Development and Validation of a Machine-Learning Deep Plasma Proteome Classifier for Early-Stage Breast Cancer Detection Authors: Alec Horrmann, Yash Travadi, Jacob Carey, Ella Boytim, Kevin Mallery, Grant Schaap, Carissa Rungkittikhun, Kaylee Judith Kamalanathan, Nathaniel R Bristow, Catalina Galeano-Garces, Adam Groth, Harrison Ball, Alexa R Hesch, Pooja Advani, Justin Hwang, Badrinath R Konety, Justin M Drake Published: 2026 ABSTRACT AbstractIntroductionProteome-guided liquid biopsy tests hold immense promise for the future of early cancer detection. Here, we analyzed the plasma proteome of 1,259 biobanked samples consisting of healthy women and women with breast cancer. The Astrin Biosciences’ breast cancer early detection test is a laboratory developed test (LDT) that uses a protein-based machine learning classifier to identify breast cancer with high accuracy.MethodsThe classifier was trained on 845 women (466 healthy and 379 with newly diagnosed, treatment naïve breast cancer) and validated on 397 women (195 healthy and 202 breast cancer) from the same collection cohort (held-out validation). All plasma samples were processed in an automated, blinded manner coupled with semi-quantitative, label-free mass spectrometry (MS)-based analysis.ResultsThe held-out validation performance achieved 92.3% specificity (180/195; 95% Wilson CI: 87.7–95.3%), 92.6% sensitivity (187/202; 95% Wilson CI: 88.1–95.4%) and an AUC of 0.975 (95% Bootstrap CI: 0.961–0.987). Observed sensitivity remained high across all breast cancer stages and pathological and molecular subtypes, albeit with small sample sizes for some subtypes. Gene set enrichment analyses (GSEA) identified epithelial-to-mesenchymal transition (EMT) and PI3K-AKT signaling as enriched in the breast cancer samples, highlighting that our test may possibly identify cancer-related proteins in early-stage patients. A simulated population demonstrates the utility of our test as a supplement to mammography, detecting nearly all (93%) breast cancers missed by mammography and reducing the number of false positives relative to MRI and Contrast-Enhanced Mammography (CEM) alone by >10-fold.DiscussionOverall, our proteomic data demonstrates high sensitivity and specificity in women with breast cancer, especially at early stages, and is a favorable supplemental test post mammogram.Keywords: liquid biopsy, proteomics, mass spectrometry, dense breast tissue, early detection, breast cancer Introduction Breast cancer remains the most diagnosed cancer among women globally, with approximately 2.3 million new cases and 670,000 deaths recorded in 2022, and projections estimating a rise to 3.2 million cases and 1.1 million deaths annually by 2050.1,2 Mammography is the gold standard for population-based breast cancer screening. However, its detection sensitivity is significantly reduced in women with dense breast tissue, who face an elevated risk of developing breast cancer.3–6 Recent updates to the FDA’s mammography regulations7 underscore the increasing focus on optimizing screening strategies in this patient population (representing approximately 50% of the US screening population). MRI and Contrast-Enhanced Mammography (CEM) have recently shown potential as a promising supplement to traditional mammography screens.8 A consensus on the best imaging strategy for individuals with dense breasts has not been established. Despite the expanding role of functional imaging such as PET/CT in breast cancer management, conventional and advanced imaging modalities alike face limitations across disease stages, underscoring the clinical need for complementary molecular-based detection strategies.9 Although supplemental imaging such as breast MRI can improve cancer detection in women with dense breasts, its use is not routinely recommended for average-risk individuals due to increased rates of false positives, over-diagnosis, additional biopsies, concern for contrast related toxicity, patient anxiety, and cost. Therefore, supplemental MRI imaging is often reserved for use in high-risk patients (with a lifetime risk of breast cancer ≥20%). There is currently a lack of sufficient evidence to justify the benefit of use in patients with high breast density who otherwise have a negative screening mammogram,10 creating an opportunity for new supplemental screening tests to help bridge this gap for high-risk patients. The implementation of liquid biopsy-based tests for early cancer detection is poised to be an important component of the supplemental screening landscape, including breast cancer. These tests can help reduce unnecessary imaging or other diagnostic interventions, thus decreasing costs related to current supplemental screening guidelines while increasing detection rates. The low sensitivity for breast cancer at early stages with nucleotide-based tests11 offers opportunities to bring new technologies and approaches, such as proteomics, to at-risk women where supplemental screening would typically be implemented. Proteomics is an emerging field in the clinical diagnostics space with niche uses for blood–based diagnostic testing.12–14 This has typically been through analyses of a single protein or a few proteins to a few hundred proteins, such as the antibody-based affinity platform Olink, for single or multi-cancer detection.15–17 Pre-analytical variability, or fragility, of the plasma proteome and the lack of high throughput assessment of the proteome has hindered significant clinical incorporation. However, recent advancements in mass spectrometry (MS) technology and a better understanding of the pre-analytical nuances around proteome-based assessment have allowed for the incorporation of unbiased (or shotgun) proteomics using (MS)-based technologies for deep proteome analyses (>5,000 proteins). Here, we assessed a large cohort of biobanked samples consisting of 1,259 women to detect breast cancer, as early as stage 0, from healthy controls through liquid biopsy-based deep proteomic profiling via MS as a scale up and validation of our previous work.18 We have further refined and validated our automated processing pipeline to continuously prepare and analyze hundreds of plasma samples with low sample error rates and high reproducibility while obtaining sensitivities near 90% across all stages of breast cancer using our proprietary spectral library with an improved machine learning algorithm. While this workflow is currently limited to breast cancer detection, this pipeline can be broadly adapted to other types of cancer and diseases. Methods Sample Acquisition and Processing Treatment naïve breast cancer (Stages 0-IV) and healthy plasma samples were purchased from two different biobanks (ProteoGenex and BioIVT) and shipped frozen to Astrin Biosciences. Cancer samples were selected based on the following criteria: 1) treatment naïve, 2) no prior history of cancer, and 3) a confirmed diagnosis of breast cancer based on histopathological analysis of subsequent biopsy or resected tissue samples. Healthy female donors were selected to create an age-matched population. The banked plasma samples were collected between 2004 and 2025 (median 2022; interquartile range 2021–2023), with approximately 80% collected from 2020 onwards. Samples were prepared exactly as described by Horrmann et al.18 Briefly, 200μL plasma aliquots, process controls, digestion controls, and positive and negative controls were loaded onto a plate using a Microlab Prep liquid handler (Hamilton). For the controls, we purchased an Enolase digest (Waters), HeLa lysates (Thermo Fisher), and pools of nine different biobanked positives and negatives. Next, each sample received a proprietary functionalized superparamagnetic bead solution. These beads were designed to leave behind high abundance contaminating proteins while maintaining high capture efficiency for lower abundance analytes. Using a Kingfisher Flex (Thermo Fisher) the samples were washed and subsequently transferred to a lysis solution followed by lysis on a thermal shaker. Using the Microlab Prep, the supernatant was then transferred to a 96-well LoBinding plate (Eppendorf, 0030129512). A series of buffer exchanges to remove contaminants were performed, with the last solution being a digestion buffer (pH 8.0). The proteins were digested with Lys-C and trypsin overnight. The following morning, the peptides were removed and dried using a vacuum concentrator (Labconco). Characterization of Breast Cancer Molecular Subtypes Typical subtyping requires pathology reports that include ER/PR/HER2 expression, proliferation information via Ki-67, and histopathology of the tumor. We used the following characterization based on available information shared to us from each biobank on the reporting of hormone status. HER2 was considered positive with staining intensity of 3+: Hormone receptor positive (ie, estrogen receptor or progesterone receptor) (HR+) and HER2 positive (HER2+); HR+ and HER2 negative (HER2-); HR negative (HR-) and HER2+; and HR- and HER2-. Information related to tumor grading and pathological subtypes was received directly from the biobanks. Peptide Quantification and Mass Spectrometry Parameters Prepared peptides were resuspended in 0.1% formic acid in water and quantified using a quantitative fluorometric peptide assay (Thermo Fisher, 23290). EvoTips were prepared as directed by the manufacturer with the exception that solvent volumes were doubled. 250 ng per sample were injected and analyzed on an Orbitrap Astral mass spectrometer (Thermo Fisher) coupled to an EvoSep One liquid chromatography (LC) system (EvoSep) running the whisper zoom 40 SPD with an Ion Opticks column (AUR3-15075C18-XT) set to 1,600 V and default parameters from the Label Free Quantification using Data Independent Acquisition (LFQ DIA) preset with 3 Th windows. Mass Spectrometry Data Analysis Running in data-independent acquisition (DIA) mode, the Orbitrap Astral produces raw mass spectra from which peptide quantities can be computed based on a spectral library. Peptide quantities are then analyzed to produce gene group abundance data that are ultimately used for the classifier. The spectral library used herein was generated from an independent set of 280 patient samples consisting of individuals with breast cancer and healthy controls as previously described.18 All peptide quantification and gene group abundance data were generated from DIA-NN (2.1.0) with default parameters except for the following settings: Maximum mass accuracy fixed to 11 and 10 ppm (MS1 and MS2, respectively), scan window of 7, up to one variable modification of methionine oxidation or N-terminal acetylation. A random subset of 100 patient samples used for spectral library generation was designated as “normalizing samples” for protein quantification. QuantUMS parameters were trained on these 100 samples and fixed for subsequent analysis.19 The samples were each processed in independent DIA-NN sessions alongside all 100 normalizing samples (ie, total of 101 samples) to obtain the MaxLFQ-normalized quantifications.20 The gene group-level abundance data from DIA-NN was preprocessed as follows. Non-missing abundances were log-transformed, and any remaining missing values were imputed using the minimum observed abundance for each gene-group minus one. Finally, the abundances were standardized by gene-group (zero mean and unit variance) to generate features suitable for machine learning classification models. We developed quality control (QC) metrics for both the plate level and the individual sample level. At the plate level, we evaluated mass spectrometry controls that included commercial enolase standards for retention time and total precursor ion intensity. Commercial HeLa cell lysates were evaluated for peptide and protein IDs and selected peptides were assessed for consistent abundance values across the plates, and blank injections were added to compare total ion chromatograms (TIC) between blank and patient samples. Processing controls, commercial yeast lysates, and positive and negative controls were analyzed to ensure all sample processing steps were sufficiently completed. At the sample level, we required that each sample meet minimum peptide and protein IDs. Plates or samples not meeting these QC thresholds were re-ran. Two unsuccessful runs of the same sample resulted in discarding that sample. Reproducibility and Repeatability Analysis For reproducibility experiments, two different operators ran the same samples from 8 donors (4 healthy and 4 breast cancer) over two different days. For repeatability experiments, the first operator ran samples from 4 donors (2 healthy and 2 breast cancer) in triplicate across three different plates, while a second operator ran a different set of samples from 4 donors (2 healthy and 2 breast cancer) in triplicate on a single plate. In total, 16 measurements were run for reproducibility, and 24 measurements were run for repeatability. Performance of each sample was assessed where each of the samples was classified as healthy or breast cancer. If the same sample was classified correctly across operators or days, then the sample was positive. If the samples were classified differently across operators or days, then the sample was negative. Training and Validation Split We randomly assigned 845 of the samples to the training portion and the remaining 397 samples were assigned to the validation portion (a roughly 70:30 split), where their clinical information was blinded from sample preparation and analysis. Validation samples were batched separately from training samples to mimic real world situations around sample acquisition for an early detection test. Seventeen samples that did not pass QC were removed from the analysis. Machine Learning and Cross-Validation Analysis To avoid the potential of confounding by source, we implemented an Exponentiated Gradient method,21 ensembling L2-norm regularized logistic regression, with the constraint of demographic parity across two sources. The default parameters in fairlearn library were used. To make predictions as part of cross-validation and to assess performance on the holdout validation, predictions of each component logistic regression were weighted according to the model weights estimated by the Exponentiated Gradient method. Source was treated as a sensitive feature. We performed five-fold cross-validation of this model to assess the performance of our assay in distinguishing breast cancer from healthy samples in the training portion. At each iteration of the cross-validation, the data pre-processing described above was learned without the held-out fold, and the learned statistics were then applied to the held-out fold. The lambda parameter (strength of L2 regularization) was tuned by performing nested cross-validation (five folds) and selecting the parameter which maximized accuracy. A threshold for >90% specificity was determined using the outer-fold cross-validated predictions in healthy samples. The model was then trained on all of the training data, using the most frequently selected lambda from the previous step, and predictions were made on the validation samples. The model was trained using the scikit-learn library (version 1.4.2) and the fairlearn library (version 0.13.0). Statistical Analysis The number of proteins detected per sample was compared between breast cancer and healthy groups using Welch’s two-sample t-test (unequal variances). Validation samples were scored using the model trained on the entire training set and threshold for 90% cross-validation specificity was used to classify the samples - positive if the predicted score exceeded the threshold and negative otherwise. The ROC curve was generated using the predicted scores, and the corresponding AUC and 95% bootstrap confidence interval (using 1,000 simulations) were calculated. Sensitivity and specificity for breast cancer and healthy groups, respectively, were estimated along with the corresponding 95% Wilson confidence intervals. Sensitivity estimates for subgroups of pathological stage, molecular subtype, and pathological subtypes within breast cancer samples were calculated along with the corresponding 95% Wilson confidence intervals. Sensitivity and specificity estimate by age group (<40, 40–49, 50–59, 60–69, 70–79, 80+), source of specimens, race, and batch of processing (plate ID) were also calculated along with the corresponding 95% Wilson confidence intervals. For sensitivity by stage, a Cochran-Armitage test was performed to test for significance of trend. Due to limited sample size for certain stages, a permutation test was used: the detection labels were randomly permuted across samples (50,000 permutations) while holding the per-stage sample sizes fixed, the Cochran-Armitage statistic was recomputed for each permutation, and the p-value was calculated as (1 + number of permutations with a statistic ≥ observed)/(1 + total permutations). The negative predicted value (NPV) was estimated using an assumed breast cancer prevalence of 1.25% and the calculation is shown in the Healthcare Utility section of the methods. The analysis was performed in Python (version 3.12.4) using SciPy library (version 1.15.3) and statsmodels library (version 0.14.5). We assessed the robustness of our classifier by looking for associations between the cross-validated labels of the healthy samples using the held-out folds, and potential confounders including source, race, age of the donor at the time of collection, and batch. This confounder analysis was repeated in the validation set, using the predicted labels of healthy samples. We performed a chi-squared contingency table test for all these factors, dichotomizing by classification of the machine learning model. The age of the donor was binarized based on the mean of the healthy samples. Gene Set Enrichment Analysis Preranked gene set enrichment analysis (GSEA, version 4.4.0) was performed comparing all the healthy control samples to the breast cancer samples (including training and validation sets). Limma (version 3.64.1) of non-imputed data was used to perform differential expression of proteins (DEP) between the two groups, where Empirical Bayes moderation of the standard errors was applied with robust variance estimation and a trend prior (trend = TRUE, robust = TRUE). The ranked list was made using sign(logFC)*-log10(adj.P.Val) to rank the genes, and this ranked list was then used as input to preranked GSEA. Healthcare Utility To model the impact of the Astrin Biosciences’ breast cancer early detection test on finding dense breast cancers missed by the current screening guidelines, we simulated a population of 100,000 women with dense breasts (BI-RADS C or D) screened via annual mammogram. Based on the literature, 43% of women have BI-RADS C/D breast tissue,22 and we simulated this population of women to have the following characteristics: 84% have BI-RADS category C (heterogeneously dense) and 16% have BI-RADS category D (extremely dense).23 84% have BI-RADS category C (heterogeneously dense) and 16% have BI-RADS category D (extremely dense).23 1.24% of BI-RADS category C have incident breast cancer and 1.30% of BI-RADS category D have incident breast cancer.23 1.24% of BI-RADS category C have incident breast cancer and 1.30% of BI-RADS category D have incident breast cancer.23 Performance of diagnostics for dense breast cancer was derived from published literature to incorporate into the simulation: Sensitivity and specificity of screening mammogram in BI-RADS category C are 69% (62–76%) and 90% respectively; sensitivity and specificity in BI-RADS category D are 47% (30–65%) and 90%, respectively.23,24 Sensitivity and specificity of screening mammogram in BI-RADS category C are 69% (62–76%) and 90% respectively; sensitivity and specificity in BI-RADS category D are 47% (30–65%) and 90%, respectively.23,24 Sensitivity and specificity of MRI are 97% (86–99%) and 69% (46–85%), respectively.25 Sensitivity and specificity of MRI are 97% (86–99%) and 69% (46–85%), respectively.25 Sensitivity and specificity of CEM are 91% (77–97%) and 74% (52–89%), respectively.25 Sensitivity and specificity of CEM are 91% (77–97%) and 74% (52–89%), respectively.25 Sensitivity and specificity of ABUS are 67% (53–79%) and 89.9% (89.1–90.6%), respectively.26 Sensitivity and specificity of ABUS are 67% (53–79%) and 89.9% (89.1–90.6%), respectively.26 Using the above as well as the sensitivity and specificity of our test as measured in the validation study, we simulated the number of false negative breast cancers and false positives in women with dense breasts. In this population, we assume all women receive a mammogram, and then following a negative mammogram, women may receive one of four different supplemental screens, with adherences ranging from 1% to 100%. This simulation is repeated 100,000 times, and we summarize the number of false negatives and false positives by median, 2.5 percentile, and 97.5 percentile. Uncertainty of sensitivity and specificity for all diagnostics is incorporated into the simulation. Negative and positive predictive values are calculated at each iteration following the mammogram and each supplemental screen. Study Approval Samples were obtained from commercial biobank locations who consent patients under appropriate institution review board approved protocols. All the patient data was provided to Astrin Biosciences in a de-identified manner. This study was conducted in accordance with the Declaration of Helsinki. Results Development of a High Throughput Pipeline for Protein-Based Detection from Plasma For MS-based proteomics to be considered a viable platform for early cancer screening, it was necessary to develop and rigorously validate a high throughput assay capable of analyzing hundreds of patient samples per week. To do this, we utilized an automated robotics platform that allows up to 76 samples to be processed in parallel (Figure 1A and B, Supplemental Figure 1). To validate that sample preparation and analysis meet strict criteria around repeatability (under same conditions) and reproducibility (under different conditions), multiple controls were included. Utilizing commercial digests of yeast enolase and HeLa lysates, we can measure the quantification of specific peptides and show that the retention time stability across all of our sample batches is consistent (Supplemental Figure 2A–F). To confirm that we are minimizing sample carryover from run to run on the MS, we analyzed the TIC of blank injections and patient samples (healthy and cancer). The blank injections were <1% of the total TIC signal from patient samples (Supplemental Figure 2G). Overview of the sample processing methodology (Plasma Preparation), (A) after receiving the banked plasma samples through injection onto the mass spectrometer followed by data processing, normalization and then cancer prediction (Data Analysis), (B). Created using BioRender. The number of peptide (C) and proteins (D) across health and breast cancer samples (both training and validation cohorts) were nearly identical. Spearman correlation (E) and principal component analysis (PCA), (F) were assessed for any discrepancies across the samples which again were nearly identical. While this demonstrates that the Evosep One liquid chromatography (LC) system and the Orbitrap Astral mass spectrometer (MS) are well suited to the challenging conditions required to meet the high standards required for a clinical assay, sample preparation remained as the most likely source for variation. To circumvent this, we prepared and included a known concentration of intact yeast protein extract that was spiked onto each plate and processed in parallel with the patient samples. This allowed us to track sample recovery, reduction and alkylation efficiency, peptide recovery, and protein and peptide identifications (Supplemental Figure 3A). In addition, positive and negative controls made from pools of 9 late-stage breast cancer samples (stage III and IV) and healthy patient samples, respectively, were analyzed. These control pools provided information on the consistency of the preparation process over time including peptide and protein IDs (Supplemental Figure 3B and C). These samples were not included as part of the training data, but we verified that they classified correctly once the model training was completed (Supplemental Figure 3D). This laboratory developed assay was validated in accordance with the Clinical Laboratory Improvement Amendments (CLIA) standards for high-complexity testing, and the assay’s analytical performance has been characterized and documented. Indeed, repeatability and reproducibility analysis showed full agreement between repeated measurements of the same samples. The n = 8 samples measured in triplicate showed an average positive agreement (APA) of 100% (4/4), and an average negative agreement (ANA) of 100% (4/4). The samples measured by two different operators also had APA of 100% (4/4) and ANA of 100% (4/4). Qualitatively, model scores showed strong agreement within donors whether from the same operator or across operators (Supplemental Figure 4). Using this high throughput proteomic pipeline, we trained on plasma samples that passed QC metrics for peptide and protein counts (Figure 1C and D). This results in a total of 845 women, including 379 newly diagnosed, treatment naïve breast cancer patients (pathological stage 0: n = 45, I: n = 130, II: n = 152, III: n = 35, IV: n = 14) and 466 healthy controls. The patient demographics are summarized in Table 1. The plasma samples were de-identified, randomized, and blinded before processing and running on the MS. All MS files were run sequentially and normalized to identify and quantify each individual patient’s proteome. Table 1Patient Demographics and Splits Used in the Training and Validation SetsDemographicsTraining SetValidation SetHealthyCancerHealthyCancerSample sizeCount, n466379195202AgeMean (SD)55.4 (10.8)57.2 (12.4)56.4 (8.8)59.8 (13.0)BMIMean (SD)26.2 (4.4)27.8 (6.3)26.2 (4.3)26.7 (5.4)pStage0, n (%)45 (11.9%)14 (6.9%)I, n (%)130 (34.3%)70 (34.7%)II, n (%)152 (40.1%)98 (48.5%)III, n (%)35 (9.2%)16 (7.9%)IV, n (%)14 (3.7%)2 (1.0%)Unknown, n (%)3 (0.8%)2 (1.0%)PathologyLCIS, n (%)1 (0.3%)1 (0.5%)DCIS, n (%)38 (10.0%)11 (5.4%)ILC, n(%)33 (8.7%)20 (9.9%)IDC, n(%)270 (71.3%)153 (75.7%)Mixed, n(%)4 (1.0%)1 (0.5%)Other/Unknown, n(%)33 (8.7%)16 (8.0%)SubtypeHR+/HER2-190 (50.1%)121 (59.9%)HR-/HER2-35 (9.2%)16 (7.9%)HR+/HER2+27 (7.1%)6 (3.0%)HR-/HER2+19 (5.0%)3 (1.5%)Unknown108 (28.6%)56 (27.7%)RaceWhite, n (%)429 (92.1%)355 (93.6%)176 (90.3%)197 (97.5%)Black/African American, n (%)12 (2.6%)8 (2.1%)4 (2.1%)1 (0.5%)Asian, n (%)21 (4.5%)15 (4.0%)10 (5.1%)4 (2.0%)Other/Unknown, n (%)4 (0.8%)1 (0.3%)5 (2.5%)0 (0.0%)SourceProteoGenex, n (%)316 (67.8%)240 (63.3%)105 (53.8%)129 (63.9%)BioIVT, n (%)150 (32.2%)139 (36.7%)90 (46.2%)73 (36.1%) Patient Demographics and Splits Used in the Training and Validation Sets For all patient samples that were ran, the median (Q1-Q3) number of proteins detected between breast cancer and healthy samples were 7,064 (6,904–7,160) and 7,054 (6,876–7,187) proteins, respectively, with 9,266 unique proteins observed across both cohorts. Most patients had more than 6,500 proteins detected per sample (92.5%, 1149/1242) with no significant differences in means across the two cohorts (Welch’s t-test p-value: 0.81 (training), 0.50 (validation)). This demonstrates both the consistency and quality of our sample and batch preparation coupled to the sensitivity of our MS method (Figure 1C and D). Semi-quantitative assessment of protein intensities showed that we captured a dynamic range of protein abundances spanning over 8 orders of magnitude (Supplemental Figure 5), highlighting the workflow’s potential to capture low abundance protein fractions. Spearman correlation (Figure 1E) and principal component analysis (PCA) of all proteins in the proteome (Figure 1F) resulted in no differences between the healthy and breast cancer samples, suggesting that patient-to-patient variability in the proteomic signatures dominates the underlying cancer signal. However, a supervised approach can still reliably detect the cancer signal from patients’ proteomic profiles as demonstrated in the subsequent figures. Model Training and Held-Out Validation for Breast Cancer Prediction The model training and held-out validation were performed at gene-group-level: out of 11,000 unique gene groups, a median (Q1-Q3) of 8,406 (8,197–8,533) were quantified per sample and 2,594 (2,467–2,803) were imputed per sample. In cross-validation of the training samples, we observed an AUC of 0.961 (95% Bootstrap CI: 0.947–0.97) (Figure 2A) in distinguishing breast cancer from healthy samples. A receiver operating curve (ROC) showed strong discriminative performance, demonstrating the model’s accuracy for breast cancer prediction (Figure 2A). Across all patient samples in the training cohort, the model achieved a sensitivity of 89.2% (95% Wilson CI: 85.7–91.9%) with a specificity of 90.1% (95% Wilson CI: 87.1–92.5%) (Figure 2B). Training and validation performance. Receiver operator curve (ROC) was assessed for the training (A) with performance of sensitivity and specificity for the training (B) samples. ROC was also assessed for the validation set (C) with performance of sensitivity and specificity for the validation samples (D). Sensitivity remained high across all breast cancer stages (E). The error bars denote the 95% Wilson confidence intervals. Abbreviation: AUC, area under the curve. To validate our training model, we analyzed a held-out cohort of 397 samples consisting of 195 healthy and 202 breast cancer patients (pathological stage 0: 14, stage I: 70, stage II: 98, stage III:16, and stage IV: 2; stage unknown: 2) as a held-out validation set. Overall performance was 92.6% sensitivity (95% Wilson CI: 88.1–95.4%) at 92.3% specificity (95% Wilson CI: 87.7–95.3%) with an AUC of 0.975 (95% Bootstrap CI: 0.961–0.987) and a negative predictive value (NPV) of 99.87% (95% Bootstrap CI: 99.79–99.92%) (Figure 2C and D). When broken down by stage, the estimated sensitivities (95% Wilson CI) for stage 0-IV were: 85.7% (60.1–96.0%), 88.6% (79.0–94.1%), 96.9% (91.4–99.0%), 87.5% (64.0–96.5%), and 100.0% (34.2–100%), respectively) (Figure 2E). We observed no differences in performance accuracy across stages 0-III (Cochran-Armitage permutation trend test p-value = 0.204). Stage IV was not considered due to limited sample size. Across molecular and pathological subtypes of breast cancer, the estimated sensitivities (95% Wilson CI) were HR+/HER2-: 95.9% (90.7–98.2%), HR-/HER2-: 93.8% (71.7–98.9%), HR+/HER2+: 66.7% (30.0–90.3%), HR-/HER2+: 100% (43.9–100%), LCIS: 100% (20.7–100%), DCIS: 90.9% (62.3–98.4%), ILC: 100% (83.9–100%), IDC: 94.8% (90.0–97.3%), and mixed histology: 100% (20.7–100%) (Figure 3A and B). In the pathological subtypes analysis, LCIS was a grade 3 and two stage 0 cancers were diagnosed as Paget’s disease and were subtyped as “Other/Unknown”. Potential confounders were also assessed in our training and validation cohorts including age, source, race, and plate (or batch) (Figure 4A and B and Supplemental Figure 6A, B). We observed no association between the predicted label and the tested variables among healthy samples in either the training or validation (Table 2). Sensitivity across breast cancer subtypes. Sensitivity analyses of breast cancer molecular (A) and pathological (B) subtypes show no differences among the different groups. The error bars denote the 95% Wilson confidence intervals. Assessment of confounders based on age and sample source. To ensure that the training model was not influenced by confounders, we assessed validation performance by age (A) and sample source (B). The error bars denote the 95% Wilson confidence intervals. Table 2Confounder AnalysisConfounderTraining SetValidation Setp-valuesp-valuesAge of donor0.4841.000Plate0.6290.282Race0.9400.394Source0.1500.721 Confounder Analysis Pathway Analyses of Breast Cancer Samples To ensure we are identifying and enriching cancer-associated proteins and pathways, we performed gene set enrichment analyses (GSEA) by combining both the training and held-out validation breast cancer samples. Based on 50 Hallmark pathways, we first assessed if the number of proteins detected met the minimum recommended threshold of 15 genes for each of the pathways (Figure 5A). While this threshold was met for almost all pathways, on average ~68% of the genes in each pathway were represented. GSEA identified enriched cancer-associated pathways including epithelial-to-mesenchymal transition (EMT), PI3K-AKt signaling, KRAS up signaling, and WNT/beta-catenin signaling (Figure 5B–D). Importantly, several immune-activation or regulatory pathways (Interferon gamma, alpha, TGF beta) were not significantly enriched in breast cancer samples, or even decreased in these plasma samples (IL6, IL2, TNFalpha signaling) (Supplemental Figure 7A–C). Altogether, the plasma proteins captured were associated with known breast cancer pathways with limited or even negative associations with immune cell activity. Pathway analysis of breast cancer enriched proteins. Gene set enrichment analysis was performed on all pathways with >15 genes (A) and found enrichment of pathways including epithelial-to-mesenchymal transition (EMT) (B and C) and PI3K-AKT signaling (B and D). Healthcare Utility of Astrin Biosciences’ Breast Cancer Early Detection Test While breast density was lacking in our particular cohort to assess model performance, we have published in an earlier manuscript that the performance of our test is independent of breast density.18 Based on that information and our sensitivity and specificity, we modeled the utility of our test to detect breast cancers in women with dense breasts in a screening population. We observed a dramatic increase in the proportion of dense breast cancers detected when incorporating our test as a supplemental screening test following a negative mammogram (Figure 6A). In our analysis, we found that without supplemental screening, 431 (95% CI: 347–521) breast cancers were estimated to be false negatives. Increasing uptake in all pathways resulted in fewer missed breast cancers, but with ABUS, 142 (95% CI: 87–212) breast cancers were still missed at 100% uptake. Conversely, while MRI and CEM missed only 13 (95% CI: 2–49) and 39 (95% CI: 11–105) breast cancers respectively, there were 37,433 (95% CI: 23,013–57,566) and 32,984 (95% CI: 19,988–53,369) false positives at 100% uptake when combined with mammography (Figure 6A and B). Meanwhile, false positives remain low with our deep proteome analysis (16,591; 95% CI: 13,771–20,371) at 100% uptake when combined with mammography with a similarly low number of breast cancers (31; 95% CI: 15–54) estimated to be false negatives. Even with a moderate uptake (40%) of our test, we would reduce the number of false negative/undiagnosed breast cancers in women with dense breasts to 271 (95% CI: 214–332) (Figure 6B). Healthcare utility of deep proteome-based breast cancer supplemental screening test. Women with dense breasts were assumed to receive standard of care with low uptake of supplemental screening or receive our test following a negative mammogram (A) and found to have a much higher rate of detected dense breast cancers with low rates of false positives when screened via Astrin (B). The uncertainty in (B) is derived from a Monte Carlo simulation. Discussion We have developed and built a reproducible and repeatable protein-based liquid biopsy assay that is robust for breast cancer detection, especially at early stages including pathological stages 0, I, and II. There were also no differences in performance across molecular or pathological subtypes. Our machine learning classifier was able to distinguish cancer-specific signals and was not confounded by other variables including age, sample source, and batch. This assay is automated and capable of analyzing several hundred samples in parallel per week. Held-out validation of this test maintained high sensitivity and specificity, paving the way for such a test to be offered as a supplemental screening option for women with a negative mammogram with moderate to high risk of harboring breast cancer. The results of our findings may be especially important for women with an invasive lobular carcinoma (ILC) pathology, which are extremely difficult to detect via mammogram due to their distinct biology and growth pattern as compared to invasive ductal carcinoma (IDC).27 Even with the limited sample size of our validation set (N = 20), our test identified 100% of the ILC cases (see Figure 3B). We plan to enrich this histology in future prospective studies to confirm performance as it will be critical to have an early detection test sensitive enough to detect ILCs. For over four decades, mammograms have been used to screen for breast cancer, enabling early diagnosis in many individuals, thereby increasing breast cancer survival. While mammogram sensitivity in women with non-dense breast tissue is high, women with dense breast tissue are at risk of missed tumors as mammograms have been shown to miss more than half of diagnosed cancers (8.4/1000 for mammogram6 and 17.4/1000 abbreviated MRIs (AB-MRIs) per 1000 supplemental exams).28 This underscores the considerable opportunity to enhance cancer detection in women with dense breast tissue. Currently, women with dense breasts are referred to an additional screening modality, typically ultrasound, which increases the number of cancers detected but still with low sensitivity. Although supplemental imaging techniques such as automated whole breast ultrasound, CEM, and AB-MRI each offer incremental gains, they are constrained by important trade-offs including operator dependence, cost, higher recall burdens, limited accessibility, and reliance on contrast agents. These persistent challenges highlight the need for innovative approaches capable of improving detection which are convenient, affordable, specific, and do not strain clinical workflows. In this context, our proposed breast cancer early detection test offers a promising path forward with a high sensitivity in early-stage breast cancers, including women with dense breasts, while maintaining high specificity and NPV. As a liquid biopsy test, implementation of the plasma derived deep proteome analysis into the clinic would be straightforward and could be widely implemented after mammogram and before MRI as a supplemental screening test. For example, a dense breast patient who needs supplemental screening would receive our test following a negative mammogram with the goal of reducing unnecessary MRIs. In the case of a positive result, the patient would then be referred to an MRI for tumor identification and biopsy. In recent years, many blood-based cancer screening tests have entered the market. ShieldTM by Guardant specifically detects the presence of colorectal cancer while FirstLook by DELFI Diagnostics detects lung cancer. Notably, Grail Galleri® and Exact Sciences CancerGuardTM are both multi-cancer early detection (MCED) tests with varying sensitivities based on the cancer type evaluated and cancer stage. All of these tests detect cancer via cfDNA, such as fragmentomics, methylation, or somatic alteration as the primary analyte, with some supplemented by a panel of blood-based proteins, such as in the CancerGuardTM test. Importantly, it should be emphasized that the sensitivity for early-stage breast cancer detection is low for both Galleri® and CancerGuardTM implying that cfDNA-based assays, even with a small sub-sampling of proteins, may not be sufficient as the primary analyte for early-stage breast cancer detection, likely due to the low shedding rate of cfDNA in early-stage cancers.29 Indeed, the sensitivity for early-stage breast cancer detection using Galleri® is 2.6 and 47.5% for stages I and II, respectively,11 and for CancerGuardTM is 8.3 and 40.0% for stages I and II, respectively. Further, the MCED tests require several milliliters (mL) to several tubes of blood. The described breast cancer early detection test has the unique advantage of finding early-stage breast cancer with <1 mL of blood with sensitivities approaching or exceeding 90% across stages 0-II and will be comparable in cost to MRI or other screening tests. There are a few notable limitations to this study. First, the analysis is conducted on samples purchased from biobanks. While we ruled out confounding with available clinical information, we cannot rule out the possibility that our classifier is confounded by other unmeasured factors, such as selection bias and variations in collection and storage techniques. Second, this analysis has been conducted on women where BI-RADS density was not recorded. Importantly, our analysis of a separate study showed that it is likely that our test performance is independent of breast density.18 Future studies will be required to confirm performance and suitability for supplemental analysis of dense breast cancer patients at the same time. Third, for validation, we randomly allocated our dataset into training and validation. These validation samples were batched separately from the training samples and so represent generalization to future batches. However, future validations will be required to confirm generalizability to future groups of individuals from those observed in training using external datasets. Fourth, the racial demographics of the training and validation sets were predominantly white (92.8% and 94.0%, respectively), potentially limiting the validity of these findings for minority groups. Future efforts will be aimed at prospective collection of samples that increase source variety with more complete clinical annotations and reducing racial disparities in our validation cohorts. MS-based throughput remains a significant challenge for large-scale implementation. Significant capital would be required for the purchase of several mass spectrometers that run in parallel to perform a given test at scale. Alternatively, the development of a targeted protein panel will increase test throughput substantially without significant capital requirements. For our healthcare utility analysis, we emphasize that these conclusions rely on simulated data rather than real-world clinical outcomes directly tied to our test as well as our prior published observation on our test performance on dense breast tissue.18 As stated previously, BI-RADS density was not recorded in this cohort and we extrapolated our findings from our prior published work to assume similar performance on dense breast tissue for this simulation. Finally, our current methodology requires blood to be drawn into EDTA tubes, centrifuged, and frozen within 24 hours. Future work will also evaluate other ways to mitigate protein variations that come from blood stored in EDTA tubes, with the goal of increasing the accessibility and consistency of recruited samples. Conclusions We describe a novel deep proteome-based breast cancer early detection test that can serve as a supplemental screening test particularly in patients with dense breasts (BI-RADS C/D) as seen on screening mammography. The test has a high sensitivity and high specificity with an NPV of 99.87% with the potential to reduce a significant number of unnecessary MRIs and biopsies. Future trials and studies will focus on external validation in larger, more diverse populations.