XpressO: An explanatory deep learning pipeline for the prediction and visualization of gene expression heterogeneity in breast tumors

XpressO: An explanatory deep learning pipeline for the prediction and visualization of gene expression heterogeneity in breast tumors | 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 Article XpressO: An explanatory deep learning pipeline for the prediction and visualization of gene expression heterogeneity in breast tumors Shrey S. Sukhadia, Digvijay Yadav, Vibha Rao, Meher Kalra, Kristen Muller, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6167948/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Spatial genetic heterogeneity plays a critical role in tumor evolution and therapeutic resistance, yet traditional histopathological characterization remains challenging and time-consuming. Here, we present an explainable deep learning pipeline “XpressO” that predicts and visualizes gene expression directly from whole slide images (WSIs), providing spatial resolution of tumor transcriptomics. Using histopathological image features of WSIs of invasive breast cancer as data and associated bulk RNA sequencing data from The Cancer Genome Atlas (TCGA) as expression labels, our model forms complex associations between tissue phenotype and gene expression. By generating high-resolution expression maps, our approach reveals both spatial variation and predicted gene activity across tumor samples, capturing patterns that are often lost in bulk profiling. The interpretability framework further highlights histological regions that contribute to specific gene expression signals, bridging the gap between tumor histology and genetic heterogeneity. This method offers a promising tool for integrating imaging and transcriptomics, enabling data-driven biomarker discovery and advancing precision oncology through spatially-informed molecular profiling. Biological sciences/Cancer Biological sciences/Cancer/Breast cancer Health sciences/Biomarkers/Prognostic markers Biological sciences/Molecular biology/Transcriptomics Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Deep Learning (DL) has revolutionized the analysis of whole slide images (WSIs) in oncology, offering unprecedented capabilities in cancer diagnosis, prognosis, and treatment planning [ 1 ]. By leveraging complex neural networks, DL models can process high-resolution WSIs to identify intricate patterns and features that may be challenging for human observers to discern [ 2 ]. In recent years, substantial progress has been made in applying DL techniques to WSIs, such as the development of sophisticated algorithms capable of segmenting tumor regions, classifying cancer subtypes, and predicting patient outcomes based on histopathological data [ 3 – 5 ]. These rapid advancements have exhibited tremendous clinical potential across cancer types, enhancing diagnostic precision and prognostic evaluations in myriad cancers, including cancers of the skin [ 6 ], lung [ 7 ], colon [ 8 ], cervix [ 9 ], prostate [ 10 ], and breast [ 11 – 12 ]. Despite these advancements, the explainability and interpretability remain key challenges in applying DL to cancer WSIs, where the “black box” nature of DL models renders them difficult to understand and interpret, ultimately creating a barrier to the incorporation of DL models into clinical practice [ 13 ][ 14 ][ 15 ]. Breast cancer is one of the most commonly diagnosed cancer worldwide [ 16 ], and is steadily increasing in incidence, most notably in women under 50 [ 17 ]. Invasive breast carcinoma (IBC) is a type of breast cancer that has spread beyond the ducts or lobules of the breast into the surrounding breast tissue. As IBC has the potential to spread to regional draining lymph nodes and other distant parts of the body [ 18 ], early detection and diagnosis is paramount to survival. IBC is typically detected and diagnosed through a combination of imaging, histopathological examination of needle biopsy samples, and molecular profiling, which helps determine the appropriate course of treatment [ 19 ], where gene expression profiles in both tumor tissues and extratumoral microenvironment have been shown to predict outcomes [ 20 ][ 21 ]. Importantly, given the high degree of genetic heterogeneity, both spatially and temporally within individual tumors and across different patients [ 22 ][ 23 ][ 24 ], there is a pressing need to expand biomarker profiles to gain a deeper understanding of genetic heterogeneity and achieve the goals of precision medicine—providing targeted treatments based on specific biomarkers. This urgency has accelerated the incorporation of DL into breast cancer research, fueling technological advancements [ 25 ][ 26 ]. To this end, convolutional neural networks (CNNs) have recently been utilized to classify histopathological images into categories such as benign or malignant, and to identify specific subtypes of breast cancer such as ductal, lobular, mucinous, and papillary carcinomas [ 27 ]. Moreover, hybrid models combining multiple DL architectures have demonstrated improved performance in classifying breast cancer from histopathological images [ 28 ][ 29 ]. Beyond classification, DL approaches have been applied to predict gene expression profiles from WSIs to aid in the determination of molecular subtypes and to predict early stage breast cancer recurrence using hematoxylin and eosin-stained WSIs. These techniques approximate traditional gene-based methods and provide a faster route to molecular characterization, underscoring the crucial role of combining DL and gene expression in the selection of appropriate postoperative treatments [ 30 , 31 ]. Furthermore, the ongoing issue of explainability has catalyzed development of interpretable models that provide insights into their decision-making processes and is a critical area of ongoing research [ 32 ]. While interpretable DL models applied to WSIs offer transformative potential for predicting gene expression profiles in breast cancer [ 33 ], the selection of genes for model training poses a major challenge and has often relied on predefined gene sets, such as those identified through Gene Set Enrichment Analysis (GSEA) [ 34 ] or established panels like PAM50 profiles [ 35 ][ 36 ]. Although these approaches provide valuable insights into pathway-level perturbations and tumor subtyping, they may constrain the identification of novel, actionable, independent biomarkers that do not align with predefined pathways or signatures. In contrast, by predicting gene expression profiles directly from WSI features and subsequently integrating these results into breast cancer pathways, we can create a more data-driven and holistic framework in which each gene is represented equally, irrespective of prior biological assumptions. Thus, to expand biomarker profiles in breast cancer research using DL and address barriers in explainability in DL models, we created a transparent DL pipeline ‘XpressO’ that i) demonstrates a step-by-step workflow with customized scripts implemented to bridge the gaps between publicly available DL methodologies, and ii) uses RNA expression data from the tissue of each WSI to create a spatial heat map that visualizes the predicted expression of biomarkers based on tumor phenotype. These visualizations highlight the probability of high or low expression of genes in specific regions of interest on WSIs, potentially bypassing traditional and time-consuming histopathological techniques and making predictions interpretable for clinical applications. Materials and Methods Collection and Processing of WSIs The explainable DL pipeline created in the present study can be seen in Figure 1. First, we accessed WSIs of IBC from 191 patients from The Cancer Genome Atlas (TCGA) web-portal [37]. RNA-seq data for each of the 191 WSIs were previously derived by isolating and sequencing RNA from WSI tissues. For the present study. RNA-seq data were downloaded in the form of fragments per kilobase per million reads (FPKM) values from the TCGA portal. The DL-based weakly supervised algorithm Clustering-constrained Attention Multiple Instance Learning (CLAM) was employed to segment tumor regions of interest (ROIs) on the WSIs [38]. CLAM leverages attention mechanisms to automatically detect sub-regions with high diagnostic significance for accurate whole-slide classification. Additionally, it employs instance-level clustering on the identified representative regions to constrain and refine the feature space, enhancing model interpretability, and performance. As scanned WSIs result in large files and high-resolution images, they were divided into smaller and more manageable patches (~1000 patches per WSI) using the ‘opencv’ module in Python [39]. Using a sliding window approach, the ROI mask guided the extraction of fixed-size square patches (e.g., 256 × 256 pixels) at the desired magnification level of 20x. To extract features from these masks, we utilized a pre-trained Unified Network for Instance-level Representation Learning (UNI) model (Figure 1) [40]. The pre-trained UNI model is a weakly supervised Vision Transformer (ViT-L/16) [41] via DINOv2 [42], optimized for histopathological image analysis, enabling robust extraction of high-dimensional features (i.e., feature embeddings from WSI patches). The resulting feature embeddings offered a computationally efficient and accurate representation of histopathological patterns that were fed into the downstream task of classification of gene expression for the WSIs (Figure 1). Collection and Processing of RNA-seq data The gene expression (i.e., RNA-seq) data for the 191 WSIs of IBC tissues were downloaded from The Cancer Genome Atlas (TCGA) in the form of fragments per kilobase per million reads (FPKMs). Genes with FPKM values not reported in n≥1 samples or with FPKM values ≤5 were filtered out. This filtration eliminated the missing and low value FPKMs that could potentially hinder interpretability of the respective genes by the model. Further, a literature review was conducted to identify both common and novel genes of interest that could be expressed at either lower or higher levels in breast tumors than in breast normal tissues [45-93]. FPKM values were then binarized into "high" and "low" categories based on the median expression value of each selected gene (label) using a custom script [43]. Following the binarization of FPKMs, the dataset was split into training, validation, and testing subsets using a predefined script that was run for a range of ‘k’ folds (i.e. k=[4, 14] folds) for each gene of interest. For instance, for a k-fold of 10, the dataset was split into 80:10:10 ratio, defining 80% for training, 10% for validation (for k-fold cross-validation during the training round), and 10% for the testing set. The binary label mappings of high (“1”) and low (“0”) expressions were applied to training, validation, and testing sets depending on the tissue slides (or WSIs) that were populated in each of these sets. This prepared the slide-level labels ready for the task of gene expression classification, marking each WSI as either high-expression or low-expression based on the binarized gene expression levels assigned to them. Model Training The CLAM SB model was used for the classification of the binarized gene expressions using the feature-embeddings extracted from the ROIs of tissue slides [44]. The training phase involved weakly supervised learning, using slide-level labels without specific annotations for individual patches. This enabled the model to automatically identify patches within the ROIs of each slide for which feature-embeddings may or may not be contributing to the high or low expression of the relevant biomarker (or gene). Training parameters were set as follows: Learning rate = 2×10 −4 , Epochs = 200, Optimizer = Adam, Dropout = 25%, and cross validation-fold range = [4, 12]. Model Evaluation and Metrics The trained model was evaluated on independent test sets for the genes of interest based on the data-split based on the best performing ‘k’ fold during the training round (e.g., the test set size would differ for the k =10 [80:10:10=training:validation:testing ratio] vs. k=12 [76:12:12 ratio]). The performance metrics such as area under the receiver operating characteristic curve ( AUC-ROC) , accuracy, precision (i.e., how many of the predicted positive instances are actually positive) , recall (i.e., how many of the actual positive instances the model is able to correctly identify) , and F1-score (i.e., the harmonic mean of precision and recall) were assessed for the testing set for each gene of interest. Visualization the Distribution of High and Low Gene Expression on WSIs Gene expression heatmaps were generated on the testing set (WSIs) using the heatmap module pf CLAM to visualize the attention areas identified on WSIs by the CLAM SB model. Attention scores to depict aggregated probabilities (p0 or p1) of expression were constructed for the top ten highest performing WSI-patches. The scores were subsequently overlaid on the original WSIs and interpreted further, highlighting the regions of high diagnostic significance for each gene of interest. Results Genes of interest Based on our literature search for both known and novel genes of interest that have been shown to be significantly under- or overexpressed in breast tumors compared to normal breast tissues, we identified an initial set of 36 genes (Table 1). Table 1: Genes found to be under- or over-expressed in breast tumors in the literature. Gene Symbol Full Gene Name References MMP11 Matrix metalloproteinase 11 [45][46] FCGR2A Fc gamma receptor IIa [47] HER2 Human epidermal growth factor receptor 2 [48][49] ESR1 Estrogen receptor 1 [50][51] HLA-DRA Major histocompatibility complex, class II, DR alpha [52] C3AR1 Complement component 3a receptor 1 [53] FYB FYN binding protein 1 [54] RRM2 Ribonucleotide reductase M2 [55] PLAU Plasminogen activator, urokinase [56] COL1A1 Collagen alpha 1 type 1 chain [57][58] KRT81 Keratin 81 [59][60] S100A7 S100 calcium binding protein A7 [61][62] SCGB1D2 Secretoglobin family 1D member 2 [63][64] KRT18 Keratin 18 [65] NOTCH3 Notch receptor 3 [66][67] PIK3CA Phosphatidylinositol-4,5-bisphosphate 3-kinase catalytic subunit alpha [68] TP53 Tumor protein 53 [69] KMT2C Lysine-specific methyltransferase 2C [70] MAP3K1 Mitogen-activated protein kinase kinase kinase 1 [71] GATA3 GATA binding protein 3 [72] PTEN Phosphatase and tensin homolog [73] SF3B1 Splicing factor 3b subunit 1 [74] COL11A1 Collagen type XI alpha 1 chain [75][76] COL10A1 Collagen Type X Alpha 1 Chain [77] PCOLCE2 Procollagen C-endopeptidase enhancer 2 [78] LAMA2 Laminin subunit alpha 2 [79][80] TMTC1 Transmembrane O-mannosyltransferase targeting cadherins 1 [81] ADAMTS5 ADAM metallopeptidase with thrombospondin type 1 motif 5 [82] TIMP4 TIMP metallopeptidase inhibitor 4 [83] RSPO3 R-spondin 3 [84] OR2W3 Olfactory receptor family 2 subfamily W member 3 [85][86] OR2B6 Olfactory Receptor Family 2 Subfamily B Member 6 [85] OR2T8 Olfactory Receptor Family 2 Subfamily T Member 8 [85] MMP3 Matrix metalloproteinase-3 [87][88] KLK8 Kallikrein-related peptidase 8 [89][90] PTPRZ1 Protein tyrosine phosphatase receptor type Z1 [91][92][93] After filtering genes with FPKM values that were either not reported in n≥1 samples or reported at ≤5 FPKM, the following nine genes were retained in the FPKM file for further analyses: ESR1 (estrogen receptor 1), S100A7 (S100 calcium binding protein A7), SCGB1D2 (secretoglobin family 1D member 2), COL11A1 (collagen type XI alpha 1), GATA3 (GATA binding protein 3), TIMP4 (tissue inhibitor of metalloproteinases 4), MMP3 (matrix metallopeptidase 3), KLK8 (kallikrein-related peptidase 8), and PTPRZ1 (protein tyrosine phosphatase receptor type Z1). Model Evaluation of Gene Expression The performance metrics for the predicted expression of nine genes from the testing set can be found in Table 2. The model successfully predicted the expression of COL11A1 , KLK8 , and GATA3 , with AUC values >0.8 (95% CI: [0.67, 1.0], [0.65, 1.0], and [0.79, 1.0], respectively), indicating strong predictive performance. COL11A1 and GATA3 had precision, recall, and F1 scores >0.8, while KLK8 exhibited slightly lower values between 0.6 and 0.8. Additionally, ESR1 and S100A7 were predicted at a moderate AUC of 0.75 (95% CI: [0.52, 0.99] and [0.53, 0.97], respectively) with considerably high performance metrics (precision, recall, and F1 ≥ 0.7). Other genes performed with AUC values ranging from 0.69 to 0.73, where TIMP4 exhibited the highest precision at 0.84. Table 2 : Performance metrics for the predicted expression of nine genes in the testing set as a function of the best performing kth fold for each gene. Biomarker Best performing fold AUC CI range Test Error AUC Precision Recall F1 GATA3 11th (0.79, 1.0) 0.17 0.90 0.83 0.83 0.83 COL11A1 11th (0.68, 1.0) 0.16 0.84 0.91 0.84 0.82 KLK8 5th (0.65, 1.0) 0.33 0.82 0.67 0.67 0.67 ESR1 11th (0.52, 0.99) 0.22 0.75 0.77 0.75 0.75 S100A7 10th (0.53, 0.97) 0.28 0.75 0.74 0.72 0.72 TIMP4 7th (0.45, 1.0) 0.28 0.73 0.84 0.69 0.67 MMP3 7th (0.48, 0.95) 0.39 0.72 0.63 0.61 0.61 PTPRZ1 9th (0.49, 0.95) 0.33 0.72 0.68 0.67 0.67 SCGB1D2 9th (0.43, 0.95) 0.39 0.69 0.62 0.61 0.59 AUC: area under the curve; CI: confidence interval Prediction and Visualization of Gene Expression on WSIs The attention scores from the heatmap module of CLAM depicted aggregated probabilities (p0 or p1) of ROIs across the top 10 high-performing patches for the classification of high versus low gene expression for WSIs belonging to the testing set of the respective gene or biomarker (Table S1). The resulting attention maps (i.e., heatmaps) aided the evaluation of heterogeneity of biomarker expression within and across several WSIs (Figures2-4). The heatmaps showing predicted biomarker expression across WSIs can be seen in Figure 2, while WSI names and aggregated probabilities for both high and low expression are reported in Table S1. Our analyses show that COL11A1 was predicted to be expressed consistently high in the test tissues. KLK8 was predicted to be expressed at low levels across the test tissues, except for WSI ‘TCGA-BH-A0BM’, where it exhibited moderately high predicted expression (68.5%). The predicted expression of MMP3 varied across WSIs, whereas the expression of TIMP4 was predicted to be consistently high across the test tissues. More interestingly, the two tissue pairs ‘TCGA-E2-A15H’ and ‘TCGA-BH-A0BP’ exhibited opposite predicted expression patterns for TIMP4 and MMP3 (Figure 4 a and b). Additionally, GATA3 exhibited varying levels of predicted expression across the test tissues, while ESR1 exhibited mainly low predicted expression, except for high predicted expression in WSI ‘TCGA-E2-A1L8’ (76%) probability. Notably, this WSI was found to be positive for estrogen receptor (ER) expression per IHC results in TCGA. Most interestingly, the expression of both GATA3 and ESR1 were predicted to be low in the two tissues ‘TCGA-BH-A0RX’and ‘TCGA-E2-A150’ and exhibited similar patterns of low expression (Figure 3 a and b). In agreement with these findings, both of these WSIs were found to be negative for ER expression according to the IHC results available in TCGA. Furthermore, as shown in Figure 2 and Table S1, the predicted expression of PTPRZ1 was mainly low across the test tissues except for one tissue (“TCGA-AN-A0FN”), in which it was predicted to have high levels of expression. S100A7 was predicted to be expressed at high levels in two out of four test tissues, whereas the remaining two exhibited a mixed level of predicted expression. Lastly, SCGB1D2 oscillated between high and low predicted expression across several tissues. Discussion This study presents a novel deep learning (DL) pipeline that integrates multiple computational techniques to predict gene expression directly from WSIs of IBC tumors, offering a high-throughput alternative to traditional bulk transcriptomics. By leveraging weakly supervised learning and attention-based interpretability methods, our approach enables the spatial mapping of genetic heterogeneity within and across tumor tissues based on phenotypic characteristics. Notably, the model achieved high predictive performance for key biomarkers such as COL11A1 , KLK8 , and GATA3 , which are associated with tumor aggressiveness and poor prognosis. The high expression of COL11A1 has been significantly correlated with aggressive tumors with poor prognosis [ 94 ][ 95 ]. In agreement with the literature, we observed both high and low predicted gene expression of KLK8 , which has been shown to be differentially expressed in breast cancer cells [ 96 – 98 ]. Additionally, MMP3 exhibited varied expression, whereas TIMP4 exhibited consistently high predicted expression in WSIs in the present study. As the TIMP4 protein is known to inhibit MMP3 and prevent tumor growth [ 99 ][ 100 ], the high expression of the TIMP4 gene could indicate an active role in suppressing tumor growth and metastasis. GATA3 exhibited varied expression in several tissues, whereas ESR1 exhibited low expression consistently across the tissues. Notably, the low predicted expression of GATA3 and ESR1 coincided between two WSIs (i.e., TCGA-BH-A0RX and TCGA-E2-A150). As GATA3 is an ESR1 co-operating transcriptional factor, their expression has been shown to be correlated in breast cancer tissues [ 101 ][ 102 ]. GATA3 and ESR1 were consistently co-predicted with low expression across multiple tissue samples, which aligns with their transcriptional co-regulation in breast cancer. Given that GATA3 is a known cooperating transcription factor for ESR1, their simultaneous low expression patterns could indicate a transition toward an ESR1-independent phenotype, which is often linked to endocrine therapy resistance. This could have significant implications for hormone-based treatments, such as tamoxifen or fulvestrant, which rely on ESR1 activity for efficacy. In addition, PTPRZ1 —which was predicted to be expressed at low levels across most of the tissues in the present study—is a tumor suppressor, and thus low expression could indicate higher cancer cell proliferation and invasion and has exhibited lower expression in a breast cancer cohort than a normal breast cohort [ 103 ]. High expression of S100A7 —which exhibited a high probability of expression in two WSIs and mixed expression levels in others—has been previously correlated with tumor growth and may lead to metastasis in tissues with high levels of expression [ 104 ]. Lastly, in agreement with the literature, the predicted expression of SCGB1D2 was variable in the present study, as SCGB1D2 has been shown to be highly expressed in ductal carcinoma in situ (DCIS) breast cancer with invasive capability compared to the DCIS that is not invasive [ 105 ]. Beyond prediction accuracy and predictions of localized gene expression, our DL pipeline enhances model explainability, bridging the gap between existing computational pathology tools and their clinical applications. By integrating a pre-trained UNI model for feature extraction, CLAM for region-based classification and interpretation of gene-expression categorized using custom scripts, this framework provides clinicians with intuitive visualizations of predicted gene expression patterns on WSIs, potentially aiding both diagnostic decision-making and targeted therapeutic development. This method reduces reliance on labor- and time-intensive immunohistochemistry and cytochemical methodologies, which require separate staining and imaging for each biomarker, making multiplexed analysis difficult. By leveraging DL-based in silico predictions, our approach enables the simultaneous assessment of multiple biomarkers on a single WSI, substantially improving efficiency in histopathology workflows and enhancing high-throughput molecular characterization in a scalable and cost-effective manner. It is important to note that the present study has several limitations. IHC results for the eight out of nine biomarkers identified in the present study are not available from TCGA, and thus we were unable to validate the spatial gene expression predictions for those biomarkers on the tumor patches of the respective WSIs. Additionally, TCGA also does not have spatial transcriptomics data to validate the gene expression for spatial patches. Future work will close this gap by conducting extensive IHC or spatial transcriptomics on in-house breast tumor samples to clinically validate XpressO in the near future. Conclusions The ability to predict region-specific gene expression patterns on WSIs suggests that the ‘XpressO’ pipeline created in the present study could be a valuable tool for biomarker discovery and precision oncology. Our findings reveal novel predicted co-expression patterns of key biomarkers using tissue phenotypes, suggesting that tissue phenotypes can serve as surrogate biomarkers for potential mechanistic insights into IBC progression and treatment resistance regulated by the relevant genes. Future work will focus on validating these findings through spatial transcriptomics or extensive IHC staining, allowing for a region-of-interest (ROI)-specific assessment of gene expression in IBC tissues. Additionally, the freely available online resource developed here will facilitate broader access to the ‘XpressO’ pipeline, ensuring seamless integration into pathology labs and pharmaceutical research for biomarker-driven precision medicine. Declarations Competing Interests JH was paid partially as an editor/writer of this manuscript. Author Contribution SS conceptualized the idea, participated in the construction of data analysis pipeline, interpreted results and contributed majorly to the writing of the manuscript. DY participated in the construction and execution of the data analysis pipeline and contributed to the writing of the manuscript. VR participated in the execution of the data analysis pipeline and contributed majorly to the writing of the manuscript. MK participated in the execution of the data analysis pipeline and reviewed the manuscript. KM reviewed and edited the manuscript. JH contributed majorly to the writing of the manuscript. Acknowledgement This study was funded partially by William H. Neukom 1964 Institute for Computational Science. The funder played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript. Data Availability All data generated or analysed during this study are either included in this published article [and its supplementary information files] or at the repository: https://github.com/skr1/XpressO References Tran, Khoa A., et al. "Deep learning in cancer diagnosis, prognosis and treatment selection." Genome Medicine 13 (2021): 1-17. Li, Mengfang, et al. "Medical image analysis using deep learning algorithms." Frontiers in Public Health 11 (2023): 1273253. Mostafa, Ghada, et al. "The power of deep learning in simplifying feature selection for hepatocellular carcinoma: a review." BMC Medical Informatics and Decision Making 24.1 (2024): 287. Wang, Ching-Wei, et al. "Deep learning for endometrial cancer subtyping and predicting tumor mutational burden from histopathological slides." npj Precision Oncology 8.1 (2024): 287. Wang, Jun, et al. "Advances in Multiple Instance Learning for Whole Slide Image Analysis: Techniques, Challenges, and Future Directions." arXiv preprint arXiv:2408.09476 (2024). "High-fidelity detection, subtyping, and localization of five skin neoplasms using supervised and semi-supervised learning." Journal of Pathology Informatics 14 (2023): 100159. Wang, Xi, et al. "Weakly supervised deep learning for whole slide lung cancer image analysis." IEEE transactions on cybernetics 50.9 (2019): 3950-3962 Hamida, A. Ben, et al. "Deep learning for colon cancer histopathological images analysis." Computers in Biology and Medicine 136 (2021): 104730. Cheng, Shenghua, et al. "Robust whole slide image analysis for cervical cancer screening using deep learning." Nature communications 12.1 (2021): 5639. Singhal, Nitin, et al. "A deep learning system for prostate cancer diagnosis and grading in whole slide images of core needle biopsies." Scientific reports 12.1 (2022): 3383. Jiang, Bitao, et al. "Deep learning applications in breast cancer histopathological imaging: diagnosis, treatment, and prognosis." Breast Cancer Research 26.1 (2024): 137. Zeiser, Felipe André, et al. "DeepBatch: A hybrid deep learning model for interpretable diagnosis of breast cancer in whole-slide images." Expert Systems with Applications 185 (2021): 115586. Teng, Qiaoying, et al. "A survey on the interpretability of deep learning in medical diagnosis." Multimedia Systems 28.6 (2022): 2335-2355. Dhar, Tribikram, et al. "Challenges of deep learning in medical image analysis—improving explainability and trust." IEEE Transactions on Technology and Society 4.1 (2023): 68-75. Salahuddin, Zohaib, et al. "Transparency of deep neural networks for medical image analysis: A review of interpretability methods." Computers in biology and medicine 140 (2022): 105111. Arnold, Melina, et al. "Current and future burden of breast cancer: Global statistics for 2020 and 2040." The Breast 66 (2022): 15-23. American Cancer Society. Cancer Facts and Figures 2025. Atlanta: American Cancer Society; 2025. Srikantamurthy, Mahati Munikoti, et al. "Classification of benign and malignant subtypes of breast cancer histopathology imaging using hybrid CNN-LSTM based transfer learning." BMC Medical Imaging 23.1 (2023): 19. Qasrawi, Radwan, et al. "Hybrid ensemble deep learning model for advancing breast cancer detection and classification in clinical applications." Heliyon 10.19 (2024). Román-Pérez, Erick, et al. "Gene expression in extratumoral microenvironment predicts clinical outcome in breast cancer patients." Breast Cancer Research 14 (2012): 1-12. Huang, Erich et al. “Gene expression predictors of breast cancer outcomes.” Lancet (London, England) vol. 361,9369 (2003): 1590-6. doi:10.1016/S0140-6736(03)13308-9. Turashvili, Gulisa, and Edi Brogi. "Tumor heterogeneity in breast cancer." Frontiers in medicine 4 (2017): 227. Lüönd, Fabiana, Stefanie Tiede, and Gerhard Christofori. "Breast cancer as an example of tumour heterogeneity and tumour cell plasticity during malignant progression." British journal of cancer 125.2 (2021): 164-175. Kudelova, Eva, et al. "Genetic heterogeneity, tumor microenvironment and immunotherapy in triple-negative breast cancer." International Journal of Molecular Sciences 23.23 (2022): 14937. Chugh, Gunjan, Shailender Kumar, and Nanhay Singh. "Survey on machine learning and deep learning applications in breast cancer diagnosis." Cognitive Computation 13.6 (2021): 1451-1470. Dar, Rayees Ahmad, Muzafar Rasool, and Assif Assad. "Breast cancer detection using deep learning: Datasets, methods, and challenges ahead." Computers in biology and medicine 149 (2022): 106073. Zerouaoui, Hasnae, and Ali Idri. "Deep hybrid architectures for binary classification of medical breast cancer images." Biomedical Signal Processing and Control 71 (2022): 103226. Lee, Sung Hak, and Hyun-Jong Jang. "Deep learning-based prediction of molecular cancer biomarkers from tissue slides: A new tool for precision oncology." Clinical and Molecular Hepatology 28.4 (2022): 754. Lee, Geongyu, et al. "Predicting the risk of early-stage breast cancer recurrence using H\&E-stained tissue images." arXiv preprint arXiv:2406.06650 (2024). Muhammad, Dost, and Malika Bendechache. "Unveiling the black box: A systematic review of Explainable Artificial Intelligence in medical image analysis." Computational and structural biotechnology journal (2024). Levy-Jurgenson, Alona, et al. "Spatial transcriptomics inferred from pathology whole-slide images links tumor heterogeneity to survival in breast and lung cancer." Scientific reports 10.1 (2020): 18802. Schmauch, Benoît, et al. "A deep learning model to predict RNA-Seq expression of tumours from whole slide images." Nature communications 11.1 (2020): 3877. Wang, Yinxi, et al. "Predicting molecular phenotypes from histopathology images: a transcriptome-wide expression–morphology analysis in breast cancer." Cancer research 81.19 (2021): 5115-5126. Parker, Joel S., et al. "Supervised risk predictor of breast cancer based on intrinsic subtypes." Journal of clinical oncology 27.8 (2009): 1160-1167. SJ, SCHNITT. "Pathology of invasive breast cancer." Diseases of the Breast (2000). American College of Radiology, et al. "Standards for diagnosis and management of invasive breast carcinoma." CA: A cancer journal for clinicians 48.2 (1998): 83-107. https://www.cancer.gov/tcga Lu, Ming Y., et al. "Data-efficient and weakly supervised computational pathology on whole-slide images." Nature biomedical engineering 5.6 (2021): 555-570. https://www.python.org/ Chu, Jiaming, et al. "UniParser: Multi-Human Parsing with Unified Correlation Representation Learning." IEEE Transactions on Image Processing (2024). Chen, R.J., Ding, T., Lu, M.Y. et al. Towards a general-purpose foundation model for computational pathology. Nat Med 30, 850–862 (2024). https://doi.org/10.1038/s41591-024-02857-3. Oquab, Maxime, et al. "Dinov2: Learning robust visual features without supervision." arXiv preprint arXiv:2304.07193 (2023). Sukhadia, Shrey S., et al. "Machine Learning Enabled Prediction of Biologically Relevant Gene Expression Using CT‐Based Radiomic Features in Non‐Small Cell Lung Cancer." Cancer Medicine 13.24 (2024): e70509. Fillioux, Leo, et al. "Structured state space models for multiple instance learning in digital pathology." International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2023. González de Vega, Raquel et al. “MMP-11 as a biomarker for metastatic breast cancer by immunohistochemical-assisted imaging mass spectrometry.” Analytical and bioanalytical chemistry vol. 411,3 (2019): 639-646. doi:10.1007/s00216-018-1365-3 Zhuang, Ying et al. “MMP11 promotes the proliferation and progression of breast cancer through stabilizing Smad2 protein.” Oncology reports vol. 45,4 (2021):16. doi:10.3892/or.2021.7967 Gavin, Patrick G et al. “Association of Polymorphisms in FCGR2A and FCGR3A With Degree of Trastuzumab Benefit in the Adjuvant Treatment of ERBB2/HER2- Positive Breast Cancer: Analysis of the NSABP B-31 Trial.” JAMA oncology vol. 3,3 (2017): 335-341. doi:10.1001/jamaoncol.2016.4884 Klocker, Eva Valentina, and Christoph Suppan. “Biomarkers in Her2- Positive Disease.” Breast care (Basel, Switzerland) vol. 15,6 (2020): 586-593. doi:10.1159/000512283 Carney, Walter P. “HER2 status is an important biomarker in guiding personalized HER2 therapy.” Personalized medicine vol. 2,4 (2005): 317-324. doi:10.2217/17410541.2.4.317 Brett, Jamie O., et al. "ESR1 mutation as an emerging clinical biomarker in metastatic hormone receptor-positive breast cancer." Breast Cancer Research 23 (2021): 1-15. Carausu, Marcela, et al. "ESR1 mutations: a new biomarker in breast cancer." Expert review of molecular diagnostics 19.7 (2019): 599-611. Saraiva, Diana P et al. “Expression of HLA-DR in Cytotoxic T Lymphocytes: A Validated Predictive Biomarker and a Potential Therapeutic Strategy in Breast Cancer.” Cancers vol. 13,15 3841. 30 Jul. 2021, doi:10.3390/cancers13153841 Song, Guiqin, et al. "Identification of aberrant gene expression during breast ductal carcinoma in situ progression to invasive ductal carcinoma." Journal of International Medical Research 48.1 (2020): 0300060518815364. Chou, Ching-Wen et al. “Identified the novel resistant biomarkers for taxane-based therapy for triple-negative breast cancer.” International journal of medical sciences vol. 18,12 2521-2531. 26 Apr. 2021, doi:10.7150/ijms.59177 Abdel-Rahman, Manar Ahmed, Mena Mahfouz, and Hany Onsy Habashy. "RRM2 expression in different molecular subtypes of breast cancer and its prognostic significance." Diagnostic Pathology 17 (2022): 1-8. Sarno, Federica et al. “Functional Validation of the Putative Oncogenic Activity of PLAU.” Biomedicines vol. 11,1 102. 30 Dec. 2022, doi:10.3390/biomedicines11010102 Liu, Jing et al. “Collagen 1A1 (COL1A1) promotes metastasis of breast cancer and is a potential therapeutic target.” Discovery medicine vol. 25,139 (2018): 211-223. Ma, Hon-Ping et al. “Collagen 1A1 (COL1A1) Is a Reliable Biomarker and Putative Therapeutic Target for Hepatocellular Carcinogenesis and Metastasis.” Cancers vol. 11,6 786. 7 Jun. 2019, doi:10.3390/cancers11060786 Yan, Zhideng et al. “The prognostic marker KRT81 is involved in suppressing CD8 + T cells and predicts immunotherapy response for triple-negative breast cancer.” Cancer biology & therapy vol. 25,1 (2024): 2355705. doi:10.1080/15384047.2024.2355705 Nanashima, Naoki, et al. "Hair keratin KRT81 is expressed in normal and breast cancer cells and contributes to their invasiveness." Oncology reports 37.5 (2017): 2964-2970. Emberley, Ethan D et al. “S100A7 and the progression of breast cancer.” Breast cancer research : BCR vol. 6,4 (2004): 153-9. doi:10.1186/bcr816 Al-Haddad, Sahar, et al. "Psoriasin (S100A7) expression and invasive breast cancer." The American journal of pathology 155.6 (1999): 2057-2066. Zafrakas, Menelaos et al. “Expression analysis of mammaglobin A (SCGB2A2) and lipophilin B (SCGB1D2) in more than 300 human tumors and matching normal tissues reveals their co-expression in gynecologic malignancies.” BMC cancer vol. 6 88. 9 Apr. 2006, doi:10.1186/1471-2407-6-88 Sharma, P., Bhattacharyya, D.K. & Kalita, J. Disease biomarker identification from gene network modules for metastasized breast cancer. Sci Rep 7, 1072 (2017). https://doi.org/10.1038/s41598-017-00996-x Takan, Işıl et al. “"In the light of evolution:" keratins as exceptional tumor biomarkers.” PeerJ vol. 11 e15099. 17 Mar. 2023, doi:10.7717/peerj.15099 Chen, Wen-Jia, et al. "NOTCH3 inhibits transcription factor ZEB1 expression and metastasis of breast cancer cells via transcriptionally upregulating miR-223." Journal of Cancer 15.1 (2024): 192. Chen, Min-Na et al. “Notch3 restricts metastasis of breast cancers through regulation of the JAK/STAT5A signaling pathway.” BMC cancer vol. 23,1 1257. 20 Dec. 2023, doi:10.1186/s12885-023-11746-w Liu, Yi-Rong et al. “PIK3CA mutations define favorable prognostic biomarkers in operable breast cancer: a systematic review and meta-analysis.” OncoTargets and therapy vol. 7 543-52. 10 Apr. 2014, doi:10.2147/OTT.S60115 Silwal-Pandit, Laxmi, et al. "TP53 mutation spectrum in breast cancer is subtype specific and has distinct prognostic relevance." Clinical Cancer Research 20.13 (2014): 3569-3580. Liu, Xinhua et al. “KMT2C is a potential biomarker of prognosis and chemotherapy sensitivity in breast cancer.” Breast cancer research and treatment vol. 189,2 (2021): 347- 361. doi:10.1007/s10549-021-06325-1 Narang, Binish. "MAP3K1 Identified as a Prognostic Biomarker in Breast Cancer After Multi-omics Bioinformatics Analysis." (2023). Takaku, Motoki et al. “GATA3 in Breast Cancer: Tumor Suppressor or Oncogene?.” Gene expression vol. 16,4 (2015): 163-8. doi:10.3727/105221615X14399878166113 Derkyi-Kwarteng, Leonard, et al. "A retrospective analysis suggests PTEN expression is associated with favorable clinicopathological features of breast cancer." Scientific Reports 14.1 (2024): 21645. Maguire, Sarah L., et al. "SF3B1 mutations constitute a novel therapeutic target in breast cancer." The Journal of pathology 235.4 (2015): 571-580. Luo, Qi et al. “COL11A1 serves as a biomarker for poor prognosis and correlates with immune infiltration in breast cancer.” Frontiers in genetics vol. 13 935860. 9 Sep. 2022, doi:10.3389/fgene.2022.935860 Shi, Wenjie, et al. "COL11A1 as an novel biomarker for breast cancer with machine learning and immunohistochemistry validation." Frontiers in Immunology 13 (2022): 937125. Zhang, Mingdi, et al. "Bioinformatics analysis of prognostic significance of COL10A1 in breast cancer." Bioscience reports 40.2 (2020): BSR20193286. Thomas, Michael J., et al. "PCPE2: Expression of Multifunctional Extracellular Glycoprotein Associates With Diverse Cellular Functions." Journal of Lipid Research (2024): 100664. Fu, Tao et al. “LAMC2 as a prognostic biomarker in human cancer: a systematic review and meta-analysis.” BMJ open vol. 12,11 e063682. 17 Nov. 2022, doi:10.1136/bmjopen-2022-063682 Hu, J., and S. Li. "The role of Lama2 in cancer: Current perspectives." Cancer Research Journal 10.4 (2022): 85-88. Ali, Rafat et al. “Identification of New Key Genes and Their Association with Breast Cancer Occurrence and Poor Survival Using In Silico and In Vitro Methods.” Biomedicines vol. 11,5 1271. 25 Apr. 2023, doi:10.3390/biomedicines11051271 Tavousi, Najmeh, et al. "ADAMTS5 modulates breast cancer development as a diagnostic biomarker and potential tumor suppressor, regulating by BAIAP2-AS1, VTI1B, CRNDE, and hsa-miR-135b-3p: integrated systems biology and experimental approach." (2022). Liss, Michaelann, et al. "Tissue inhibitor of metalloproteinase-4 is elevated in early-stage breast cancers with accelerated progression and poor clinical course." The American journal of pathology 175.3 (2009): 940-946. Ter Steege, Eline J., et al. "R-spondin-3 promotes proliferation and invasion of breast cancer cells independently of Wnt signaling." Cancer Letters 568 (2023): 216301. Weber, Lea, et al. "Olfactory receptors as biomarkers in human breast carcinoma tissues." Frontiers in oncology 8 (2018): 33. Masjedi, Shirin, Laurence J. Zwiebel, and Todd D. Giorgio. "Olfactory receptor gene abundance in invasive breast carcinoma." Scientific reports 9.1 (2019): 13736.. Argote Camacho, Angela Ximena, et al. "Metalloproteinases 1 and 3 as potential biomarkers in breast cancer development." International journal of molecular sciences 22.16 (2021): 9012. Ławicki, Paweł, et al. "Plasma Levels of Metalloproteinase 3 (MMP-3) and Metalloproteinase 7 (MMP-7) as New Candidates for Tumor Biomarkers in Diagnostic of Breast Cancer Patients." Journal of Clinical Medicine 12.7 (2023): 2618. Tailor, Prashant D., et al. "Diagnostic and prognostic biomarker potential of kallikrein family genes in different cancer types." Oncotarget 9.25 (2018): 17876. Watrowski, Rafał, et al. "Gene expression of kallikreins in breast cancer cell lines." Anticancer Research 40.5 (2020): 2487-2495. Xia, Zhenkun, et al. "The expression, functions, interactions and prognostic values of PTPRZ1: a review and bioinformatic analysis." Journal of Cancer 10.7 (2019): 1663. Thalor, Anamika, et al. "Machine learning assisted analysis of breast cancer gene expression profiles reveals novel potential prognostic biomarkers for triple-negative breast cancer." Computational and structural biotechnology journal 20 (2022): 1618-1631. Chen, Min, et al. "Integrin α6β4 Upregulates PTPRZ1 Through UCHL1-Mediated Hif-1α Nuclear Accumulation to Promote Triple-Negative Breast Cancer Cell Invasive Properties." Cancers 16.21 (2024): 3683. Li, Xia, Yue Jin, and Jian Xue. "Unveiling Collagen’s Role in Breast Cancer: Insights into Expression Patterns, Functions and Clinical Implications." International Journal of General Medicine (2024): 1773-1787. Luo, Qi, et al. "COL11A1 serves as a biomarker for poor prognosis and correlates with immune infiltration in breast cancer." Frontiers in Genetics 13 (2022): 935860. Yousef, G. M., et al. "Kallikrein gene downregulation in breast cancer." British journal of cancer 90.1 (2004): 167-172. Watrowski, Rafał, et al. "Gene expression of kallikreins in breast cancer cell lines." Anticancer Research 40.5 (2020): 2487-2495. Michaelidou, Kleita, Alexandros Ardavanis, and Andreas Scorilas. "Clinical relevance of the deregulated kallikrein-related peptidase 8 mRNA expression in breast cancer: a novel independent indicator of disease-free survival." Breast cancer research and treatment 152 (2015): 323-336. Suhaimi, Shafinah Ahmad, Soon Choy Chan, and Rozita Rosli. "Matrix metallopeptidase 3 polymorphisms: emerging genetic markers in human breast cancer metastasis." Journal of breast cancer 23.1 (2020): 1-9. Melendez-Zajgla, Jorge, et al. "Tissue inhibitor of metalloproteinases-4. The road less traveled." Molecular cancer 7 (2008): 1-11. Ricks-Santi, Luisel J., et al. "Characterization of GATA3 and Mammaglobin in breast tumors from African American Women." Archives of microbiology & immunology 7.1 (2023): 18. Theodorou, Vasiliki, et al. "GATA3 acts upstream of FOXA1 in mediating ESR1 binding by shaping enhancer accessibility." Genome research 23.1 (2013): 12-22. Huang, Peng, et al. "Chemotherapy-driven increases in the CDKN1A/PTN/PTPRZ1 axis promote chemoresistance by activating the NF-κB pathway in breast cancer cells." Cell Communication and Signaling 16 (2018): 1-12. Mishra, Sanjay, et al. "cPLA2 blockade attenuates S100A7-mediated breast tumorigenicity by inhibiting the immunosuppressive tumor microenvironment." Journal of Experimental & Clinical Cancer Research 41.1 (2022): 54. Doebar, Shusma C., et al. "Gene expression differences between ductal carcinoma in situ with and without progression to invasive breast cancer." The American journal of pathology 187.7 (2017): 1648-1655. Additional Declarations Competing interest reported. JH was paid partially as an editor/writer of this manuscript. Supplementary Files SupplementaryTable1S1.xlsx SupplementaryTable2S2.xlsx Cite Share Download PDF Status: Posted Version 1 posted 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-6167948","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":466009477,"identity":"93d10392-992d-419b-9b34-15cd287f3029","order_by":0,"name":"Shrey S. Sukhadia","email":"data:image/png;base64,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","orcid":"","institution":"Dartmouth Health","correspondingAuthor":true,"prefix":"","firstName":"Shrey","middleName":"S.","lastName":"Sukhadia","suffix":""},{"id":466009478,"identity":"f1907a29-560b-46a9-ac07-6f333ca8a15e","order_by":1,"name":"Digvijay Yadav","email":"","orcid":"","institution":"Dartmouth Health","correspondingAuthor":false,"prefix":"","firstName":"Digvijay","middleName":"","lastName":"Yadav","suffix":""},{"id":466009479,"identity":"50341c1c-93da-4ad1-ab7e-d634afb981f8","order_by":2,"name":"Vibha Rao","email":"","orcid":"","institution":"Dartmouth Health","correspondingAuthor":false,"prefix":"","firstName":"Vibha","middleName":"","lastName":"Rao","suffix":""},{"id":466009482,"identity":"d5efef33-94d4-494f-b48b-83df69c47571","order_by":3,"name":"Meher Kalra","email":"","orcid":"","institution":"Dartmouth College","correspondingAuthor":false,"prefix":"","firstName":"Meher","middleName":"","lastName":"Kalra","suffix":""},{"id":466009484,"identity":"f2a8adff-d093-492e-b615-6dbc8041552c","order_by":4,"name":"Kristen Muller","email":"","orcid":"","institution":"Dartmouth Health","correspondingAuthor":false,"prefix":"","firstName":"Kristen","middleName":"","lastName":"Muller","suffix":""},{"id":466009485,"identity":"fcc1bbe7-f734-4250-8366-badd95c42f63","order_by":5,"name":"Jennifer Head","email":"","orcid":"","institution":"Synergy Scientifics","correspondingAuthor":false,"prefix":"","firstName":"Jennifer","middleName":"","lastName":"Head","suffix":""}],"badges":[],"createdAt":"2025-03-06 07:23:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6167948/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6167948/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":83982829,"identity":"fb6bd6fa-d5e9-46c2-8c7a-b7372dfd0d55","added_by":"auto","created_at":"2025-06-05 10:23:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":83832,"visible":true,"origin":"","legend":"\u003cp\u003eFlow diagram of the XpressO pipeline starting with segmentation of tumor regions of interest (ROIs) on WSIs, followed by the extraction of feature-embeddings from them, accompanied by the processing of bulk gene-expression data from TCGA to categorize them into high and low expression labels, and then followed by prediction of those labels using ROI-feature-embeddings accompanied by visualization of those predictions in top performing patches/areas of those ROIs.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6167948/v1/fc59c42dee27145df8d6724c.png"},{"id":83982712,"identity":"37b63f6a-9fde-43ee-80fb-6e7a03d7deef","added_by":"auto","created_at":"2025-06-05 10:23:45","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":613959,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of predicted gene expression on WSI patches. The red arrow indicates either high or low predicted expression with the respective probability or attention score (p_0) for each biomarker. The colored borders of the boxes enclosing tissue images denote a color-match for the same tissue sample covered by multiple biomarkers. These colors correspond to the color of pairs in Supplementary Table2.xlsx.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6167948/v1/2912aab422e12b70cfaafbf2.png"},{"id":83982814,"identity":"df6d082b-2182-4878-8def-738074e4446a","added_by":"auto","created_at":"2025-06-05 10:23:48","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":546787,"visible":true,"origin":"","legend":"\u003cp\u003eOpposing expression patterns observed for TIMP4 and MMP3 gene, i.e., a) showing the high and low expression of TIMP4 and MMP3 in b.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6167948/v1/7487c14b98fafefedfa794bd.png"},{"id":83982746,"identity":"a3bc5f71-5d0f-46fb-911b-fd07ada97945","added_by":"auto","created_at":"2025-06-05 10:23:47","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":624451,"visible":true,"origin":"","legend":"\u003cp\u003eConsistent low expression observed for GATA3 and ESR1 across the two tissues : a) and b).\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6167948/v1/c5de57328f6ab6de55828b05.png"},{"id":100372567,"identity":"dee66ec4-d771-4727-92f0-dbf536ec6684","added_by":"auto","created_at":"2026-01-16 08:12:41","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2381891,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6167948/v1/5709bd02-c831-4f09-b938-1bea1e0d6a8f.pdf"},{"id":83982863,"identity":"fe552edd-65d0-4d11-bf8f-86fb6cb88933","added_by":"auto","created_at":"2025-06-05 10:23:50","extension":"xlsx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":13429,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTable1S1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6167948/v1/562d4d453a1c2cd5e96b46db.xlsx"},{"id":83982818,"identity":"0087b521-fd81-4789-83c5-32dfc39ed62f","added_by":"auto","created_at":"2025-06-05 10:23:48","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":14682,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTable2S2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6167948/v1/6adca164cc00508cbc868609.xlsx"}],"financialInterests":"Competing interest reported. JH was paid partially as an editor/writer of this manuscript.","formattedTitle":"XpressO: An explanatory deep learning pipeline for the prediction and visualization of gene expression heterogeneity in breast tumors","fulltext":[{"header":"Introduction","content":"\u003cp\u003eDeep Learning (DL) has revolutionized the analysis of whole slide images (WSIs) in oncology, offering unprecedented capabilities in cancer diagnosis, prognosis, and treatment planning [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. By leveraging complex neural networks, DL models can process high-resolution WSIs to identify intricate patterns and features that may be challenging for human observers to discern [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. In recent years, substantial progress has been made in applying DL techniques to WSIs, such as the development of sophisticated algorithms capable of segmenting tumor regions, classifying cancer subtypes, and predicting patient outcomes based on histopathological data [\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. These rapid advancements have exhibited tremendous clinical potential across cancer types, enhancing diagnostic precision and prognostic evaluations in myriad cancers, including cancers of the skin [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], lung [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], colon [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], cervix [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], prostate [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], and breast [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Despite these advancements, the explainability and interpretability remain key challenges in applying DL to cancer WSIs, where the \u0026ldquo;black box\u0026rdquo; nature of DL models renders them difficult to understand and interpret, ultimately creating a barrier to the incorporation of DL models into clinical practice [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e][\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e][\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBreast cancer is one of the most commonly diagnosed cancer worldwide [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], and is steadily increasing in incidence, most notably in women under 50 [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Invasive breast carcinoma (IBC) is a type of breast cancer that has spread beyond the ducts or lobules of the breast into the surrounding breast tissue. As IBC has the potential to spread to regional draining lymph nodes and other distant parts of the body [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], early detection and diagnosis is paramount to survival. IBC is typically detected and diagnosed through a combination of imaging, histopathological examination of needle biopsy samples, and molecular profiling, which helps determine the appropriate course of treatment [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], where gene expression profiles in both tumor tissues and extratumoral microenvironment have been shown to predict outcomes [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e][\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Importantly, given the high degree of genetic heterogeneity, both spatially and temporally within individual tumors and across different patients [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e][\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e][\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e], there is a pressing need to expand biomarker profiles to gain a deeper understanding of genetic heterogeneity and achieve the goals of precision medicine\u0026mdash;providing targeted treatments based on specific biomarkers.\u003c/p\u003e \u003cp\u003eThis urgency has accelerated the incorporation of DL into breast cancer research, fueling technological advancements [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e][\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. To this end, convolutional neural networks (CNNs) have recently been utilized to classify histopathological images into categories such as benign or malignant, and to identify specific subtypes of breast cancer such as ductal, lobular, mucinous, and papillary carcinomas [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Moreover, hybrid models combining multiple DL architectures have demonstrated improved performance in classifying breast cancer from histopathological images [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e][\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Beyond classification, DL approaches have been applied to predict gene expression profiles from WSIs to aid in the determination of molecular subtypes and to predict early stage breast cancer recurrence using hematoxylin and eosin-stained WSIs. These techniques approximate traditional gene-based methods and provide a faster route to molecular characterization, underscoring the crucial role of combining DL and gene expression in the selection of appropriate postoperative treatments [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFurthermore, the ongoing issue of explainability has catalyzed development of interpretable models that provide insights into their decision-making processes and is a critical area of ongoing research [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. While interpretable DL models applied to WSIs offer transformative potential for predicting gene expression profiles in breast cancer [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], the selection of genes for model training poses a major challenge and has often relied on predefined gene sets, such as those identified through Gene Set Enrichment Analysis (GSEA) [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] or established panels like PAM50 profiles [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e][\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Although these approaches provide valuable insights into pathway-level perturbations and tumor subtyping, they may constrain the identification of novel, actionable, independent biomarkers that do not align with predefined pathways or signatures. In contrast, by predicting gene expression profiles directly from WSI features and subsequently integrating these results into breast cancer pathways, we can create a more data-driven and holistic framework in which each gene is represented equally, irrespective of prior biological assumptions.\u003c/p\u003e \u003cp\u003eThus, to expand biomarker profiles in breast cancer research using DL and address barriers in explainability in DL models, we created a transparent DL pipeline \u0026lsquo;XpressO\u0026rsquo; that i) demonstrates a step-by-step workflow with customized scripts implemented to bridge the gaps between publicly available DL methodologies, and ii) uses RNA expression data from the tissue of each WSI to create a spatial heat map that visualizes the predicted expression of biomarkers based on tumor phenotype. These visualizations highlight the probability of high or low expression of genes in specific regions of interest on WSIs, potentially bypassing traditional and time-consuming histopathological techniques and making predictions interpretable for clinical applications.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003e\u003cstrong\u003eCollection and Processing of WSIs\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe explainable DL pipeline created in the present study can be seen in Figure 1. First, we accessed WSIs of IBC from 191 patients from The Cancer Genome Atlas (TCGA) web-portal [37]. RNA-seq data for each of the 191 WSIs were previously derived by isolating and sequencing RNA from WSI tissues. For the present study. RNA-seq data were downloaded in the form of fragments per kilobase per million reads (FPKM) values from the TCGA portal. The DL-based weakly supervised algorithm Clustering-constrained Attention Multiple Instance Learning (CLAM) was employed to segment tumor regions of interest (ROIs) on the WSIs [38]. CLAM leverages attention mechanisms to automatically detect sub-regions with high diagnostic significance for accurate whole-slide classification. Additionally, it employs instance-level clustering on the identified representative regions to constrain and refine the feature space, enhancing model interpretability, and performance.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAs scanned WSIs result in large files and high-resolution images, they were divided into smaller and more manageable patches (~1000 patches per WSI) using the \u0026lsquo;opencv\u0026rsquo; module in Python [39]. Using a sliding window approach, the ROI mask guided the extraction of fixed-size square patches (e.g., 256 \u0026times; 256 pixels) at the desired magnification level of 20x.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eTo extract features from these masks, we utilized a pre-trained Unified Network for Instance-level Representation Learning (UNI) model (Figure 1) [40]. The pre-trained UNI model is a weakly supervised Vision Transformer (ViT-L/16) [41] via DINOv2 [42], optimized for histopathological image analysis, enabling robust extraction of high-dimensional features (i.e., feature embeddings from WSI patches). The resulting feature embeddings offered a computationally efficient and accurate representation of histopathological patterns that were fed into the downstream task of classification of gene expression for the WSIs (Figure 1).\u003c/p\u003e\n\u003ch4\u003e\u003cstrong\u003eCollection and Processing of RNA-seq data\u003c/strong\u003e\u003c/h4\u003e\n\u003cp\u003eThe gene expression (i.e., RNA-seq) data for the 191 WSIs of IBC tissues were downloaded from The Cancer Genome Atlas (TCGA) in the form of fragments per kilobase per million reads (FPKMs). Genes with FPKM values not reported in n\u0026ge;1 samples or with FPKM values \u0026le;5 were filtered out. This filtration eliminated the missing and low value FPKMs that could potentially hinder interpretability of the respective genes by the model. Further, a literature review was conducted to identify both common and novel genes of interest that could be expressed at either lower or higher levels in breast tumors than in breast normal tissues [45-93].\u003c/p\u003e\n\u003cp\u003eFPKM values were then binarized into \u0026quot;high\u0026quot; and \u0026quot;low\u0026quot; categories based on the median expression value of each selected gene (label) using a custom script [43]. Following the binarization of FPKMs, the dataset was split into training, validation, and testing subsets using a predefined script that was run for a range of \u0026lsquo;k\u0026rsquo; folds (i.e. k=[4, 14] folds) for each gene of interest. For instance, for a k-fold of 10, the dataset was split into 80:10:10 ratio, defining 80% for training, 10% for validation (for k-fold cross-validation during the training round), and 10% for the testing set. The binary label mappings of high (\u0026ldquo;1\u0026rdquo;) and low (\u0026ldquo;0\u0026rdquo;) expressions were applied to training, validation, and testing sets depending on the tissue slides (or WSIs) that were populated in each of these sets. This prepared the slide-level labels ready for the task of gene expression classification, marking each WSI as either high-expression or low-expression based on the binarized gene expression levels assigned to them.\u003c/p\u003e\n\u003ch4\u003e\u003cstrong\u003eModel Training\u003c/strong\u003e\u003c/h4\u003e\n\u003cp\u003eThe CLAM SB model was used for the classification of the binarized gene expressions using \u0026nbsp;the feature-embeddings extracted from the ROIs of tissue slides [44]. The training phase involved weakly supervised learning, using slide-level labels without specific annotations for individual patches. This enabled the model to automatically identify patches within the ROIs of each slide for which feature-embeddings may or may not be contributing to the high or low expression of the relevant biomarker (or gene). Training parameters were set as follows: Learning rate = 2\u0026times;10\u003csup\u003e\u0026minus;4\u003c/sup\u003e, Epochs = 200, Optimizer = Adam, Dropout = 25%, and cross validation-fold range = [4, 12].\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel Evaluation and Metrics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe trained model was evaluated on independent test sets for the genes of interest based on the data-split based on the best performing \u0026lsquo;k\u0026rsquo; fold during the training round (e.g., the test set size would differ for the k =10 [80:10:10=training:validation:testing ratio] vs. k=12 [76:12:12 ratio]). The performance metrics such as area under the receiver operating characteristic curve (\u003cstrong\u003eAUC-ROC)\u003c/strong\u003e, \u003cstrong\u003eaccuracy, precision\u0026nbsp;\u003c/strong\u003e(i.e., how many of the predicted positive instances are actually positive)\u003cstrong\u003e, recall\u0026nbsp;\u003c/strong\u003e(i.e., how many of the actual positive instances the model is able to correctly identify)\u003cstrong\u003e, and F1-score\u0026nbsp;\u003c/strong\u003e(i.e., the harmonic mean of precision and recall)\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003ewere assessed for the testing set for each gene of interest.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVisualization the Distribution of High and Low Gene Expression on WSIs\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGene expression heatmaps were generated on the testing set (WSIs) using the heatmap module pf CLAM to visualize the attention areas identified on WSIs by the CLAM SB model. Attention scores to depict aggregated probabilities (p0 or p1) of expression were constructed for the top ten highest performing WSI-patches. The scores were subsequently overlaid on the original WSIs and interpreted further, highlighting the regions of high diagnostic significance for each gene of interest.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003eGenes of interest\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBased on our literature search for both known and novel genes of interest that have been shown to be significantly under- or overexpressed in breast tumors compared to normal breast tissues, we identified an initial set of 36 genes (Table 1).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable 1: Genes found to be under- or over-expressed in breast tumors in the literature.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"624\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eGene Symbol\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eFull Gene Name\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eReferences\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eMMP11\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eMatrix metalloproteinase 11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[45][46]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eFCGR2A\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eFc gamma receptor IIa\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[47]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eHER2\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eHuman epidermal growth factor receptor 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[48][49]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eESR1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eEstrogen receptor 1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[50][51]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eHLA-DRA\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eMajor histocompatibility complex, class II, DR alpha\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[52]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eC3AR1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eComplement component 3a receptor 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[53]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eFYB\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eFYN binding protein 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[54]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eRRM2\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eRibonucleotide reductase M2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[55]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003ePLAU\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003ePlasminogen activator, urokinase\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[56]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eCOL1A1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eCollagen alpha 1 type 1 chain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[57][58]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eKRT81\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eKeratin 81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[59][60]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eS100A7\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eS100 calcium binding protein A7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[61][62]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eSCGB1D2\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eSecretoglobin family 1D member 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[63][64]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eKRT18\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eKeratin 18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[65]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eNOTCH3\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eNotch receptor 3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[66][67]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003ePIK3CA\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003ePhosphatidylinositol-4,5-bisphosphate 3-kinase catalytic subunit alpha\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[68]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eTP53\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eTumor protein 53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[69]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eKMT2C\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eLysine-specific methyltransferase 2C\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[70]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eMAP3K1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eMitogen-activated protein kinase kinase kinase 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[71]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eGATA3\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eGATA binding protein 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[72]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003ePTEN\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003ePhosphatase and tensin homolog\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[73]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eSF3B1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eSplicing factor 3b subunit 1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[74]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eCOL11A1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eCollagen type XI alpha 1 chain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[75][76]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eCOL10A1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eCollagen Type X Alpha 1 Chain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[77]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003ePCOLCE2\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eProcollagen C-endopeptidase enhancer 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[78]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eLAMA2\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eLaminin subunit alpha 2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[79][80]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eTMTC1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eTransmembrane O-mannosyltransferase targeting cadherins 1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[81]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eADAMTS5\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eADAM metallopeptidase with thrombospondin type 1 motif 5\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[82]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eTIMP4\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eTIMP metallopeptidase inhibitor 4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[83]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eRSPO3\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eR-spondin 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[84]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eOR2W3\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eOlfactory receptor family 2 subfamily W member 3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[85][86]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eOR2B6\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eOlfactory Receptor Family 2 Subfamily B Member 6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[85]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eOR2T8\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eOlfactory Receptor Family 2 Subfamily T Member 8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[85]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eMMP3\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eMatrix metalloproteinase-3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[87][88]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003eKLK8\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eKallikrein-related peptidase 8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[89][90]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e\u003cem\u003ePTPRZ1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003eProtein tyrosine phosphatase receptor type Z1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 208px;\"\u003e\n \u003cp\u003e[91][92][93]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAfter filtering genes with FPKM values that were either not reported in n\u0026ge;1 samples or reported at \u0026le;5 FPKM, the following nine genes were retained in the FPKM file for further analyses: \u003cem\u003eESR1\u0026nbsp;\u003c/em\u003e(estrogen receptor 1), \u003cem\u003eS100A7\u0026nbsp;\u003c/em\u003e(S100 calcium binding protein A7), \u003cem\u003eSCGB1D2\u0026nbsp;\u003c/em\u003e(secretoglobin family 1D member 2), \u003cem\u003eCOL11A1\u0026nbsp;\u003c/em\u003e(collagen type XI alpha 1), \u003cem\u003eGATA3\u0026nbsp;\u003c/em\u003e(GATA binding protein 3), \u003cem\u003eTIMP4\u0026nbsp;\u003c/em\u003e(tissue inhibitor of metalloproteinases 4), \u003cem\u003eMMP3\u0026nbsp;\u003c/em\u003e(matrix metallopeptidase 3), \u003cem\u003eKLK8\u003c/em\u003e (kallikrein-related peptidase 8), and \u003cem\u003ePTPRZ1\u0026nbsp;\u003c/em\u003e(protein tyrosine phosphatase receptor type Z1).\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel Evaluation of Gene Expression\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe performance metrics for the predicted expression of nine genes from the testing set can be found in Table 2. The model successfully predicted the expression of \u003cem\u003eCOL11A1\u003c/em\u003e, \u003cem\u003eKLK8\u003c/em\u003e, and \u003cem\u003eGATA3\u003c/em\u003e, with AUC values \u0026gt;0.8 (95% CI: [0.67, 1.0], [0.65, 1.0], and [0.79, 1.0], respectively), indicating strong predictive performance. \u003cem\u003eCOL11A1\u003c/em\u003e and \u003cem\u003eGATA3\u003c/em\u003e had precision, recall, and F1 scores \u0026gt;0.8, while \u003cem\u003eKLK8\u003c/em\u003e exhibited slightly lower values between 0.6 and 0.8. Additionally, \u003cem\u003eESR1\u003c/em\u003e and \u003cem\u003eS100A7\u003c/em\u003e were predicted at a moderate AUC of 0.75 (95% CI: [0.52, 0.99] and [0.53, 0.97], respectively) with considerably high performance metrics (precision, recall, and F1 \u0026ge; 0.7). Other genes performed with AUC values ranging from 0.69 to 0.73, where \u003cem\u003eTIMP4\u003c/em\u003e exhibited the highest precision at 0.84.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u003c/strong\u003e: Performance metrics for the predicted expression of nine genes in the testing set as a function of the best performing kth fold for each gene.\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBiomarker\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBest performing fold\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC CI range\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTest Error\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cem\u003eGATA3\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e11th\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e(0.79, 1.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cem\u003eCOL11A1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e11th\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e(0.68, 1.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cem\u003eKLK8\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e5th\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e(0.65, 1.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cem\u003eESR1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e11th\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e(0.52, 0.99)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cem\u003eS100A7\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e10th\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e(0.53, 0.97)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.72\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cem\u003eTIMP4\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e7th\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e(0.45, 1.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.73\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cem\u003eMMP3\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e7th\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e(0.48, 0.95)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cem\u003ePTPRZ1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e9th\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e(0.49, 0.95)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cem\u003eSCGB1D2\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e9th\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e(0.43, 0.95)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAUC: area under the curve; CI: confidence interval\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePrediction and Visualization of Gene Expression on WSIs\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe attention scores from the heatmap module of CLAM depicted aggregated probabilities (p0 or p1) of ROIs across the top 10 high-performing patches for the classification of high versus low gene expression for WSIs belonging to the testing set of the respective gene or biomarker (Table S1). The resulting attention maps (i.e., heatmaps) aided the evaluation of heterogeneity of biomarker expression within and across several WSIs (Figures2-4).\u003c/p\u003e\n\u003cp\u003eThe heatmaps showing predicted biomarker expression across WSIs can be seen in Figure 2, while WSI names and aggregated probabilities for both high and low expression are reported in Table S1. Our analyses show that \u003cem\u003eCOL11A1\u003c/em\u003e was predicted to be expressed consistently high in the test tissues. \u003cem\u003eKLK8\u003c/em\u003e was predicted to be expressed at low levels across the test tissues, except for WSI \u0026lsquo;TCGA-BH-A0BM\u0026rsquo;, where it exhibited moderately high predicted expression (68.5%). The predicted expression of \u003cem\u003eMMP3\u003c/em\u003e varied across WSIs, whereas the expression of \u003cem\u003eTIMP4\u003c/em\u003e was predicted to be consistently high across the test tissues. More interestingly, the two tissue pairs \u0026lsquo;TCGA-E2-A15H\u0026rsquo; and \u0026lsquo;TCGA-BH-A0BP\u0026rsquo; exhibited opposite predicted expression patterns for \u003cem\u003eTIMP4\u003c/em\u003e and \u003cem\u003eMMP3\u003c/em\u003e (Figure 4 a and b).\u003cem\u003e\u0026nbsp;\u003c/em\u003eAdditionally, \u003cem\u003eGATA3\u003c/em\u003e exhibited varying levels of predicted expression across the test tissues, while \u003cem\u003eESR1\u003c/em\u003e exhibited mainly low predicted expression, except for high predicted expression in WSI \u0026lsquo;TCGA-E2-A1L8\u0026rsquo; (76%) probability. Notably, this WSI was found to be positive for estrogen receptor (ER) expression per IHC results in TCGA. Most interestingly, the expression of both \u003cem\u003eGATA3\u003c/em\u003e and \u003cem\u003eESR1\u003c/em\u003e were predicted to be low in the two tissues \u0026lsquo;TCGA-BH-A0RX\u0026rsquo;and \u0026lsquo;TCGA-E2-A150\u0026rsquo; and exhibited similar patterns of low expression (Figure 3 a and b). In agreement with these findings, both of these WSIs were found to be negative for ER expression according to the IHC results available in TCGA.\u003c/p\u003e\n\u003cp\u003eFurthermore, as shown in Figure 2 and Table S1, the predicted expression of \u003cem\u003ePTPRZ1\u003c/em\u003e was mainly low across the test tissues except for one tissue (\u0026ldquo;TCGA-AN-A0FN\u0026rdquo;), in which it was predicted to have high levels of expression. \u003cem\u003eS100A7\u003c/em\u003e was predicted to be expressed at high levels in two out of four test tissues, whereas the remaining two exhibited a mixed level of predicted expression. Lastly, \u003cem\u003eSCGB1D2\u003c/em\u003e oscillated between high and low predicted expression across several tissues.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study presents a novel deep learning (DL) pipeline that integrates multiple computational techniques to predict gene expression directly from WSIs of IBC tumors, offering a high-throughput alternative to traditional bulk transcriptomics. By leveraging weakly supervised learning and attention-based interpretability methods, our approach enables the spatial mapping of genetic heterogeneity within and across tumor tissues based on phenotypic characteristics. Notably, the model achieved high predictive performance for key biomarkers such as \u003cem\u003eCOL11A1\u003c/em\u003e, \u003cem\u003eKLK8\u003c/em\u003e, and \u003cem\u003eGATA3\u003c/em\u003e, which are associated with tumor aggressiveness and poor prognosis.\u003c/p\u003e \u003cp\u003eThe high expression of \u003cem\u003eCOL11A1\u003c/em\u003e has been significantly correlated with aggressive tumors with poor prognosis [\u003cspan citationid=\"CR94\" class=\"CitationRef\"\u003e94\u003c/span\u003e][\u003cspan citationid=\"CR95\" class=\"CitationRef\"\u003e95\u003c/span\u003e]. In agreement with the literature, we observed both high and low predicted gene expression of \u003cem\u003eKLK8\u003c/em\u003e, which has been shown to be differentially expressed in breast cancer cells [\u003cspan additionalcitationids=\"CR97\" citationid=\"CR96\" class=\"CitationRef\"\u003e96\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR98\" class=\"CitationRef\"\u003e98\u003c/span\u003e]. Additionally, \u003cem\u003eMMP3\u003c/em\u003e exhibited varied expression, whereas \u003cem\u003eTIMP4\u003c/em\u003e exhibited consistently high predicted expression in WSIs in the present study. As the TIMP4 protein is known to inhibit MMP3 and prevent tumor growth [\u003cspan citationid=\"CR99\" class=\"CitationRef\"\u003e99\u003c/span\u003e][\u003cspan citationid=\"CR100\" class=\"CitationRef\"\u003e100\u003c/span\u003e], the high expression of the \u003cem\u003eTIMP4\u003c/em\u003e gene could indicate an active role in suppressing tumor growth and metastasis. \u003cem\u003eGATA3\u003c/em\u003e exhibited varied expression in several tissues, whereas \u003cem\u003eESR1\u003c/em\u003e exhibited low expression consistently across the tissues. Notably, the low predicted expression of \u003cem\u003eGATA3\u003c/em\u003e and \u003cem\u003eESR1\u003c/em\u003e coincided between two WSIs (i.e., TCGA-BH-A0RX and TCGA-E2-A150). As \u003cem\u003eGATA3\u003c/em\u003e is an \u003cem\u003eESR1\u003c/em\u003e co-operating transcriptional factor, their expression has been shown to be correlated in breast cancer tissues [\u003cspan citationid=\"CR101\" class=\"CitationRef\"\u003e101\u003c/span\u003e][\u003cspan citationid=\"CR102\" class=\"CitationRef\"\u003e102\u003c/span\u003e]. GATA3 and ESR1 were consistently co-predicted with low expression across multiple tissue samples, which aligns with their transcriptional co-regulation in breast cancer. Given that GATA3 is a known cooperating transcription factor for ESR1, their simultaneous low expression patterns could indicate a transition toward an ESR1-independent phenotype, which is often linked to endocrine therapy resistance. This could have significant implications for hormone-based treatments, such as tamoxifen or fulvestrant, which rely on ESR1 activity for efficacy.\u003c/p\u003e \u003cp\u003eIn addition, \u003cem\u003ePTPRZ1\u003c/em\u003e\u0026mdash;which was predicted to be expressed at low levels across most of the tissues in the present study\u0026mdash;is a tumor suppressor, and thus low expression could indicate higher cancer cell proliferation and invasion and has exhibited lower expression in a breast cancer cohort than a normal breast cohort [\u003cspan citationid=\"CR103\" class=\"CitationRef\"\u003e103\u003c/span\u003e]. High expression of \u003cem\u003eS100A7\u003c/em\u003e\u0026mdash;which exhibited a high probability of expression in two WSIs and mixed expression levels in others\u0026mdash;has been previously correlated with tumor growth and may lead to metastasis in tissues with high levels of expression [\u003cspan citationid=\"CR104\" class=\"CitationRef\"\u003e104\u003c/span\u003e]. Lastly, in agreement with the literature, the predicted expression of \u003cem\u003eSCGB1D2\u003c/em\u003e was variable in the present study, as \u003cem\u003eSCGB1D2\u003c/em\u003e has been shown to be highly expressed in ductal carcinoma \u003cem\u003ein situ\u003c/em\u003e (DCIS) breast cancer with invasive capability compared to the DCIS that is not invasive [\u003cspan citationid=\"CR105\" class=\"CitationRef\"\u003e105\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBeyond prediction accuracy and predictions of localized gene expression, our DL pipeline enhances model explainability, bridging the gap between existing computational pathology tools and their clinical applications. By integrating a pre-trained UNI model for feature extraction, CLAM for region-based classification and interpretation of gene-expression categorized using custom scripts, this framework provides clinicians with intuitive visualizations of predicted gene expression patterns on WSIs, potentially aiding both diagnostic decision-making and targeted therapeutic development. This method reduces reliance on labor- and time-intensive immunohistochemistry and cytochemical methodologies, which require separate staining and imaging for each biomarker, making multiplexed analysis difficult. By leveraging DL-based \u003cem\u003ein silico\u003c/em\u003e predictions, our approach enables the simultaneous assessment of multiple biomarkers on a single WSI, substantially improving efficiency in histopathology workflows and enhancing high-throughput molecular characterization in a scalable and cost-effective manner.\u003c/p\u003e \u003cp\u003eIt is important to note that the present study has several limitations. IHC results for the eight out of nine biomarkers identified in the present study are not available from TCGA, and thus we were unable to validate the spatial gene expression predictions for those biomarkers on the tumor patches of the respective WSIs. Additionally, TCGA also does not have spatial transcriptomics data to validate the gene expression for spatial patches. Future work will close this gap by conducting extensive IHC or spatial transcriptomics on in-house breast tumor samples to clinically validate XpressO in the near future.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThe ability to predict region-specific gene expression patterns on WSIs suggests that the \u0026lsquo;XpressO\u0026rsquo; pipeline created in the present study could be a valuable tool for biomarker discovery and precision oncology. Our findings reveal novel predicted co-expression patterns of key biomarkers using tissue phenotypes, suggesting that tissue phenotypes can serve as surrogate biomarkers for potential mechanistic insights into IBC progression and treatment resistance regulated by the relevant genes. Future work will focus on validating these findings through spatial transcriptomics or extensive IHC staining, allowing for a region-of-interest (ROI)-specific assessment of gene expression in IBC tissues. Additionally, the freely available online resource developed here will facilitate broader access to the \u0026lsquo;XpressO\u0026rsquo; pipeline, ensuring seamless integration into pathology labs and pharmaceutical research for biomarker-driven precision medicine.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eCompeting Interests\u003c/h2\u003e\n\u003cp\u003eJH was paid partially as an editor/writer of this manuscript.\u003c/p\u003e\n\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\n\u003cp\u003eSS conceptualized the idea, participated in the construction of data analysis pipeline, interpreted results and contributed majorly to the writing of the manuscript. DY participated in the construction and execution of the data analysis pipeline and contributed to the writing of the manuscript. VR participated in the execution of the data analysis pipeline and contributed majorly to the writing of the manuscript. MK participated in the execution of the data analysis pipeline and reviewed the manuscript. KM reviewed and edited the manuscript. JH contributed majorly to the writing of the manuscript.\u003c/p\u003e\n\u003ch2\u003eAcknowledgement\u003c/h2\u003e\n\u003cp\u003eThis study was funded partially by William H. Neukom 1964 Institute for Computational Science. The funder played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript.\u003c/p\u003e\n\u003ch2\u003eData Availability\u003c/h2\u003e\n\u003cp\u003eAll data generated or analysed during this study are either included in this published article [and its supplementary information files] or at the repository: https://github.com/skr1/XpressO\u003c/p\u003e"},{"header":"References","content":"\u003col start=\"1\" type=\"1\"\u003e\n\u003cli\u003eTran, Khoa A., et al. \u0026quot;Deep learning in cancer diagnosis, prognosis and treatment selection.\u0026quot; Genome Medicine 13 (2021): 1-17.\u003c/li\u003e\n\u003cli\u003eLi, Mengfang, et al. \u0026quot;Medical image analysis using deep learning algorithms.\u0026quot; Frontiers in Public Health 11 (2023): 1273253.\u003c/li\u003e\n\u003cli\u003eMostafa, Ghada, et al. \u0026quot;The power of deep learning in simplifying feature selection for hepatocellular carcinoma: a review.\u0026quot; BMC Medical Informatics and Decision Making 24.1 (2024): 287.\u003c/li\u003e\n\u003cli\u003eWang, Ching-Wei, et al. \u0026quot;Deep learning for endometrial cancer subtyping and predicting tumor mutational burden from histopathological slides.\u0026quot; npj Precision Oncology 8.1 (2024): 287.\u003c/li\u003e\n\u003cli\u003eWang, Jun, et al. \u0026quot;Advances in Multiple Instance Learning for Whole Slide Image Analysis: Techniques, Challenges, and Future Directions.\u0026quot; arXiv preprint arXiv:2408.09476 (2024).\u003c/li\u003e\n\u003cli\u003e\u0026quot;High-fidelity detection, subtyping, and localization of five skin neoplasms using supervised and semi-supervised learning.\u0026quot; Journal of Pathology Informatics 14 (2023): 100159.\u003c/li\u003e\n\u003cli\u003eWang, Xi, et al. \u0026quot;Weakly supervised deep learning for whole slide lung cancer image analysis.\u0026quot; IEEE transactions on cybernetics 50.9 (2019): 3950-3962\u003c/li\u003e\n\u003cli\u003eHamida, A. Ben, et al. \u0026quot;Deep learning for colon cancer histopathological images analysis.\u0026quot; Computers in Biology and Medicine 136 (2021): 104730.\u003c/li\u003e\n\u003cli\u003eCheng, Shenghua, et al. \u0026quot;Robust whole slide image analysis for cervical cancer screening using deep learning.\u0026quot; Nature communications 12.1 (2021): 5639.\u003c/li\u003e\n\u003cli\u003eSinghal, Nitin, et al. \u0026quot;A deep learning system for prostate cancer diagnosis and grading in whole slide images of core needle biopsies.\u0026quot; Scientific reports 12.1 (2022): 3383.\u003c/li\u003e\n\u003cli\u003eJiang, Bitao, et al. \u0026quot;Deep learning applications in breast cancer histopathological imaging: diagnosis, treatment, and prognosis.\u0026quot; Breast Cancer Research 26.1 (2024): 137.\u003c/li\u003e\n\u003cli\u003eZeiser, Felipe Andr\u0026eacute;, et al. \u0026quot;DeepBatch: A hybrid deep learning model for interpretable diagnosis of breast cancer in whole-slide images.\u0026quot; Expert Systems with Applications 185 (2021): 115586.\u003c/li\u003e\n\u003cli\u003eTeng, Qiaoying, et al. \u0026quot;A survey on the interpretability of deep learning in medical diagnosis.\u0026quot; Multimedia Systems 28.6 (2022): 2335-2355.\u003c/li\u003e\n\u003cli\u003eDhar, Tribikram, et al. \u0026quot;Challenges of deep learning in medical image analysis\u0026mdash;improving explainability and trust.\u0026quot; IEEE Transactions on Technology and Society 4.1 (2023): 68-75.\u003c/li\u003e\n\u003cli\u003eSalahuddin, Zohaib, et al. \u0026quot;Transparency of deep neural networks for medical image analysis: A review of interpretability methods.\u0026quot; Computers in biology and medicine 140 (2022): 105111.\u003c/li\u003e\n\u003cli\u003eArnold, Melina, et al. \u0026quot;Current and future burden of breast cancer: Global statistics for 2020 and 2040.\u0026quot; The Breast 66 (2022): 15-23.\u003c/li\u003e\n\u003cli\u003eAmerican Cancer Society. Cancer Facts and Figures 2025. Atlanta: American Cancer Society; 2025.\u003c/li\u003e\n\u003cli\u003eSrikantamurthy, Mahati Munikoti, et al. \u0026quot;Classification of benign and malignant subtypes of breast cancer histopathology imaging using hybrid CNN-LSTM based transfer learning.\u0026quot; BMC Medical Imaging 23.1 (2023): 19.\u003c/li\u003e\n\u003cli\u003eQasrawi, Radwan, et al. \u0026quot;Hybrid ensemble deep learning model for advancing breast cancer detection and classification in clinical applications.\u0026quot; Heliyon 10.19 (2024).\u003c/li\u003e\n\u003cli\u003eRom\u0026aacute;n-P\u0026eacute;rez, Erick, et al. \u0026quot;Gene expression in extratumoral microenvironment predicts clinical outcome in breast cancer patients.\u0026quot; Breast Cancer Research 14 (2012): 1-12.\u003c/li\u003e\n\u003cli\u003eHuang, Erich et al. \u0026ldquo;Gene expression predictors of breast cancer outcomes.\u0026rdquo; Lancet (London, England) vol. 361,9369 (2003): 1590-6. doi:10.1016/S0140-6736(03)13308-9.\u003c/li\u003e\n\u003cli\u003eTurashvili, Gulisa, and Edi Brogi. \u0026quot;Tumor heterogeneity in breast cancer.\u0026quot; Frontiers in medicine 4 (2017): 227.\u003c/li\u003e\n\u003cli\u003eL\u0026uuml;\u0026ouml;nd, Fabiana, Stefanie Tiede, and Gerhard Christofori. \u0026quot;Breast cancer as an example of tumour heterogeneity and tumour cell plasticity during malignant progression.\u0026quot; British journal of cancer 125.2 (2021): 164-175.\u003c/li\u003e\n\u003cli\u003eKudelova, Eva, et al. \u0026quot;Genetic heterogeneity, tumor microenvironment and immunotherapy in triple-negative breast cancer.\u0026quot; International Journal of Molecular Sciences 23.23 (2022): 14937. \u003c/li\u003e\n\u003cli\u003eChugh, Gunjan, Shailender Kumar, and Nanhay Singh. \u0026quot;Survey on machine learning and deep learning applications in breast cancer diagnosis.\u0026quot; Cognitive Computation 13.6 (2021): 1451-1470.\u003c/li\u003e\n\u003cli\u003eDar, Rayees Ahmad, Muzafar Rasool, and Assif Assad. \u0026quot;Breast cancer detection using deep learning: Datasets, methods, and challenges ahead.\u0026quot; Computers in biology and medicine 149 (2022): 106073.\u003c/li\u003e\n\u003cli\u003eZerouaoui, Hasnae, and Ali Idri. \u0026quot;Deep hybrid architectures for binary classification of medical breast cancer images.\u0026quot; Biomedical Signal Processing and Control 71 (2022): 103226.\u003c/li\u003e\n\u003cli\u003eLee, Sung Hak, and Hyun-Jong Jang. \u0026quot;Deep learning-based prediction of molecular cancer biomarkers from tissue slides: A new tool for precision oncology.\u0026quot; Clinical and Molecular Hepatology 28.4 (2022): 754.\u003c/li\u003e\n\u003cli\u003eLee, Geongyu, et al. \u0026quot;Predicting the risk of early-stage breast cancer recurrence using H\\\u0026amp;E-stained tissue images.\u0026quot; arXiv preprint arXiv:2406.06650 (2024).\u003c/li\u003e\n\u003cli\u003eMuhammad, Dost, and Malika Bendechache. \u0026quot;Unveiling the black box: A systematic review of Explainable Artificial Intelligence in medical image analysis.\u0026quot; Computational and structural biotechnology journal (2024).\u003c/li\u003e\n\u003cli\u003eLevy-Jurgenson, Alona, et al. \u0026quot;Spatial transcriptomics inferred from pathology whole-slide images links tumor heterogeneity to survival in breast and lung cancer.\u0026quot; Scientific reports 10.1 (2020): 18802.\u003c/li\u003e\n\u003cli\u003eSchmauch, Beno\u0026icirc;t, et al. \u0026quot;A deep learning model to predict RNA-Seq expression of tumours from whole slide images.\u0026quot; Nature communications 11.1 (2020): 3877.\u003c/li\u003e\n\u003cli\u003eWang, Yinxi, et al. \u0026quot;Predicting molecular phenotypes from histopathology images: a transcriptome-wide expression\u0026ndash;morphology analysis in breast cancer.\u0026quot; Cancer research 81.19 (2021): 5115-5126.\u003c/li\u003e\n\u003cli\u003eParker, Joel S., et al. \u0026quot;Supervised risk predictor of breast cancer based on intrinsic subtypes.\u0026quot; Journal of clinical oncology 27.8 (2009): 1160-1167.\u003c/li\u003e\n\u003cli\u003eSJ, SCHNITT. \u0026quot;Pathology of invasive breast cancer.\u0026quot; Diseases of the Breast (2000).\u003c/li\u003e\n\u003cli\u003eAmerican College of Radiology, et al. \u0026quot;Standards for diagnosis and management of invasive breast carcinoma.\u0026quot; CA: A cancer journal for clinicians 48.2 (1998): 83-107.\u003c/li\u003e\n\u003cli\u003ehttps://www.cancer.gov/tcga\u003c/li\u003e\n\u003cli\u003eLu, Ming Y., et al. \u0026quot;Data-efficient and weakly supervised computational pathology on whole-slide images.\u0026quot; Nature biomedical engineering 5.6 (2021): 555-570.\u003c/li\u003e\n\u003cli\u003ehttps://www.python.org/\u003c/li\u003e\n\u003cli\u003eChu, Jiaming, et al. \u0026quot;UniParser: Multi-Human Parsing with Unified Correlation Representation Learning.\u0026quot; IEEE Transactions on Image Processing (2024).\u003c/li\u003e\n\u003cli\u003eChen, R.J., Ding, T., Lu, M.Y. et al. Towards a general-purpose foundation model for computational pathology. Nat Med 30, 850\u0026ndash;862 (2024). https://doi.org/10.1038/s41591-024-02857-3.\u003c/li\u003e\n\u003cli\u003eOquab, Maxime, et al. \u0026quot;Dinov2: Learning robust visual features without supervision.\u0026quot; arXiv preprint arXiv:2304.07193 (2023).\u003c/li\u003e\n\u003cli\u003eSukhadia, Shrey S., et al. \u0026quot;Machine Learning Enabled Prediction of Biologically Relevant Gene Expression Using CT‐Based Radiomic Features in Non‐Small Cell Lung Cancer.\u0026quot; Cancer Medicine 13.24 (2024): e70509.\u003c/li\u003e\n\u003cli\u003eFillioux, Leo, et al. \u0026quot;Structured state space models for multiple instance learning in digital pathology.\u0026quot; International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2023.\u003c/li\u003e\n\u003cli\u003eGonz\u0026aacute;lez de Vega, Raquel et al. \u0026ldquo;MMP-11 as a biomarker for metastatic breast cancer by immunohistochemical-assisted imaging mass spectrometry.\u0026rdquo; Analytical and bioanalytical chemistry vol. 411,3 (2019): 639-646. doi:10.1007/s00216-018-1365-3\u003c/li\u003e\n\u003cli\u003eZhuang, Ying et al. \u0026ldquo;MMP11 promotes the proliferation and progression of breast cancer through stabilizing Smad2 protein.\u0026rdquo; Oncology reports vol. 45,4 (2021):16. doi:10.3892/or.2021.7967\u003c/li\u003e\n\u003cli\u003eGavin, Patrick G et al. \u0026ldquo;Association of Polymorphisms in FCGR2A and FCGR3A With Degree of Trastuzumab Benefit in the Adjuvant Treatment of ERBB2/HER2- Positive Breast Cancer: Analysis of the NSABP B-31 Trial.\u0026rdquo; JAMA oncology vol. 3,3 (2017): 335-341. doi:10.1001/jamaoncol.2016.4884\u003c/li\u003e\n\u003cli\u003eKlocker, Eva Valentina, and Christoph Suppan. \u0026ldquo;Biomarkers in Her2- Positive Disease.\u0026rdquo; Breast care (Basel, Switzerland) vol. 15,6 (2020): 586-593. doi:10.1159/000512283\u003c/li\u003e\n\u003cli\u003eCarney, Walter P. \u0026ldquo;HER2 status is an important biomarker in guiding personalized HER2 therapy.\u0026rdquo; Personalized medicine vol. 2,4 (2005): 317-324. doi:10.2217/17410541.2.4.317\u003c/li\u003e\n\u003cli\u003eBrett, Jamie O., et al. \u0026quot;ESR1 mutation as an emerging clinical biomarker in metastatic hormone receptor-positive breast cancer.\u0026quot; Breast Cancer Research 23 (2021): 1-15.\u003c/li\u003e\n\u003cli\u003eCarausu, Marcela, et al. \u0026quot;ESR1 mutations: a new biomarker in breast cancer.\u0026quot; Expert review of molecular diagnostics 19.7 (2019): 599-611.\u003c/li\u003e\n\u003cli\u003eSaraiva, Diana P et al. \u0026ldquo;Expression of HLA-DR in Cytotoxic T Lymphocytes: A Validated Predictive Biomarker and a Potential Therapeutic Strategy in Breast Cancer.\u0026rdquo; Cancers vol. 13,15 3841. 30 Jul. 2021, doi:10.3390/cancers13153841\u003c/li\u003e\n\u003cli\u003eSong, Guiqin, et al. \u0026quot;Identification of aberrant gene expression during breast ductal carcinoma in situ progression to invasive ductal carcinoma.\u0026quot; Journal of International Medical Research 48.1 (2020): 0300060518815364.\u003c/li\u003e\n\u003cli\u003eChou, Ching-Wen et al. \u0026ldquo;Identified the novel resistant biomarkers for taxane-based therapy for triple-negative breast cancer.\u0026rdquo; International journal of medical sciences vol. 18,12 2521-2531. 26 Apr. 2021, doi:10.7150/ijms.59177\u003c/li\u003e\n\u003cli\u003eAbdel-Rahman, Manar Ahmed, Mena Mahfouz, and Hany Onsy Habashy. \u0026quot;RRM2 expression in different molecular subtypes of breast cancer and its prognostic significance.\u0026quot; Diagnostic Pathology 17 (2022): 1-8.\u003c/li\u003e\n\u003cli\u003eSarno, Federica et al. \u0026ldquo;Functional Validation of the Putative Oncogenic Activity of PLAU.\u0026rdquo; Biomedicines vol. 11,1 102. 30 Dec. 2022, doi:10.3390/biomedicines11010102\u003c/li\u003e\n\u003cli\u003eLiu, Jing et al. \u0026ldquo;Collagen 1A1 (COL1A1) promotes metastasis of breast cancer and is a potential therapeutic target.\u0026rdquo; Discovery medicine vol. 25,139 (2018): 211-223.\u003c/li\u003e\n\u003cli\u003eMa, Hon-Ping et al. \u0026ldquo;Collagen 1A1 (COL1A1) Is a Reliable Biomarker and Putative Therapeutic Target for Hepatocellular Carcinogenesis and Metastasis.\u0026rdquo; Cancers vol. 11,6 786. 7 Jun. 2019, doi:10.3390/cancers11060786\u003c/li\u003e\n\u003cli\u003eYan, Zhideng et al. \u0026ldquo;The prognostic marker KRT81 is involved in suppressing CD8 + T cells and predicts immunotherapy response for triple-negative breast cancer.\u0026rdquo; Cancer biology \u0026amp; therapy vol. 25,1 (2024): 2355705. doi:10.1080/15384047.2024.2355705\u003c/li\u003e\n\u003cli\u003eNanashima, Naoki, et al. \u0026quot;Hair keratin KRT81 is expressed in normal and breast cancer cells and contributes to their invasiveness.\u0026quot; Oncology reports 37.5 (2017): 2964-2970.\u003c/li\u003e\n\u003cli\u003eEmberley, Ethan D et al. \u0026ldquo;S100A7 and the progression of breast cancer.\u0026rdquo; Breast cancer research : BCR vol. 6,4 (2004): 153-9. doi:10.1186/bcr816\u003c/li\u003e\n\u003cli\u003eAl-Haddad, Sahar, et al. \u0026quot;Psoriasin (S100A7) expression and invasive breast cancer.\u0026quot; The American journal of pathology 155.6 (1999): 2057-2066.\u003c/li\u003e\n\u003cli\u003eZafrakas, Menelaos et al. \u0026ldquo;Expression analysis of mammaglobin A (SCGB2A2) and lipophilin B (SCGB1D2) in more than 300 human tumors and matching normal tissues reveals their co-expression in gynecologic malignancies.\u0026rdquo; BMC cancer vol. 6 88. 9 Apr. 2006, doi:10.1186/1471-2407-6-88\u003c/li\u003e\n\u003cli\u003eSharma, P., Bhattacharyya, D.K. \u0026amp; Kalita, J. Disease biomarker identification from gene network modules for metastasized breast cancer. Sci Rep 7, 1072 (2017). https://doi.org/10.1038/s41598-017-00996-x\u003c/li\u003e\n\u003cli\u003eTakan, Işıl et al. \u0026ldquo;\u0026quot;In the light of evolution:\u0026quot; keratins as exceptional tumor biomarkers.\u0026rdquo; PeerJ vol. 11 e15099. 17 Mar. 2023, doi:10.7717/peerj.15099\u003c/li\u003e\n\u003cli\u003eChen, Wen-Jia, et al. \u0026quot;NOTCH3 inhibits transcription factor ZEB1 expression and metastasis of breast cancer cells via transcriptionally upregulating miR-223.\u0026quot; Journal of Cancer 15.1 (2024): 192.\u003c/li\u003e\n\u003cli\u003eChen, Min-Na et al. \u0026ldquo;Notch3 restricts metastasis of breast cancers through regulation of the JAK/STAT5A signaling pathway.\u0026rdquo; BMC cancer vol. 23,1 1257. 20 Dec. 2023, doi:10.1186/s12885-023-11746-w\u003c/li\u003e\n\u003cli\u003eLiu, Yi-Rong et al. \u0026ldquo;PIK3CA mutations define favorable prognostic biomarkers in operable breast cancer: a systematic review and meta-analysis.\u0026rdquo; OncoTargets and therapy vol. 7 543-52. 10 Apr. 2014, doi:10.2147/OTT.S60115\u003c/li\u003e\n\u003cli\u003eSilwal-Pandit, Laxmi, et al. \u0026quot;TP53 mutation spectrum in breast cancer is subtype specific and has distinct prognostic relevance.\u0026quot; Clinical Cancer Research 20.13 (2014): 3569-3580.\u003c/li\u003e\n\u003cli\u003eLiu, Xinhua et al. \u0026ldquo;KMT2C is a potential biomarker of prognosis and chemotherapy sensitivity in breast cancer.\u0026rdquo; Breast cancer research and treatment vol. 189,2 (2021): 347- 361. doi:10.1007/s10549-021-06325-1\u003c/li\u003e\n\u003cli\u003eNarang, Binish. \u0026quot;MAP3K1 Identified as a Prognostic Biomarker in Breast Cancer After Multi-omics Bioinformatics Analysis.\u0026quot; (2023).\u003c/li\u003e\n\u003cli\u003eTakaku, Motoki et al. \u0026ldquo;GATA3 in Breast Cancer: Tumor Suppressor or Oncogene?.\u0026rdquo; Gene expression vol. 16,4 (2015): 163-8. doi:10.3727/105221615X14399878166113\u003c/li\u003e\n\u003cli\u003eDerkyi-Kwarteng, Leonard, et al. \u0026quot;A retrospective analysis suggests PTEN expression is associated with favorable clinicopathological features of breast cancer.\u0026quot; Scientific Reports 14.1 (2024): 21645.\u003c/li\u003e\n\u003cli\u003eMaguire, Sarah L., et al. \u0026quot;SF3B1 mutations constitute a novel therapeutic target in breast cancer.\u0026quot; The Journal of pathology 235.4 (2015): 571-580.\u003c/li\u003e\n\u003cli\u003eLuo, Qi et al. \u0026ldquo;COL11A1 serves as a biomarker for poor prognosis and correlates with immune infiltration in breast cancer.\u0026rdquo; Frontiers in genetics vol. 13 935860. 9 Sep. 2022, doi:10.3389/fgene.2022.935860\u003c/li\u003e\n\u003cli\u003eShi, Wenjie, et al. \u0026quot;COL11A1 as an novel biomarker for breast cancer with machine learning and immunohistochemistry validation.\u0026quot; Frontiers in Immunology 13 (2022): 937125.\u003c/li\u003e\n\u003cli\u003eZhang, Mingdi, et al. \u0026quot;Bioinformatics analysis of prognostic significance of COL10A1 in breast cancer.\u0026quot; Bioscience reports 40.2 (2020): BSR20193286.\u003c/li\u003e\n\u003cli\u003eThomas, Michael J., et al. \u0026quot;PCPE2: Expression of Multifunctional Extracellular Glycoprotein Associates With Diverse Cellular Functions.\u0026quot; Journal of Lipid Research (2024): 100664.\u003c/li\u003e\n\u003cli\u003eFu, Tao et al. \u0026ldquo;LAMC2 as a prognostic biomarker in human cancer: a systematic review and meta-analysis.\u0026rdquo; BMJ open vol. 12,11 e063682. 17 Nov. 2022, doi:10.1136/bmjopen-2022-063682\u003c/li\u003e\n\u003cli\u003eHu, J., and S. Li. \u0026quot;The role of Lama2 in cancer: Current perspectives.\u0026quot; Cancer Research Journal 10.4 (2022): 85-88.\u003c/li\u003e\n\u003cli\u003eAli, Rafat et al. \u0026ldquo;Identification of New Key Genes and Their Association with Breast Cancer Occurrence and Poor Survival Using In Silico and In Vitro Methods.\u0026rdquo; Biomedicines vol. 11,5 1271. 25 Apr. 2023, doi:10.3390/biomedicines11051271\u003c/li\u003e\n\u003cli\u003eTavousi, Najmeh, et al. \u0026quot;ADAMTS5 modulates breast cancer development as a diagnostic biomarker and potential tumor suppressor, regulating by BAIAP2-AS1, VTI1B, CRNDE, and hsa-miR-135b-3p: integrated systems biology and experimental approach.\u0026quot; (2022).\u003c/li\u003e\n\u003cli\u003eLiss, Michaelann, et al. \u0026quot;Tissue inhibitor of metalloproteinase-4 is elevated in early-stage breast cancers with accelerated progression and poor clinical course.\u0026quot; The American journal of pathology 175.3 (2009): 940-946.\u003c/li\u003e\n\u003cli\u003eTer Steege, Eline J., et al. \u0026quot;R-spondin-3 promotes proliferation and invasion of breast cancer cells independently of Wnt signaling.\u0026quot; Cancer Letters 568 (2023): 216301.\u003c/li\u003e\n\u003cli\u003eWeber, Lea, et al. \u0026quot;Olfactory receptors as biomarkers in human breast carcinoma tissues.\u0026quot; Frontiers in oncology 8 (2018): 33.\u003c/li\u003e\n\u003cli\u003eMasjedi, Shirin, Laurence J. Zwiebel, and Todd D. Giorgio. \u0026quot;Olfactory receptor gene abundance in invasive breast carcinoma.\u0026quot; Scientific reports 9.1 (2019): 13736..\u003c/li\u003e\n\u003cli\u003eArgote Camacho, Angela Ximena, et al. \u0026quot;Metalloproteinases 1 and 3 as potential biomarkers in breast cancer development.\u0026quot; International journal of molecular sciences 22.16 (2021): 9012.\u003c/li\u003e\n\u003cli\u003eŁawicki, Paweł, et al. \u0026quot;Plasma Levels of Metalloproteinase 3 (MMP-3) and Metalloproteinase 7 (MMP-7) as New Candidates for Tumor Biomarkers in Diagnostic of Breast Cancer Patients.\u0026quot; Journal of Clinical Medicine 12.7 (2023): 2618.\u003c/li\u003e\n\u003cli\u003eTailor, Prashant D., et al. \u0026quot;Diagnostic and prognostic biomarker potential of kallikrein family genes in different cancer types.\u0026quot; Oncotarget 9.25 (2018): 17876.\u003c/li\u003e\n\u003cli\u003eWatrowski, Rafał, et al. \u0026quot;Gene expression of kallikreins in breast cancer cell lines.\u0026quot; Anticancer Research 40.5 (2020): 2487-2495.\u003c/li\u003e\n\u003cli\u003eXia, Zhenkun, et al. \u0026quot;The expression, functions, interactions and prognostic values of PTPRZ1: a review and bioinformatic analysis.\u0026quot; Journal of Cancer 10.7 (2019): 1663.\u003c/li\u003e\n\u003cli\u003eThalor, Anamika, et al. \u0026quot;Machine learning assisted analysis of breast cancer gene expression profiles reveals novel potential prognostic biomarkers for triple-negative breast cancer.\u0026quot; Computational and structural biotechnology journal 20 (2022): 1618-1631.\u003c/li\u003e\n\u003cli\u003eChen, Min, et al. \u0026quot;Integrin \u0026alpha;6\u0026beta;4 Upregulates PTPRZ1 Through UCHL1-Mediated Hif-1\u0026alpha; Nuclear Accumulation to Promote Triple-Negative Breast Cancer Cell Invasive Properties.\u0026quot; Cancers 16.21 (2024): 3683.\u003c/li\u003e\n\u003cli\u003eLi, Xia, Yue Jin, and Jian Xue. \u0026quot;Unveiling Collagen\u0026rsquo;s Role in Breast Cancer: Insights into Expression Patterns, Functions and Clinical Implications.\u0026quot; International Journal of General Medicine (2024): 1773-1787.\u003c/li\u003e\n\u003cli\u003eLuo, Qi, et al. \u0026quot;COL11A1 serves as a biomarker for poor prognosis and correlates with immune infiltration in breast cancer.\u0026quot; Frontiers in Genetics 13 (2022): 935860.\u003c/li\u003e\n\u003cli\u003eYousef, G. M., et al. \u0026quot;Kallikrein gene downregulation in breast cancer.\u0026quot; British journal of cancer 90.1 (2004): 167-172.\u003c/li\u003e\n\u003cli\u003eWatrowski, Rafał, et al. \u0026quot;Gene expression of kallikreins in breast cancer cell lines.\u0026quot; Anticancer Research 40.5 (2020): 2487-2495.\u003c/li\u003e\n\u003cli\u003eMichaelidou, Kleita, Alexandros Ardavanis, and Andreas Scorilas. \u0026quot;Clinical relevance of the deregulated kallikrein-related peptidase 8 mRNA expression in breast cancer: a novel independent indicator of disease-free survival.\u0026quot; Breast cancer research and treatment 152 (2015): 323-336.\u003c/li\u003e\n\u003cli\u003eSuhaimi, Shafinah Ahmad, Soon Choy Chan, and Rozita Rosli. \u0026quot;Matrix metallopeptidase 3 polymorphisms: emerging genetic markers in human breast cancer metastasis.\u0026quot; Journal of breast cancer 23.1 (2020): 1-9.\u003c/li\u003e\n\u003cli\u003eMelendez-Zajgla, Jorge, et al. \u0026quot;Tissue inhibitor of metalloproteinases-4. The road less traveled.\u0026quot; Molecular cancer 7 (2008): 1-11.\u003c/li\u003e\n\u003cli\u003eRicks-Santi, Luisel J., et al. \u0026quot;Characterization of GATA3 and Mammaglobin in breast tumors from African American Women.\u0026quot; Archives of microbiology \u0026amp; immunology 7.1 (2023): 18.\u003c/li\u003e\n\u003cli\u003eTheodorou, Vasiliki, et al. \u0026quot;GATA3 acts upstream of FOXA1 in mediating ESR1 binding by shaping enhancer accessibility.\u0026quot; Genome research 23.1 (2013): 12-22.\u003c/li\u003e\n\u003cli\u003eHuang, Peng, et al. \u0026quot;Chemotherapy-driven increases in the CDKN1A/PTN/PTPRZ1 axis promote chemoresistance by activating the NF-\u0026kappa;B pathway in breast cancer cells.\u0026quot; Cell Communication and Signaling 16 (2018): 1-12.\u003c/li\u003e\n\u003cli\u003eMishra, Sanjay, et al. \u0026quot;cPLA2 blockade attenuates S100A7-mediated breast tumorigenicity by inhibiting the immunosuppressive tumor microenvironment.\u0026quot; Journal of Experimental \u0026amp; Clinical Cancer Research 41.1 (2022): 54.\u003c/li\u003e\n\u003cli\u003eDoebar, Shusma C., et al. \u0026quot;Gene expression differences between ductal carcinoma in situ with and without progression to invasive breast cancer.\u0026quot; The American journal of pathology 187.7 (2017): 1648-1655.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6167948/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6167948/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSpatial genetic heterogeneity plays a critical role in tumor evolution and therapeutic resistance, yet traditional histopathological characterization remains challenging and time-consuming. Here, we present an explainable deep learning pipeline \u0026ldquo;XpressO\u0026rdquo; that predicts and visualizes gene expression directly from whole slide images (WSIs), providing spatial resolution of tumor transcriptomics. Using histopathological image features of WSIs of invasive breast cancer as data and associated bulk RNA sequencing data from The Cancer Genome Atlas (TCGA) as expression labels, our model forms complex associations between tissue phenotype and gene expression. By generating high-resolution expression maps, our approach reveals both spatial variation and predicted gene activity across tumor samples, capturing patterns that are often lost in bulk profiling. The interpretability framework further highlights histological regions that contribute to specific gene expression signals, bridging the gap between tumor histology and genetic heterogeneity. This method offers a promising tool for integrating imaging and transcriptomics, enabling data-driven biomarker discovery and advancing precision oncology through spatially-informed molecular profiling.\u003c/p\u003e","manuscriptTitle":"XpressO: An explanatory deep learning pipeline for the prediction and visualization of gene expression heterogeneity in breast tumors","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-05 10:23:16","doi":"10.21203/rs.3.rs-6167948/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"0ba4f808-486a-49f9-88b0-4443bb878ead","owner":[],"postedDate":"June 5th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":49475174,"name":"Biological sciences/Cancer"},{"id":49475175,"name":"Biological sciences/Cancer/Breast cancer"},{"id":49475176,"name":"Health sciences/Biomarkers/Prognostic markers"},{"id":49475177,"name":"Biological sciences/Molecular biology/Transcriptomics"}],"tags":[],"updatedAt":"2026-01-14T20:08:44+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-05 10:23:16","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6167948","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6167948","identity":"rs-6167948","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}