further supported these findings. ANGPT2 and MPO remained significantly elevated at baseline in patients who developed CTRCD across all alternative definitions (Supplemental Figure IIC), and ENG was also elevated according to the CREC criteria using 2D and 3D echocardiographic LVEF. Assessment of pretreatment circulating plasma miRNA expression identifies distinct signatures in patients that developed CTRCD Measurement of >2,000 plasma miRNAs in all baseline samples in the discovery cohort highlighted 119 differentially expressed miRNAs (n = 99 up-regulated and n = 20 down-regulated) between those who subsequently did and did not develop CTRCD (Figure 4, Supplemental Table V). Pathway enrichment analysis of predicted messenger RNA targets of up-regulated miRNAs in the CTRCD group, performed using DIANA-miRPath, revealed significant over-representation of pathways involved in cardiomyocyte biology, such as ErbB2 signaling and arrhythmogenic right ventricular cardiomyopathy, as well as proinflammatory signaling pathways, including transforming growth factor-β and Ras signaling (Figure 4). In contrast, predicted targets of miRNAs down-regulated in the CTRCD group were enriched for pathways related to cellular metabolism (Figure 4). Plasma miRNAs Are Differentially Abundant at Baseline in Patients Who Developed CTRCD Volcano plot showing differentially expressed plasma microRNAs (miRNAs) in the entire discovery cohort (n = 136) at baseline in patients who developed cancer therapy–related cardiac dysfunction (CTRCD) compared with those who did not. The x-axis indicates log2 fold change, and the y-axis represents −log10(Q value). Dashed lines denote a ±1.5-fold change threshold. MiRNAs with Q values <0.05 are shown in red (up-regulated) or blue (down-regulated); selected miRNAs are labeled. Right panels show Kyoto Encyclopedia of Genes and Genomes pathway enrichment analysis of predicted messenger RNA targets of differentially expressed miRNAs. Pathways enriched among up-regulated miRNAs (in CTRCD) are shown in red; those enriched among down-regulated miRNAs are shown in blue. Enrichment scores are presented as −log10(Q value). miR = microRNA; TCA = tricarboxylic acid; TGF = transforming growth factor. Machine learning identifies predictive pretreatment biomarkers of CTRCD Baseline clinical variables, CMR parameters, protein biomarkers, and plasma miRNA expression profiles, available in all patients from the discovery cohort, were used to train a random forest classifier for prediction of CTRCD. Model performance, evaluated using leave-one-out cross-validation with all feature groups, demonstrated strong discriminatory power, with an AUROC curve of 0.885 (95% CI: 0.816-0.943) and specificity of 0.656 (95% CI: 0.303-0.859) (Figure 5A). Feature attribution analysis using Shapley additive explanations revealed that protein biomarkers had the greatest influence on model output, followed by CMR-derived strain parameters, miRNA expression, and clinical features (Figure 5B, Supplemental Figure IIIA). Among individual predictors, ANGPT2, MPO, and ENG were the top-ranked features associated with CTRCD risk (Figure 5B). Notably, CMR-derived LV global longitudinal and circumferential strain, as well as miRNAs such as miR-6727-5p and miR-1915-3p, were also among the top 10 features contributing to model prediction (Figure 5B, Supplemental Figure IIIA). Models trained on clinical data alone, CMR features alone, miRNAs alone, or combined clinical and CMR features yielded lower AUROC scores (Figure 5A, Supplemental Figures IIIB to IIID). In contrast, a model incorporating clinical, CMR, and protein biomarkers achieved improved performance (AUROC curve, 0.970 [95% CI: 0.942-0.990]; specificity, 0.894 [95% CI: 0.778-0.970]) (Figure 5A, Supplemental Figure IIID). Notably, a model using only protein biomarkers (ANGPT2, ENG, ET-1, sE-selectin, GDF-15, and MPO) outperformed all other models, achieving an AUROC curve of 0.978 (95% CI: 0.956-0.994) and specificity of 0.932 (95% CI: 0.818-0.990). A simplified model using only the top 3 protein features (ANGPT2, MPO, and ENG) also performed strongly, with an AUROC curve of 0.972 (95% CI: 0.948-0.991) and specificity of 0.878 (95% CI: 0.768-0.990) (Figure 5A). Visualization of pairwise and 3D combinations of ANGPT2, MPO, and ENG concentrations revealed clear separation between the CTRCD and no-CTRCD groups (Supplemental Figures IVA and IVB). Shapley additive explanations dependence plots further identified inflection points in the concentrations of these biomarkers associated with marked changes in model output (Supplemental Figure IVC). Importantly, baseline model–derived risk scores were correlated with the maximal decline in CMR-derived LVEF during treatment (Supplemental Figure VA). Finally, sensitivity analyses using alternative definitions of CTRCD, including the CREC criteria on the basis of 2D and 3D echocardiography and the European Society of Cardiology moderate CTRCD criteria, demonstrated that both the “protein only” and “ANGPT2+MPO+ENG” models retained robust predictive capability, with AUROC curve values ranging from 0.794 to 0.843 (Figure 5C, Supplemental Figures VB and VC). Machine Learning Identifies Features That Predict CTRCD Risk From Baseline Data (A) Receiver-operating characteristic curves with areas under the curves and 95% CIs for leave-one-out cross-validation of random forest models trained on baseline data from the discovery cohort (n = 136; 37 with CTRCD defined by the Cardiac Review and Evaluation Committee [CREC] using CMR). Models incorporating clinical, CMR, protein, and miRNA features are shown alone or in combination. Protein biomarkers are sE-selectin, ANGPT2, ENG, ET-1, GDF-15, and MPO. Shown are the specificities and areas under the curves with 95% CIs obtained via bootstrapping. The specificity of each model was assessed at the model threshold that achieved ≥90% sensitivity. (B) Shapley additive explanations (SHAP) value summary plot illustrating the relative contribution of individual variables to model predictions. Feature types are color coded as follows: protein (red), CMR (blue), miRNA (orange), and clinical (green). SHAP value contributions are also color mapped, with red indicating higher feature values and blue indicating lower values. (C) Sensitivity analysis using the European Society of Cardiology (ESC) recommendations with 2-dimensional (2D) and 3-dimensional (3D) echocardiography (≥10% decline in LVEF to <50%) to define CTRCD in the discovery cohort. Areas under the curves with 95% CIs are shown. (D) Validation of the “protein only” and “ANGPT2+MPO+ENG” models in a separate validation cohort of 38 early-stage HER2+ breast cancer patients using CREC or ESC criteria with echocardiography-derived 2D/3D LVEF to define CTRCD. Areas under the curves with 95% CIs are shown. GCS = global circumferential strain; GLS = global longitudinal strain; LV = left ventricular; LVESV = left ventricular end-systolic volume; MRI = magnetic resonance imaging; other abbreviations as in Figures 1, 2, and 4. Validation of predictive biomarkers To validate biomarker findings from the discovery cohort, baseline plasma levels of the 6 proteins (ie, ANGPT2, ENG, ET-1, sE-selectin, GDF-15, and MPO) constituting both the broader “protein only” model and its “ANGPT2+MPO+ENG” subset, were measured in an independent validation cohort comprising 38 patients with early-stage HER2+ breast cancer treated with anthracyclines and trastuzumab. Baseline demographic, clinical, and cardiac characteristics were largely comparable between cohorts, with significant differences observed only in cancer therapies, attributable to variations in trial protocols (Supplemental Table VIII). Applying the CREC criteria, 10 patients (26.3%) developed CTRCD, while 9 patients (23.7%) met the definition using the European Society of Cardiology moderate CTRCD criteria. Using the CREC definition, the full “protein only” model yielded an AUROC curve of 0.740 (95% CI: 0.521-0.932) and specificity of 0.679 (95% CI: 0.485-0.808), while the 3-marker subset performed similarly (AUROC, 0.725 [95% CI: 0.507-0.918]; specificity, 0.679 [95% CI: 0.495-0.818]) (Figure 5D, right). Under the European Society of Cardiology criteria, the “protein only” model achieved an AUROC curve of 0.780 (95% CI: 0.532-0.962) with specificity of 0.690 (95% CI: 0.505-0.818); the 3-marker subset again demonstrated comparable performance (AUROC, 0.777 [95% CI: 0.559-0.954]; specificity, 0.690 [95% CI: 0.505-0.828]) (Figure 5D, left). Discussion Despite decades of research, predicting CTRCD risk before or early during cancer therapy remains challenging. Circulating biomarkers can aid risk stratification during treatment, but no robust, validated options exist for pretreatment use. In the discovery EMBRACE-MRI cohort, several inflammatory (eg, MPO, interferon γ–induced protein-10, interferon-α) and endothelial (eg, ANGPT2, ENG, ET-1) activation markers were elevated at baseline, suggesting the possible presence of a pre-existing subclinical vascular inflammatory state that may increase susceptibility. In contrast, traditional cardiac biomarkers such as high-sensitivity cardiac troponin I, B-type natriuretic peptide, and GDF-15 were not predictive. Baseline whole-transcriptome miRNA profiling further revealed distinct expression patterns in patients who developed CTRCD, involving cardiomyocyte function and inflammatory pathways. These signatures aligned with the protein-level inflammatory dysregulation and point to possible early disruption of cardiovascular signaling before therapy. Although prior studies have examined miRNAs in CTRCD, their mechanistic role vs their value as sentinels of disease remains unclear.34 To integrate transcriptomic, protein, imaging, and clinical data, we applied a random forest machine learning model. Baseline ANGPT2, MPO, and ENG emerged as the strongest predictors, with smaller contributions from other protein markers, select miRNAs, and CMR parameters (Central Illustration). These associations were consistent across alternative CTRCD definitions, with the use of echocardiography-derived LVEF to detect CTRCD, and in an independent validation cohort. Notably, the validation cohort relied solely on echocardiography-based metrics rather than CMR, which may have contributed to the wider CIs observed. Pretreatment Vascular Signatures in Breast Cancer Patients Elucidate Predispositions to CTRCD Clinical parameters, cardiac imaging, and circulating biomarkers (protein and miRNA) were measured in 136 HER2+ breast cancer patients in the discovery cohort before cancer therapy (anthracyclines and trastuzumab). Features at baseline that distinguished cancer therapy–related cardiac dysfunction (CTRCD) from no-CTRCD patients are indicated. All available baseline data were used to develop a random forest machine learning model to predict CTRCD. Myeloperoxidase (MPO), angiopoietin-2 ANGPT2, and soluble CD105/endoglin (ENG) levels at baseline were found to contribute most to model performance, which was validated in an additional validation cohort of 38 HER2+ breast cancer patients. CTRCD incidence and the area under the receiver-operating characteristic curve (AUC; with 95% CI) are shown. αINF = interferon-α; CMR = cardiac magnetic resonance imaging; CREC = Cardiac Review and Evaluation Committee; EMBRACE-2 = Evaluation of Myocardial Changes During Breast Adenocarcinoma Therapy to Detect Cardiotoxicity Earlier With MRI - 2; EMBRACE MRI = Evaluation of Myocardial Changes During Breast Adenocarcinoma Therapy to Detect Cardiotoxicity Earlier With MRI; ET-1 = endothelin-1; IP-10 = interferon γ–induced protein-10; miR = microRNA; SPARE-HF = Statins for the Primary Prevention of Heart Failure in Patients Receiving Anthracycline Pilot Study; SUCCOUR = Strain Surveillance During Chemotherapy for Improving Cardiovascular Outcomes. Our results align with evidence implicating vascular injury as a key driver of CTRCD and other cardiovascular toxicities from cancer therapy.10,12 Anthracyclines, trastuzumab, and radiation are known to impair macro- and microvascular function.35, 36, 37 Elevated baseline and early-treatment levels of the neutrophil activation marker MPO and the endothelial coagulation regulator thrombomodulin have been linked to CTRCD,6,13 and our findings further support the association between inflammation, endothelial activation, and CTRCD. Notably, baseline clinical variables were not strongly associated with risk in our discovery cohort, raising questions about the origin of this inflammatory state. One possibility is tumor-driven inflammation, triggered by malignancy-derived factors or systemic immune responses, consistent with evidence that tumor characteristics, immune phenotypes, and microenvironments vary among individuals.38,39 Circulating inflammatory biomarkers may reflect a cancer-associated inflammatory phenotype that diminishes during anthracycline therapy as tumor burden decreases, the microenvironment changes, or immune activation shifts, accounting for the disappearance of group differences in inflammatory markers (eg, MPO, interferon-α) at later TPs. At the same time, several endothelial markers (eg, ANGPT2, sE-selectin, ENG, ET-1) remained elevated or increased, suggesting progressive vascular injury during treatment. This suggests that a pre-existing endothelial activation primed state may heighten susceptibility to treatment-related injury, a mechanism underexplored in cancer therapy. Within this framework, MPO, ANGPT2, and ENG emerged as the most informative biomarkers in our analyses, each with established links to vascular pathology and cardiac dysfunction. MPO, a neutrophil- and monocyte-derived enzyme involved in oxidative stress and inflammation, has been previously linked to CTRCD in multiple studies.6,8,13 Preclinical evidence further supports its pathogenic role, with murine models demonstrating MPO-mediated cardiac injury in the context of anthracycline exposure.40 ANGPT2, although not previously implicated in CTRCD, is an endothelium-derived regulator of angiogenesis and vascular inflammation.41 In nononcology populations, elevated circulating ANGPT2 levels have been associated with impaired ventricular function, adverse cardiac remodeling, and increased all-cause mortality.42,43 Similarly, ENG, expressed by activated endothelial cells, modulates proangiogenic and proinflammatory signaling pathways.44 Circulating levels have been linked to left ventricular dysfunction,45 predict cardiovascular events in patients with coronary artery disease,46 and contribute mechanistically to maladaptive remodeling. Notably, ENG expression in cardiac fibroblasts is elevated in human heart failure tissue, and ENG neutralization improves cardiac survival in murine models of pressure overload.47 Further mechanistic studies are warranted to clarify the causal role of these biomarkers and to determine whether they may serve not only as predictors of risk but also as therapeutic targets. Study limitations The results of this study should be viewed in the context of its design, cancer type (early-stage HER2+ breast cancer), treatment protocol (trastuzumab following anthracyclines rather than nonanthracycline, dual HER2 monoclonal antibodies or other HER2-targeted agents), sample size, and sociogeographic context (ie, access to advanced cardiac imaging, socioeconomic status, and race). Although we were able to validate the findings in a small cohort of early-stage HER2+ breast cancer patients treated with anthracyclines and trastuzumab, the relevance of these findings to advanced breast cancer and other types of cancers or treatments, remains to be tested. Future studies should validate these findings in larger, more diverse, multicenter cohorts of HER2+ breast cancer patients and in those with other cancer types, alternative chemotherapy regimens (including nonanthracycline protocols), and broader clinical settings. The miRNA findings should be validated using other techniques and cohorts, as combinations of baseline plasma miRNAs (although not top individual predictors) may still indicate CTRCD risk with high specificity. Conclusions Using a machine learning approach, we found that a model incorporating baseline levels of ANGPT2, MPO, and ENG was able to discriminate early-stage HER2+ breast cancer patients at high risk for CTRCD during combination treatment with anthracyclines and trastuzumab. The model’s predictive performance was confirmed in an independent validation cohort of HER2+ breast cancer patients treated with anthracyclines and trastuzumab. These findings suggest that pre-existing systemic inflammation and endothelial activation may confer vulnerability to cardiac injury during cancer treatment. Although further validation in larger and more diverse populations is needed, our results underscore the potential utility of non-cardiac-centric biomarkers in risk prediction for CTRCD. Early identification of high-risk patients before therapy could enable personalized cancer treatment, risk-adapted cardiac monitoring, and enhanced risk reduction strategies. For instance, these biomarkers might guide prevention trials using anti-inflammatory or endothelium-stabilizing therapies, while identifying low-risk patients could expand treatment options and reduce the need for intensive surveillance, minimizing delays and improving care. Data and resource sharing statement All supporting data are available within the paper and its supplemental data files. Further information and requests for resources, reagents, or patient-level data sets may be sent to the corresponding authors. COMPETENCY IN MEDICAL KNOWLEDGE: In early-stage HER2+ breast cancer patients treated sequentially with anthracyclines and trastuzumab pretreatment levels of ANGPT2, MPO and ENG measured before cancer therapy were found to be the strongest predictors of CTRCD during treatment. TRANSLATIONAL OUTLOOK: The identified biomarkers provide mechanistic insights and with further study may enable pretreatment risk stratification for CTRCD in patients with HER2+ breast cancer treated with anthracyclines and trastuzumab and help inform subsequent surveillance strategies. Future work should focus on validating these biomarkers in larger, multicenter cohorts and in diverse cancer populations and treatment regimens. Funding Support and Author Disclosures The preparation of this paper used resources provided by the Toronto General Hospital Research Institute, University Health Network. Dr Gustafson was funded by a Frederick Banting and Charles Best Canada Graduate Scholarship Doctoral Award, a Ted Rogers Centre for Heart Research Studentship, and an Ontario Graduate Scholarship. Dr Mistry was funded by a Vanier Canada Graduate Doctoral Award from the Canadian Institutes of Health Research. Dr Ching was funded by a Ted Rogers Centre for Heart Research Studentship and an Ontario Graduate Scholarship. Research in the laboratory of Dr Fish was supported by project grants from the Canadian Institutes of Health Research (PJT148487, PJT175301 and PJT198168), a Canada Foundation for Innovation Grant, the John R. Evans Leaders Fund, and the Ontario Research Fund. Dr Fish was supported by a Tier 2 Canada Research Chair in Vascular Cell and Molecular Biology. Cardio-oncology research in the laboratory of Dr Thavendiranathan was supported by the Peter Munk Cardiac Centre Innovation Grant Program, a Ted Rogers Centre for Heart Research Grant, a Heart and Stroke Foundation Project Grant-in-Aid (G-20-0029537), and project grants from Canadian Institutes of Health Research (PJT175301 and PJT198168). Dr Thavendiranathan was supported by a Tier 2 Canada Research Chair in Cardio-Oncology (CRC-2019-00097/CRC-2023-00422) and the Canadian Cancer Society/Canadian Institutes of Health Research’s W. David Hargraft grant. Dr Marwick was supported by an investigator grant (2008129) from the National Health and Medical Research Council. The funding sources had no role in the study design, data collection, analysis, interpretation, or manuscript preparation. Drs Gustafson, McIntosh, Thavendiranathan, and Fish are coinventors on a submitted patent related to these findings. All protocols and methods are freely available for academic and nonprofit research use. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose. Acknowledgments The authors thank the members of the Peter Munk Cardiac Centre Biobank, in particular Filio Billia, Katherine Tsang, Fraser Amos, and Hyeja Kim, for their support in biospecimen curation. The authors are also grateful to the participants who contributed their biospecimens and the clinical staff members involved in collection. Footnotes The authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the Author Center. Appendix For nonstandard abbreviations and acronyms, a supplemental Methods section, and supplemental tables, figures, and supplemental references, please see the online version of this paper. Contributor Information Paaladinesh Thavendiranathan, Email: dinesh.thavendiranathan@uhn.ca. Jason E. Fish, Email: jason.fish@utoronto.ca. Appendix Supplementary Material Supplementary Tables 4-6 References Associated Data Supplementary Materials Supplementary Material Supplementary Tables 4-6