Machine Learning-Based Classification of Breast Cancer using TCGA Transcriptomics Data

Vasanta Putluri^1*,   Dr. Kaushal Kapadia²

1. Texila American University.

2 Clinical Research Professional.

*Correspondence to: Vasanta Putluri, Texila American University.

Copyright

© 2025 Vasanta Putluri. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Received: 06 February 2025

Published: 15 February 2025

DOI: https://doi.org/10.5281/zenodo.14881942

Abstract:

Breast cancer (BRCA) remains one of the most prevalent and challenging malignancies worldwide, necessitating advanced approaches for diagnosis and therapeutic intervention. In this study, we utilized transcriptomics data from The Cancer Genome Atlas (TCGA-BRCA) to classify tumor versus normal breast tissue samples using machine learning techniques. RNA-seq data normalized using RSEM was employed to assess differential gene expression profiles. The dataset underwent preprocessing steps, including removing genes with zero expression and log2 transformation of normalized values to ensure data quality and consistency.

Three machine learning algorithms-Logistic Regression (LR), Neural Networks (NN), and Random Forest (RF)-were implemented to classify tumor and normal samples. The dataset was randomly split into training (70%) and testing (30%) subsets for cross-validation over 100 iterations. Classifier performance was evaluated using the Area Under Curve-Receiver Operating Characteristic Curve (AUC-ROC). Logistic regression provided interpretable insights into the key features, while neural networks and random forests offered robust, nonlinear classification capabilities. To further refine the models, we identified a minimal set of informative features capable of discriminating tumors from normal samples with high accuracy. The results underscore the potential of machine learning in leveraging high-dimensional gene expression data for precise classification, offering a foundation for further exploration into biomarkers and personalized therapeutic strategies in breast cancer.

Keywords: Breast cancer, The Cancer Genome Atlas, RNA sequencing and Machine learning.

Abbreviations: Breast cancer (BRCA), RNA sequencing (RNA-seq), The Cancer Genome Atlas (TCGA), Machine Learning (ML), Logistic Regression (LR), Neural Networks (NN), Random Forest (RF), Receiver Operating Characteristic (ROC), Area Under the Receiver Operating Characteristic Curve (AUROC), Leiomodin-1(LMOD1), NEDD4 binding protein 2 like 1(N4BP2L1) and TPX2 Microtubule Nucleation Factor (TPX2).

Introduction 
Breast cancer (BRCA) is one of the most common malignancies affecting women globally and remains a leading cause of cancer-related mortality[1-6]. Despite advancements in diagnosis and treatment, the complexity and heterogeneity of breast cancer pose significant challenges for effective disease management[7]. Distinguishing tumor tissues from normal tissues at the molecular level is crucial for understanding the disease's underlying mechanisms and developing targeted therapies. High-throughput sequencing technologies, such as RNA sequencing (RNA-seq), have provided detailed gene information along with their expression, enabling researchers to explore biomarkers associated with breast cancer. The Cancer Genome Atlas (TCGA) offers comprehensive resources to study differential gene expression between tumor and normal breast tissues[8]. These datasets allow for identifying molecular signatures and pathways implicated in breast cancer development and progression.
Machine learning (ML) techniques have emerged as powerful tools for analyzing large and complex biological datasets in recent years [9-12]. Through traditional statistical methods, ML algorithms can identify patterns and relationships in data that are not apparent. By leveraging these techniques, it is possible to develop robust models for classifying breast cancer tumors and normal samples and to identify key features driving the classification. In this study, we utilized transcriptomic data from TCGA-BRCA to classify tumor and normal breast tissue samples using three machine learning methods: Logistic Regression (LR), Neural Networks (NN), and Random Forest (RF) [13]. The data underwent preprocessing steps, including removing low-quality features and log2 transformation, to ensure consistency and reliability in analysis. Our goal was to evaluate the performance of these ML methods in accurately predicting the model and apply the dataset to build acceptable models in tumor samples compared to normal samples, along with providing the most informative features for classification. Additionally, we focused on the relevant genes identified using machine learning methods and analyzed TCGA BRCA data to predict patient survival[8]. To the best of my knowledge, this is the first study that has attempted to predict biomarkers using a machine learning approach with transcriptomic data to understand breast cancer biology and pave the way for improved diagnostic and therapeutic strategies.

Materials and Methods
Data Preprocessing:
Transcriptomics data for breast cancer (BRCA) were sourced from The Cancer Genome Atlas (TCGA)[8]. This dataset included RNA sequencing (RNA-seq) profiles from both tumor and normal breast tissue samples,[8] downloaded in RSEM-normalized [14] format from the Broad GDAC Firehose repository. The cohort comprised both tumor and normal samples, ensuring a comprehensive representation of BRCA subtypes. To ensure data quality and enhance model performance, Genes with zero expression across all samples were excluded to reduce noise and eliminate non-informative features. This step helps in focusing the analysis on genes that are actively expressed and potentially relevant to the disease. The normalized RNA-seq data were log2-transformed to stabilize variance and standardize data distribution. This transformation makes the data more compatible with machine learning models by reducing the impact of extreme values and making the data more normally distributed. Statistical checks were conducted to verify that normalization and transformations were consistent across both tumor and normal groups. This step ensures that the data preprocessing does not introduce biases and that the data from different sample groups are comparable. The total number of 1212 specimens are 1100 breast cancer samples and 112 normal samples. 

Machine Learning Models:
The following machine learning algorithms such as LR, NN, and RF, were applied to classify tumor versus normal samples. A Logistic Regression model provides interpretable coefficients to assess the contribution of each feature. Regularization techniques (L1, L2) were applied to prevent overfitting. Neural Networks (NN) is a nonlinear, multi-layer perceptron (MLP) model that was implemented. Random Forest (RF) is an ensemble model constructed with 100 decision trees. Each tree was trained on a bootstrapped sample of the dataset, and feature importance was assessed using Gini impurity scores.

In the Logistic Regression (LR) Neural Network (NN), and Random Forest (RF) models the dataset was split into training (70%) and testing (30%) subsets to train and evaluate the models whereas random state was used at 21, 0, and 0, respectively. Stratified Splitting Ensured proportional representation of tumor and normal samples in both subsets. Cross-Validation: A repeated 6 split for LR, 10 split for both of NN, and RF were used for cross-validation strategies along with over 100 iterations to enhance model robustness and prevent bias. Model performance was evaluated using the following metrics are Area Under Curve (AUC-ROC). Assesses the model's ability to differentiate between tumor and normal samples. The metrices including accuracy, precision, recall, and F1-score provided additional insights into classification performance. In the Neural Networks (NN) model we use Cross-Validation. Iterative modeling was conducted with increasing numbers of features, starting from the top two ranked features to all informative features. Models were optimized at each iteration to balance performance and computational efficiency. Python (version 3.8) was utilized within the Anaconda Navigator environment via Jupyter Notebook, and the Scikit-learn package was employed to execute the machine learning models. Survival analysis of the association between individual genes and patient survival was evaluated using the log-rank test by comparing the bottom and top percentage of patients. Survival significance was assessed by employing the package survival in the R statistical system.

Results
The analysis included RNA-seq data for tumor and normal samples from the TCGA-BRCA dataset. After preprocessing, a total of [number of genes retained] genes were retained as features for machine learning analysis. Differential expression analysis identified [number of significant genes] significantly over- or under-expressed genes between tumor and normal samples (FDR < 0.05). Three machine learning algorithms—Logistic Regression (LR), Neural Networks (NN), and Random Forest (RF)—were applied to classify tumor versus normal samples. 
    
 
Figure 1: BRCA transcriptomics (RNA Seq) data use for the Machine Learning model. In this modal we have tumor and normal samples divide to training set and validation sets for the prediction model

The models were evaluated using the Area Under the Receiver Operating Characteristic Curve (AUROC) and additional metrics such as accuracy, precision, recall, and F1-score. Cross-Validation was performed using k-fold models that are subjected to 100 iterations of cross-validation with a training set (70%) and validation set (30%) (Figure 1). Logistic regression demonstrated consistent performance, while Neural Networks and Random Forests outperformed LR in terms of nonlinear classification capabilities. A minimal feature set of genes as identified as sufficient to classify tumor and normal samples with high accuracy. Models trained with the minimal feature set achieved comparable performance to those using the full dataset, underscoring the potential for simplified models in clinical applications.
All models had the highest AUROC and accuracy, demonstrating their strength in capturing complex relationships. Random Forest provided robust performance with interpretable feature importance rankings. Logistic regression offered the most interpretable results, making it suitable for applications requiring feature-level insights.
Receiver Operating Characteristic (ROC) curves for each model showed clear separation between tumor and normal samples, with Neural Networks exhibiting the largest area under the curve. Heatmaps and boxplots of key genes indicated distinct expression patterns between tumor and normal samples, validating their biological relevance. The most informative genes identified through machine learning corresponded to key pathways implicated in breast cancer.
 
Figure 2: Significantly associated biomarkers identified using machine learning models in tumor patients compared to healthy individuals in BRCA. A) The plot shows the area under the curve (AUC) values for the top 20 significantly associated genes in tumor patients compared to healthy individuals in BRCA. B) The heatmap represents the significantly associated gene expressions and their fold changes in tumor patients compared to healthy individuals in BRCA, as predicted through machine learning.

In this study, we employed machine learning models, including Logistic Regression (LR), Neural Networks (NN), and Random Forest (RF), to analyze RNA-seq data from The Cancer Genome Atlas (TCGA) breast cancer dataset. The analysis utilized a comprehensive dataset comprising 12,633 genes to identify potential biomarkers associated with breast cancer.

Using LR, we identified >2500 genes with an area under the curve (AUC) value exceeding 0.80. Similarly, the NN model identified >3000 genes, while the RF model identified >900 genes that met the same AUC threshold. These results demonstrate the variability in gene selection across different machine learning approaches, highlighting their unique contributions to biomarker discovery.

 
Figure 3: A) Low expression of LMOD1 was associated with poor survival in TCGA BRCA cohort (log-rank p= 0.01407; Top 50% and Bottom 50%).  B) Low expression of N4BP2L1 was associated with poor survival in TCGA BRCA cohort (log-rank p= 0.03864; Top 50% and Bottom 50%). C) High expression of TPX2 was associated with poor survival in TCGA BRCA cohort (log-rank p= 0.02409; Top 40% and Bottom 60%).

After combining the results from all three machine learning models, we refined our analysis to  focus on the most robust predictors. A total of 27 genes were identified as potential biomarkers, demonstrating consistent and significant associations with breast cancer. These genes were selected based on their high predictive performance and biological relevance, with AUC values consistently above 0.95 across the models (Figure 2). To provide a comprehensive view of the selected biomarkers, we generated a bar plot illustrating the AUC values of the 27 genes. This plot highlights the predictive power of each gene in distinguishing cancer patients from healthy individuals. Additionally, we created a heatmap to visualize the expression profiles of these 27 genes, along with their corresponding fold changes. The heatmap distinctly shows the differential expression patterns of these genes between tumor and healthy samples, further validating their potential as biomarkers (Figure 2).
The integration of multiple machine learning approaches allowed for robust identification of potential biomarkers for breast cancer. The 27 genes identified in this study represent promising candidates for further validation and functional studies, offering insights into breast cancer biology and potential targets for clinical applications.
We next examined the expression of these 27 genes using TCGA BRCA transcriptomics data and analyzed their association with survival using publicly available TCGA BRCA data. The log-rank test was used to assess patient outcomes. Interestingly, Leiomodin-1(LMOD1) and 
NEDD4 binding protein 2 like 1(N4BP2L1), which were downregulated in BRCA within the TCGA BRCA cohort, were associated with poor survival (Figure 3A-B). On the other hand, TPX2 microtubule nucleation factor (TPX2), which was upregulated in BRCA within the TCGA BRCA cohort, was also associated with poor survival (Figure 3C). 
All models demonstrated consistent performance across cross-validation. External validation using independent datasets confirmed the generalizability of the models. The results highlight the utility of machine learning in distinguishing tumor and normal breast tissue samples using transcriptomic data. Neural Networks and Random Forest emerged as robust classifiers, while Logistic Regression provided interpretable insights. This study underscores the potential of combining machine learning with high-dimensional gene expression data for biomarker discovery and precision medicine in breast cancer. The genes show survival differences may need to be evaluated mechanically in future study.

Discussion
This study demonstrates the potential of machine learning techniques in classifying tumors and normal samples in breast cancer using transcriptomic data from the TCGA-BRCA dataset. By employing Logistic Regression (LR), Neural Networks (NN), and Random Forests (RF), we effectively identified key genes that differentiate tumor samples from normal tissue, offering valuable insights into breast cancer biology and potential therapeutic targets.
The comparative analysis of LR, NN, and RF models highlighted their individual strengths and limitations. Neural Networks consistently outperformed other models in terms of AUROC and accuracy, showcasing their ability to capture nonlinear relationships in high-dimensional datasets. Random Forests, with their inherent feature importance analysis, provided interpretable insights while maintaining robust classification performance. Logistic regression, although simpler, offered high interpretability, making it a valuable tool for identifying the most influential genes in the dataset. The consistent performance across cross-validation and external validation underscores the reliability of these models. The ability to achieve high classification accuracy with a minimal feature set demonstrates the potential for developing simplified diagnostic tools that retain predictive power while reducing computational complexity.
This study demonstrates the effectiveness of integrating multiple machine learning models-LR, NN, and RF models to identify robust biomarkers for breast cancer using TCGA RNA-seq data. From an initial dataset of 12,633 genes, we predicted 27 key genes with consistently high predictive performance (AUC > 0.95) across three models. AUC-ROC of 27 genes were visualize through bar plot and, their gene expression was represented in heatmap that are showed significant differential expression between tumor and healthy samples (p-value > 0.05), underscoring their potential as diagnostic biomarkers.
The diverse gene selection across models highlights their complementary strengths, with NN excelling at capturing nonlinear relationships and RF emphasizing features with strong predictive signals. By integrating results, we achieved a refined set of reliable biomarkers, offering insights into breast cancer biology and potential therapeutic targets. While the use of TCGA data provides a strong foundation, validation in independent cohorts and functional studies is needed to confirm clinical applicability. Future work could explore multi-omics integration and predictive modeling for clinical outcomes, advancing precision oncology. These findings highlight the potential of machine learning in biomarker discovery, paving the way for improved breast cancer diagnostics and treatment strategies.
The identification of key genes, such as Matrix Metallopeptidase 11(MMP11)[15-17], TPX2[18], POC1 centriolar protein A (POC1A)[19], Meis homeobox 2(MEIS2)[20], and Serum deprivation-response protein (SDPR)[21], aligns with the established understanding of breast cancer pathophysiology. These genes are implicated in processes identified through the TCGA-BRCA datasets, where a high expression of MMP11 in clinical samples was strongly associated with poor prognosis in breast cancer (BRCA) patients. Furthermore, pathway enrichment analysis revealed that the TGF-β signaling pathway is a potential downstream target of MMP11.
TPX2 is frequently overexpressed in breast cancer tissues compared to normal tissues[22]. It plays a central role in breast cancer by promoting genomic instability, abnormal mitotic activity, and tumor progression. Overexpression of TPX2 is associated with aggressive cancer phenotypes and poor patient outcomes. Ongoing research highlights its potential as a biomarker for diagnosis and prognosis, as well as a therapeutic target to improve outcomes for breast cancer patients.
Dysregulated expression of POC1A has been linked to various cancers[23, 24],. Overexpression or under expression of POC1A may disrupt normal centrosome function, leading to abnormal cell division. POC1A is being explored as a biomarker for aggressive cancer types due to its role in promoting tumor progression.
The MEIS2 gene plays a multifaceted role in breast cancer[20], influencing tumor growth, metastasis, and therapy response through its regulation of gene expression and signaling pathways. MEIS2 may have a specific role in ER-positive breast cancer [20], where it regulates gene networks involved in hormone signaling.
Downregulation of SDPR in breast cancer [25] highlights its potential as a tumor suppressor and biomarker for disease progression. Loss of SDPR expression has been associated with increased tumor growth, invasion, and metastasis. SDPR also plays a role in key cellular processes, such as DNA repair, cell cycle regulation, and hormonal signaling, which are critical in tumorigenesis.
Additionally, pathway enrichment analysis revealed significant dysregulation in pathways associated with cell proliferation, immune evasion, and metabolic reprogramming. These findings highlight the complex interplay of molecular alterations driving breast cancer progression and suggest potential targets for therapeutic intervention.
The downregulated genes (LMOD1 and N4BP2L1) and upregulated gene (TPX2) were associated with poor survival in TCGA BRCA.  Further characterization and validation of these genes may provide a prognostic marker for BRCA. LMOD1 a Tumor-Suppressive Role in Breast Cancer has been shown to effect the by the JAK2/STAT3 Pathway[26] Patients with breast cancer show significantly downregulation of N4BP2L1 in smooth muscle cells of breast tumor tissues [27] Earlier studies showed that TPX2 is a microtubule-associated protein that is strongly correlated with chromosomal instability, resulting in the development of different human tumors [22]

The TCGA-BRCA dataset has some limitations. It may not fully represent the genetic diversity of breast cancer patients from different populations. Future studies should include more diverse datasets to make the findings more applicable. The dataset also has fewer normal samples, which could affect the model's accuracy. Increasing the number of normal samples would improve the results.
Adding more types of data, like proteomics, metabolomics, and epigenomics, could give a better understanding of breast cancer. Combining these with transcriptomics could help find new biomarkers and better treatment targets. Further characterization and functional validation of the genes could provide insights into the molecular events that are related to progression of BRCA.

Conclusion
In this study, we used machine learning techniques to classify tumor and normal breast tissue samples from the TCGA-BRCA dataset. By applying Logistic Regression, Neural Networks, and Random Forest models, we achieved robust classification performance, with Neural Networks showing the highest predictive accuracy. Key genes identified, such as MMP11, TPX2, POC1A, MEIS2, and SDPR, align with known breast cancer biology and highlight critical pathways like DNA repair, cell cycle regulation, and metabolic reprogramming. Our study of normal versus cancer BRCA provided numerous novel insights into disease biology and delineated pathways that provide potential opportunities for therapeutic intervention.  This study underscores the potential of machine learning to streamline diagnostic and prognostic tools in breast cancer research, offering new perspectives on data analysis and biomarker discovery. These results may have prognostic value and provide targets for therapeutic intervention in the future.  Moving forward, expanding the dataset, incorporating multi-omics approaches, and validating findings in independent cohorts will be essential for clinical translation.

References
1. Wilkinson, L. and T. Gathani, Understanding breast cancer as a global health concern. Br J Radiol, 2022. 95(1130): p. 20211033.
2. Bose, R., et al., Activating HER2 mutations in HER2 gene amplification negative breast cancer. Cancer Discov, 2013. 3(2): p. 224-37.
3. Meena, J.K., et al., MYC Induces Oncogenic Stress through RNA Decay and Ribonucleotide Catabolism in Breast Cancer. Cancer Discov, 2024. 14(9): p. 1699-1716.
4. Dasgupta, S., et al., Metabolic enzyme PFKFB4 activates transcriptional coactivator SRC-3 to drive breast cancer. Nature, 2018. 556(7700): p. 249-254.
5. Terunuma, A., et al., MYC-driven accumulation of 2-hydroxyglutarate is associated with breast cancer prognosis. J Clin Invest, 2014. 124(1): p. 398-412.
6. Xing, Z., et al., Expression of Long Noncoding RNA YIYA Promotes Glycolysis in Breast Cancer. Cancer Res, 2018. 78(16): p. 4524-4532.
7. Carlino, F., et al., Editorial: Heterogeneity in breast cancer: clinical and therapeutic implications. Front Oncol, 2024. 14: p. 1321654.
8. Cancer Genome Atlas Research, N., et al., The Cancer Genome Atlas Pan-Cancer analysis project. Nat Genet, 2013. 45(10): p. 1113-20.
9. Isaac, S., et al., Usability, Acceptability, and Implementation of Artificial Intelligence (AI) and Machine Learning (ML) Techniques in Surgical Coaching and Training: A Scoping Review. J Surg Educ, 2024. 81(7): p. 994-1003.
10. Shakeel, C.S. and S.J. Khan, Machine learning (ML) techniques as effective methods for evaluating hair and skin assessments: A systematic review. Proc Inst Mech Eng H, 2024. 238(2): p. 132-148.
11. Ghavidel, A. and P. Pazos, Machine learning (ML) techniques to predict breast cancer in imbalanced datasets: a systematic review. J Cancer Surviv, 2023.

12. Mastrodicasa, D., et al., Artificial intelligence machine learning-based coronary CT fractional flow reserve (CT-FFR(ML)): Impact of iterative and filtered back projection reconstruction techniques. J Cardiovasc Comput Tomogr, 2019. 13(6): p. 331-335.
13. Dou, B., et al., Machine Learning Methods for Small Data Challenges in Molecular Science. Chem Rev, 2023. 123(13): p. 8736-8780.
14. Thron, C. and F. Jafari, Correcting scale distortion in RNA sequencing data. BMC Bioinformatics, 2025. 26(1): p. 32.
15. Cheng, C.W., et al., The clinical implications of MMP-11 and CK-20 expression in human breast cancer. Clin Chim Acta, 2010. 411(3-4): p. 234-41.
16. Eiro, N., et al., MMP1 and MMP11 Expression in Peripheral Blood Mononuclear Cells upon Their Interaction with Breast Cancer Cells and Fibroblasts. Int J Mol Sci, 2020. 22(1).
17. Eiro, N., et al., MMP11 expression in intratumoral inflammatory cells in breast cancer. Histopathology, 2019. 75(6): p. 916-930.
18. Liu, X., et al., Integrative analysis based on the cell cycle-related genes identifies TPX2 as a novel prognostic biomarker associated with tumor immunity in breast cancer. Aging (Albany NY), 2024. 16(8): p. 7188-7216.
19. Zhao, Q., et al., POC1A, prognostic biomarker of immunosuppressive microenvironment in cancer. Aging (Albany NY), 2022. 14(12): p. 5195-5210.
20. Xiao, Y., et al., MEIS2 suppresses breast cancer development by downregulating IL10. Cancer Rep (Hoboken), 2024. 7(5): p. e2064.
21. Tian, Y., et al., Serum deprivation response inhibits breast cancer progression by blocking transforming growth factor-beta signaling. Cancer Sci, 2016. 107(3): p. 274-80.
22. Jiang, Y., et al., TPX2 as a Novel Prognostic Indicator and Promising Therapeutic Target in Triple-negative Breast Cancer. Clin Breast Cancer, 2019. 19(6): p. 450-455.
23.Chodary Khameneh, S., et al., Weighted correlation network analysis revealed novel long non-coding RNAs for colorectal cancer. Sci Rep, 2022. 12(1): p. 2990.
24. Sala, C., et al., An interaction network of inner centriole proteins organised by POC1A-POC1B heterodimer crosslinks ensures centriolar integrity. Nat Commun, 2024. 15(1): p. 9857.
25. Ozturk, S., et al., SDPR functions as a metastasis suppressor in breast cancer by promoting apoptosis. Proc Natl Acad Sci U S A, 2016. 113(3): p. 638-43.
26. Fang, X., et al., LMOD1 Exerts a Tumor-Suppressive Role in Breast Cancer by Restraining the JAK2/STAT3 Pathway. J Biochem Mol Toxicol, 2025. 39(1): p. e70092.
27. Dastsooz, H., et al., Involvement of N4BP2L1, PLEKHA4, and BEGAIN genes in breast cancer and muscle cell development. Front Cell Dev Biol, 2024. 12: p. 1295403.