Feature Dimensionality Outweighs Model Complexity in Breast Cancer Subtype Classification Using TCGA-BRCA Gene Expression Data
arXiv:2605.06562v1 [cs.LG] 7 May 2026
Meena Al Hasani Independent Researcher
Abstract
biological settings remains unclear. Additionally, traditional evaluation metrics such as accuracy may obscure poor performance on minority subtypes. Metrics such as macro F1 score, which equally weights all classes, provide a more informative assessment in imbalanced datasets. This study investigates:
Accurate classification of breast cancer subtypes from gene expression data is critical for diagnosis and treatment selection. However, such datasets are characterized by high dimensionality and limited sample size, posing challenges for machine learning models. In this study, we evaluate the impact of model complexity and feature selection on subtype classification performance using TCGA-BRCA gene expression data. Logistic regression, random forest, and support vector machine (SVM) models were trained using varying numbers of highly variable genes (50 to 20,518). Performance was evaluated using stratified 5-fold cross-validation and assessed with accuracy and macro F1 score. While all models achieved high accuracy, macro F1 analysis revealed substantial differences in subtype-level performance. Logistic regression demonstrated the most stable and balanced performance across subtypes, including improved detection of rare classes. Random forest underperformed on minority subtypes despite strong overall accuracy, while SVM showed sensitivity to feature dimensionality. These findings highlight the importance of model simplicity, evaluation metrics, and feature selection in high-dimensional biological classification tasks.
1
1. Whether increased model complexity improves classification performance 2. How feature selection improves model behavior 3. How evaluation metrics influence model interpretation 4. Subtype-specific performance differences across models
2
Related Work
Machine learning has been increasingly applied to breast cancer subtype classification using gene expression data, building on early work demonstrating the feasibility of cancer classification from gene expression profiles [5, 14]. More broadly, machine learning has become a central tool in genomics for analyzing high-dimensional biological data [8]. Yu et al. [17] applied multiple machine learning models to RNA-seq data from TCGA to classify breast cancer into the five intrinsic molecular subtypes, demonstrating the feasibility of automated subtype identification from high-dimensional genomic profiles. Similarly, Wu and Hicks [16] evaluated support vector machines, K-nearest neighbor, Naive Bayes, and decision tree classifiers on TCGA gene expression data, finding that SVM outperformed competing models. However, both studies primarily rely on accuracy as the evaluation metric, which may obscure poor performance on minority subtypes in imbalanced datasets. Deep learning approaches have also been explored for this problem, reflecting broader trends in applying deep learning to healthcare and biomedical data [4]. Chhikara et al. [2] proposed a two-stage framework combining an autoencoder for dimensionality reduction with a deep neural network classifier, reducing the feature space from over 20,000 genes to 500 and achieving strong classification accuracy on TCGA data. While this work demonstrates the value of dimensionality reduction as a preprocessing step, it does not systematically evaluate how different levels of feature reduction affect model behavior across multiple classifier types, nor does it report subtype-level performance metrics that reveal minority class failure modes. Feature selection is widely recognized as a critical step in
Introduction
Breast cancer is a heterogeneous disease composed of multiple molecular subtypes, including Luminal A, Luminal B, HER2-enriched, Basal-like, and Normal-like [13]. These subtypes differ in prognosis and therapeutic response, making accurate classification essential for clinical decision-making. Standardized subtype classification was further enabled by the development of the PAM50 gene signature, a 50-gene panel that established a reproducible framework for molecular subtyping in clinical settings [9]. Gene expression profiling enables subtype classification but presents unique challenges due to the high dimensionality of genomic data relative to sample size. This imbalance increases the risk of overfitting, particularly for complex machine learning models. While advanced models such as random forest and support vector machines are often assumed to outperform simpler models, their effectiveness in small-sample, high-dimensional 1
machine learning workflows for high-dimensional omics data. Sanz et al. [11] demonstrated that removing redundant and irrelevant genes improves model generalization and reduces overfitting in high-dimensional biological classification tasks. Despite this, existing studies on breast cancer subtype classification rarely investigate the relationship between feature dimensionality and classifier performance in a systematic and comparative manner. A consistent limitation across prior work is the reliance on aggregate metrics such as accuracy, which obscures subtypespecific performance differences. In datasets with significant class imbalance such as TCGA-BRCA, where Luminal A samples constitute over half the cohort, high accuracy can be achieved by models that effectively ignore minority subtypes. This study addresses these gaps by systematically evaluating three classifiers of increasing complexity across five feature dimensionality levels, using macro F1 score as the primary evaluation metric and reporting per-subtype performance to expose failure modes invisible to accuracy-based evaluation.
3
metrics provide complementary perspectives on model performance, particularly in the presence of class imbalance. Accuracy measures the proportion of correctly classified samples. While it provides a general measure of performance, it can be misleading in imbalanced datasets where dominant classes disproportionately influence the score, a well-documented issue in machine learning [7]. Macro F1 score computes the F1 score independently for each class and then averages across all classes, assigning equal weight to each subtype. This makes it more suitable for evaluating performance in imbalanced datasets. Prior work has shown that metric choice significantly affects the interpretation of classification performance, particularly in imbalanced settings [12]. This metric is therefore used as the primary evaluation metric in this study. Weighted F1 score accounts for class imbalance by weighting each class according to its number of samples. While it provides a compromise between accuracy and macro F1, it still tends to favor dominant classes. Per-subtype F1 scores were analyzed to assess model performance on individual breast cancer subtypes. This enables identification of subtype-specific weaknesses that may not be apparent from aggregate metrics.
Dataset
Gene expression data and subtype annotations were obtained from the TCGA-BRCA dataset [15]. After aligning expression and clinical data by patient ID and removing missing subtype labels, a total of 981 samples were retained across five subtypes, as shown in Table 1. Gene expression features were derived from RNA-seq data, where each feature represents the expression level of a gene, resulting in 20,518 features per sample. The dataset presents several challenges for machine learning classification. It is high-dimensional relative to the number of samples, increasing the risk of overfitting. Additionally, the subtype distribution is imbalanced, with BRCA LumA comprising the majority of samples and BRCA Normal representing a small minority. These characteristics motivate the use of robust evaluation metrics and cross-validation strategies.
4.1
Accuracy
Accuracy measures the overall proportion of correctly classified samples across all subtypes. In imbalanced datasets such as TCGA-BRCA, accuracy is highly influenced by dominant classes such as BRCA LumA. As a result, models can achieve high accuracy even when performing poorly on minority subtypes. Accuracy =
4.2
TP + TN TP + TN + FP + FN
(1)
Macro F1-Score
Table 1: Detailed distribution of subtypes from the TCGA- Macro F1 score evaluates model performance by computing the BRCA dataset. F1 score for each subtype independently and averaging across all classes. This ensures that each class contributes equally, Subtype Sample Count regardless of its frequency. As a result, Macro F1 provides a more reliable measure of performance in imbalanced datasets BRCA Basal 171 where minority subtypes are clinically important. BRCA Her2 78 BRCA LumA BRCA LumB BRCA Normal
499 197 36
Total
981
C
Macro F1 =
4.3
4
Evaluation Metrics
1 X 2 · P c · Rc C c=1 Pc + Rc
(2)
Weighted F1-Score
Weighted F1 score computes the average F1 score across all classes, weighted by the number of samples in each class. This Model performance was evaluated using accuracy, macro F1 provides a balance between overall accuracy and class-level score, weighted F1 score, and per-subtype F1 scores. These performance but still biases results toward dominant subtypes. 2
4.4
Per-Subtype F1-Score
5.3
Accuracy increased with the number of genes for logistic regression and random forest, with both models achieving peak performance at 1000–20,000 genes (∼0.85–0.86). In contrast, SVM performance peaked at intermediate feature sizes (around 1000 genes) and declined at higher dimensionality, indicating sensitivity to large feature spaces.
Per-subtype F1 scores directly evaluate performance on each individual breast cancer subtype. This enables detailed analysis of model behavior across classes and highlights weaknesses in minority subtype classification.
4.5
Performance Across Feature Sizes
Metric Selection
Given the strong imbalance in TCGA-BRCA, macro F1 score 5.4 Macro F1 Reveals Model Differences was selected as the primary evaluation metric. Accuracy and weighted F1 were used as supporting metrics, while per-subtype Macro F1 analysis revealed substantial differences not captured F1 scores provide additional insight into model behavior across by accuracy alone. Logistic regression consistently achieved the highest macro F1 scores (∼0.79–0.80), indicating balanced individual subtypes. performance across subtypes. Random forest underperformed (∼0.69–0.70), suggesting bias toward dominant classes. SVM showed competitive performance at moderate feature sizes but 5 Baseline Evaluation degraded at high dimensionality. A majority class baseline was used to provide a reference point for model performance. This baseline predicts all samples as BRCA LumA, the dominant subtype in the dataset (499 out of 981 samples). While this approach yields relatively high accuracy, it fails to capture meaningful subtype distinctions and performs poorly under macro F1 evaluation. All models were evaluated using stratified 5-fold cross-validation to preserve subtype distribution, with mean and standard deviation reported across folds.
5.5
Subtype-Specific Performance
Subtype-level analysis revealed substantial variation in model behavior across classes. Logistic regression maintained consistent performance across all subtypes, including minority classes (BRCA Normal F1 ≈ 0.62). Random forest achieved strong performance on dominant subtypes but performed poorly on rare classes (BRCA Normal F1 ≈ 0.21). SVM showed improved performance on certain subtypes such as LumB and Normal but lacked overall stability. Notably, initial single train5.1 Majority Class Baseline test split results suggested that logistic regression failed on the BRCA Normal subtype. However, cross-validation revealed The majority class baseline highlights the limitations of accuthat this was an artifact of data splitting, with true performance racy as a standalone metric in imbalanced datasets. Although it substantially higher. This highlights the importance of robust achieves an accuracy of 0.509, its macro F1 score is near zero, evaluation strategies in high-dimensional, imbalanced datasets. reflecting complete failure to classify minority subtypes.
5.2
6
Evaluation and Results
Main Approach
As shown in Table 2, all machine learning models substantially outperform the majority class baseline. Logistic regression achieves the highest overall performance, with both strong accuracy and macro F1 score, indicating balanced classification across subtypes. Random forest achieves comparable accuracy but lower macro F1, suggesting bias toward dominant classes. SVM demonstrates lower overall performance, particularly under macro F1 evaluation.
This study evaluates the interaction between feature dimensionality and model complexity in breast cancer subtype classification using gene expression data. The workflow consists of preprocessing, variance-based feature selection, model training, and cross-validated evaluation across multiple feature sizes. Three models of increasing complexity were compared: logistic regression (linear), random forest (ensemble), and support vector machine with an RBF kernel (nonlinear). This design enables direct comparison of how model complexity Table 2: Baseline and model performance at full feature set interacts with high-dimensional feature spaces. (20,518 genes). Model Majority Class Baseline Logistic Regression Random Forest SVM
Accuracy
Macro F1
0.509 0.861 0.850 0.691
0.140 0.795 0.689 0.587
6.1
Preprocessing
Samples with missing subtype labels were removed. Gene expression features were standardized using z-score normalization for models sensitive to feature scaling (logistic regression and SVM), while no scaling was applied to random forest. 3
6.2
Feature Selection
6.3.1
Logistic Regression
subsampling [1]. It is capable of capturing nonlinear relationships and interactions between genes. A total of 500 estimators To address the high dimensionality of gene expression data, were used with balanced class weights. However, in imbalfeatures were selected based on variance within the training anced datasets, it may preferentially model dominant classes, data. Variance-based feature selection is a widely used heuris- leading to reduced performance on minority subtypes. tic in genomic studies, operating under the assumption that genes exhibiting greater expression variability across samples are more likely to carry subtype-discriminative signal [11, 6]. Feature selection is particularly important in bioinformatics applications, where datasets often contain thousands of genes but relatively few samples [10]. While this approach does not guarantee selection of the most biologically meaningful fea6.3.3 Support Vector Machine tures, it provides a computationally efficient and parameter-free method for dimensionality reduction that avoids data leakage when applied within each training fold. For each experiment, genes were ranked by variance A support vector machine with a radial basis function (RBF) and the top N genes were selected, where N ∈ kernel was used to model nonlinear decision boundaries [3] (C {50, 75, 100, 1000, 20518}. This range allows for systematic = 1.0, gamma = scale, default). The RBF kernel enables the evaluation of model performance under progressively reduced model to capture complex relationships between genes. Howfeature spaces, from aggressive dimensionality reduction to the ever, SVMs are sensitive to high-dimensional feature spaces full gene set. Importantly, feature selection was performed in- and limited sample sizes, which can lead to reduced stability dependently within each training fold during cross-validation to and performance when using very large feature sets. prevent data leakage and preserve the integrity of performance estimates. To assess the biological relevance of the selected features, genes consistently selected across all five cross-validation folds at the 1,000-gene threshold were examined. Of the 885 stably selected genes, several are established breast cancer subtype markers, including ERBB2, which defines the HER2-enriched 6.4 Cross-Validation Strategy subtype, KRT8, KRT14, and KRT19, which distinguish luminal from basal subtypes, and SCGB2A2 and XBP1, which are associated with luminal breast cancer. This suggests that variance- Model performance was evaluated using stratified 5-fold crossbased selection captures genuine subtype-discriminative signal validation to preserve subtype distribution. Within each fold, rather than arbitrary biological or technical variation. How- feature selection was performed using only the training data ever, this analysis serves as a sanity check rather than a formal before applying the model to the validation fold. This ensures biomarker discovery, and the presence of known markers does that performance estimates reflect true generalization rather not preclude the inclusion of non-subtype-related signal in the than benefiting from information leakage. To assess whether selected feature set. observed performance differences between models were statistically significant, pairwise Wilcoxon signed-rank tests were applied to per-fold macro F1 scores at the 1,000-gene feature 6.3 Models size. Three models of increasing complexity were evaluated to assess how model capacity interacts with feature dimensionality in high-dimensional biological data.
Logistic regression serves as a linear baseline model. It assumes 6.5 Performance Summary a linear decision boundary in feature space and is well-suited for high-dimensional datasets due to its use of L2 regularization (C = 1.0, default), which helps prevent overfitting. Class weights Table 3 summarizes cross-validated performance across feawere balanced to account for subtype imbalance. ture sizes. Logistic regression consistently achieves the highest macro F1 scores, indicating balanced performance across subtypes. Random forest attains comparable accuracy but lower 6.3.2 Random Forest macro F1, reflecting bias toward dominant classes. SVM perRandom forest is an ensemble learning method that constructs forms competitively at intermediate feature sizes but degrades multiple decision trees using bootstrap sampling and feature at high dimensionality. 4
Table 3: Cross-validated model performance across feature subset of highly variable genes. However, performance gains sizes. plateaued beyond approximately 1000 genes, indicating diminishing returns from additional features. This trend is consistent Genes Model Acc Macro F1 Wtd F1 with Figures 1 and 2, where performance stabilizes beyond approximately 1000 genes. Notably, SVM performance de20518 LR 0.861 0.795 0.860 clined at higher dimensionality, suggesting sensitivity to highRF 0.850 0.689 0.831 dimensional feature spaces and limited sample size. SVM 0.691 0.587 0.693
7
1000
LR RF SVM
0.838 0.833 0.804
0.789 0.700 0.736
0.838 0.812 0.806
100
LR RF SVM
0.777 0.808 0.766
0.716 0.662 0.677
0.782 0.784 0.768
75
LR RF SVM
0.775 0.811 0.769
0.720 0.657 0.678
0.781 0.787 0.771
50
LR RF SVM
0.721 0.790 0.740
0.656 0.608 0.658
0.734 0.758 0.743
Discussion
The results demonstrate that increasing model complexity does not necessarily improve performance in high-dimensional, small-sample biological datasets. Logistic regression consis- Figure 1: Model accuracy versus number of genes used for tently achieved the strongest overall performance in terms of feature selection. Error bars indicate standard deviation across macro F1 score, indicating balanced classification across sub- 5-fold cross-validation. types. Random forest achieved comparable accuracy but lower macro F1, suggesting that its performance was driven by dominant classes. SVM demonstrated competitive performance at intermediate feature sizes but showed reduced stability at higher dimensionality. These findings highlight the importance of jointly evaluating feature dimensionality, model complexity, and evaluation metrics rather than optimizing for accuracy alone. Pairwise Wilcoxon signed-rank tests were conducted to assess the statistical significance of performance differences between models. Logistic regression achieved higher macro F1 scores than both random forest and SVM across crossvalidation folds, however neither comparison reached statistical significance at the 0.05 threshold (LR vs RF: p = 0.0625, LR vs SVM: p = 0.0625, RF vs SVM: p = 1.000). This is attributable to the limited statistical power of 5-fold crossvalidation, where the minimum achievable p-value is 0.0625 with five paired observations. The performance differences should therefore be interpreted as consistent practical trends rather than formally significant results.
7.1
Effect of Feature Reduction Across Models
Feature dimensionality had a significant impact on model performance. All models improved as the number of genes in- Figure 2: Macro F1 score versus number of genes. Unlike creased from 50 to 1000, suggesting that important subtype- accuracy, macro F1 equally weights all subtypes and exposes discriminative information is captured within a relatively small performance differences obscured by class imbalance. 5
Figure 4: Per-subtype F1 scores at 1,000 genes averaged across 5-fold cross-validation. Random forest achieves the lowest F1 on BRCA Normal (0.21) while logistic regression maintains the most consistent performance across all subtypes.
7.3
Model errors were primarily concentrated in minority and biologically overlapping subtypes. The BRCA Normal subtype, which has the smallest number of samples, exhibited the most variability in classification performance across models. Random forest consistently underperformed on this class, while logistic regression and SVM showed moderate improvement. The BRCA LumB subtype also demonstrated reduced performance compared to BRCA LumA, suggesting partial overlap in gene expression patterns between these subtypes. In contrast, BRCA Basal was consistently classified with high performance across all models, indicating strong separability in gene expression space. These results indicate that classification errors are driven by both class imbalance and intrinsic biological similarity between subtypes, reinforcing the need for evaluation metrics that capture subtype-level performance.
Figure 3: Per-subtype F1 scores across feature sizes for logistic regression, random forest, and SVM. Each line represents one breast cancer subtype. BRCA Basal achieves consistently high F1 across all models while BRCA Normal shows the greatest variability.
7.2
Error Analysis
Subtype-Specific Performance
7.4
Subtype-level analysis revealed that classification performance varies significantly across breast cancer subtypes. BRCA Basal and BRCA LumA were consistently classified with high F1 scores across all models, indicating strong separability in gene expression space. In contrast, BRCA LumB and BRCA Her2 showed moderate performance, suggesting that these subtypes may share overlapping gene expression patterns, making them more difficult to separate using standard machine learning models.
Model Recommendation
Based on the results of this study, logistic regression with L2 regularization trained on the top 1,000 highest-variance genes represented the strongest overall configuration evaluated in this study for breast cancer subtype classification using TCGABRCA RNA-seq data. This configuration consistently achieved the highest macro F1 score across all feature sizes, demonstrated the most balanced per-subtype performance including minority classes, and showed the greatest robustness to class The BRCA Normal subtype exhibited the greatest variabil- imbalance. Its computational efficiency and interpretability ity, reflecting both its small sample size and biological simi- further support its suitability for high-dimensional biological larity to other subtypes. Logistic regression achieved the most classification tasks. consistent performance across subtypes, while random forest Random forest is not recommended as a primary classifier struggled with minority classes. SVM demonstrated moderate in this setting. Despite achieving comparable overall accuimprovement for certain subtypes but lacked overall consis- racy, it systematically underperformed on minority subtypes, tency. particularly BRCA Normal, suggesting that its ensemble mech6
anism favors dominant classes in imbalanced datasets. SVM 10 Future Work demonstrated competitive performance at intermediate feature sizes but showed significant degradation at full dimensionality, Future work could explore the use of larger datasets or multimaking it sensitive to the choice of feature set and less reliable cohort integration to improve generalization. Incorporating additional data types, such as methylation or proteomics, may across varying experimental conditions. further enhance subtype classification performance. More advanced techniques for handling class imbalance, such as synthetic oversampling or cost-sensitive learning, could 8 Limitations improve performance on minority subtypes. Additionally, exploring alternative feature selection methods beyond variance, Several limitations of this study should be acknowledged. First, such as pathway-based approaches, may provide more biologivariance-based feature selection is a heuristic that does not cally interpretable results. distinguish between subtype-discriminative signal and other Finally, extending this framework to deep learning models sources of biological or technical variation, such as patient could further evaluate the trade-off between model complexity age, tumor stage, or batch effects. Genes selected by variance and data limitations. may capture non-subtype-related variation, and future work incorporating pathway-based or supervised feature selection methods may improve biological interpretability. Author Contributions Second, all models were evaluated using default hyperparameters without systematic tuning. A grid search or Bayesian op- The author conceived the study, implemented the models, pertimization may alter the relative performance between models, formed the data analysis, and wrote the manuscript. particularly for SVM where kernel parameters can substantially affect behavior in high-dimensional spaces. References Third, this study relies on a single dataset, TCGA-BRCA, and results may not generalize to other cohorts with different [1] Leo Breiman. Random forests. Machine Learning, 45(1):5–32, patient populations, sequencing protocols, or sample collection 2001. methods. External validation on an independent dataset such as METABRIC would strengthen the generalizability of these [2] Bhupender Singh Chhikara et al. Deep learning based model for breast cancer subtype classification. arXiv preprint findings. arXiv:2111.03923, 2021. Fourth, the BRCA Normal subtype is represented by only 36 samples, making performance estimates for this class less [3] Corinna Cortes and Vladimir Vapnik. Support-vector networks. Machine Learning, 20(3):273–297, 1995. reliable than for larger subtypes regardless of cross-validation strategy. [4] Andre Esteva et al. A guide to deep learning in healthcare. Finally, pairwise Wilcoxon signed-rank tests did not reach Nature Medicine, 25:24–29, 2019. statistical significance at the 0.05 threshold (LR vs RF: p = 0.0625, LR vs SVM: p = 0.0625), likely due to the limited [5] Todd R. Golub, Donna K. Slonim, Pablo Tamayo, Christine Huard, Michelle Gaasenbeek, Jill P. Mesirov, Hilary Coller, statistical power of 5-fold cross-validation. The observed perMignon L. Loh, James R. Downing, Mark A. Caligiuri, Clara D. formance differences should be interpreted as practically meanBloomfield, and Eric S. Lander. Molecular classification of caningful trends rather than formally validated conclusions. Future cer: class discovery and class prediction by gene expression work using repeated cross-validation or larger datasets may monitoring. Science, 286(5439):531–537, 1999. provide sufficient power for formal significance testing.
9
[6] Isabelle Guyon and Andre Elisseeff. An introduction to variable and feature selection. Journal of Machine Learning Research, 3:1157–1182, 2003.
Conclusion
[7] Haibo He and Edwardo A. Garcia. Learning from imbalanced data. IEEE Transactions on Knowledge and Data Engineering, This study demonstrates that, in gene expression-based breast 21(9):1263–1284, 2009. cancer subtype classification, simpler models such as logistic regression can outperform more complex models in terms of [8] Maxwell W. Libbrecht and William S. Noble. Machine learning balanced performance. The results show that accuracy alone is applications in genetics and genomics. Nature Reviews Genetics, insufficient for evaluation, and that macro F1 provides a more 16(6):321–332, 2015. reliable metric in the presence of class imbalance. Additionally, cross-validation is essential to avoid misleading conclusions, [9] Joel S Parker, Michael Mullins, Maggie CU Cheang, et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. particularly for minority subtypes, and variance-based feature Journal of Clinical Oncology, 27(8):1160–1167, 2009. selection improves model stability without sacrificing performance. These findings highlight practical considerations for [10] Yvan Saeys, Iñaki Inza, and Pedro Larrañaga. A review of applying machine learning to high-dimensional biomedical data feature selection techniques in bioinformatics. Bioinformatics, and emphasize the importance of robust evaluation strategies. 23(19):2507–2517, 2007.
7
[11] Hector Sanz, Clarissa Valim, Esteban Vegas, Josep M Oller, and Ferran Reverter. Sprefuged: A method to select a predictive function for a gene expression data based feature selection workflow. PLOS ONE, 12(12), 2017. [12] Marina Sokolova and Guy Lapalme. A systematic analysis of performance measures for classification tasks. Information Processing & Management, 45(4):427–437, 2009. [13] Therese Sorlie, Charles M Perou, Robert Tibshirani, et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proceedings of the National Academy of Sciences, 98(19):10869–10874, 2001. [14] Alexander Statnikov, Constantin F. Aliferis, Ioannis Tsamardinos, Douglas Hardin, and Shawn Levy. A comprehensive evaluation of multicategory classification methods for microarray gene expression cancer diagnosis. Bioinformatics, 21(5):631–643, 2005. [15] John N. Weinstein et al. The cancer genome atlas pan-cancer analysis project. Nature Genetics, 45(10):1113–1120, 2013. [16] Jiande Wu and Chindo Hicks. Breast cancer type classification using machine learning. Journal of Personalized Medicine, 11(2):61, 2021. [17] Zhezhou Yu, Zhuo Wang, Xiangchun Yu, and Zhe Zhang. Rnaseq-based breast cancer subtypes classification using machine learning approaches. Computational and Mathematical Methods in Medicine, 2020:4737969, 2020.
8