Triple-negative breast cancer (TNBC) is defined by the absence of three molecular targets: estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2). This combination makes TNBC the most aggressive and invasive breast cancer subtype, but also the hardest to treat, because none of the standard receptor-targeted therapies apply.
Without reliable molecular targets, oncologists must rely on chemotherapy, which is effective in some patients and not others. Accurately classifying TNBC into its biological subtypes could reveal which patients are likely to respond to specific treatment regimens, enabling true personalized oncology. However, no unified, validated classification method currently exists.
Traditional classification methods rely on clinical pathological features and limited molecular markers, but these approaches capture only a narrow slice of tumor biology. Any single data type -- whether gene expression alone or imaging alone -- misses critical information encoded at other biological levels.
This study hypothesizes that multi-omics data integration -- combining mRNA expression, miRNA expression, gene mutations, and DNA methylation -- will produce a more accurate and generalizable TNBC classification model than any single-omics approach, and tests this hypothesis with state-of-the-art deep learning architectures.
Molecular data were downloaded from the TCGA (The Cancer Genome Atlas) database, specifically the TCGA-BRCA dataset. This included copy number variation (CNV) data for 1,089 cases, mutation data for 977 cases, methylation data for 1,097 cases, miRNA expression for 1,078 cases, and mRNA expression for 1,093 cases. Samples were labeled as TNBC based on confirmed negativity for ER, PR, and HER2.
MRI imaging data came from two sources in the TCIA (The Cancer Imaging Archive): the Duke-Breast-Cancer-MRI dataset (874 usable samples, used for training and validation) and the TCGA-BRCA-MRI dataset (84 samples, used as an independent external test set). Tumor regions in all MRI images were manually annotated by two experienced radiologists.
For the multi-omics genomic data, 840 samples were retained after filtering for excessive missing data. The study compared four model types: single-omics machine learning models (one per data type), a deep learning model integrating all four omics layers, and an MRI radiomics deep learning model. This design allows head-to-head comparison of data breadth vs. model complexity.
All models were evaluated using cross-validation on internal data and then tested on an external dataset to assess transfer learning capability -- a critical test of whether a model generalizes beyond the data it was trained on.
Each single-omics model was built using LASSO regression for feature selection -- a method that shrinks less informative features toward zero, producing a sparse, interpretable gene set. Selected features were then fed into multivariate logistic regression classifiers. For mRNA data, 22 genes were selected; for miRNA, 12; for mutations, 5; and for DNA methylation, 20 CpG sites.
For MRI images, a two-step pipeline was used. First, a Mask R-CNN model (a deep convolutional network for object detection and pixel-level segmentation) located and outlined tumor regions in each MRI slice. MRI images were preprocessed using CLAHE (contrast-limited adaptive histogram equalization) to reduce noise and normalize intensity. Then, SE-ResNet101 -- a deep residual network augmented with Squeeze-and-Excitation attention modules -- classified the detected regions as TNBC or non-TNBC.
The SE module works by learning channel-specific importance weights: it compresses spatial feature maps into a global descriptor (Squeeze), then uses a small fully connected network to generate per-channel scale factors (Excitation), and applies these to reweight the feature map before classification. This mechanism allows the network to focus on the most informative feature channels for each input.
For the integrated multi-omics model, all 59 molecular markers (22 mRNA + 12 miRNA + 5 mutations + 20 methylation sites) were combined, and Bayesian hyperparameter optimization was used to automatically tune the neural network's architecture. The final network had two hidden layers with 7 nodes each, trained using scaled conjugate gradient backpropagation with cross-entropy loss. Neighborhood Component Analysis (NCA) was used for dimensionality reduction prior to training.
Single-omics models showed modest but meaningful performance. The mRNA model achieved an AUC of 0.892 on the training set and 0.731 on the validation set. The miRNA model reached 0.759 and 0.719 respectively. The mutation model was notably stronger, with AUCs of 0.978 (training) and 0.909 (validation). The DNA methylation model achieved 0.773 and 0.707. Each model thus captured a real but incomplete signal about TNBC status.
The integrated multi-omics deep learning model substantially outperformed all single-omics approaches: 98.0% accuracy on the training set, 97.0% on the validation set, and 91.0% on the external test set. AUC values exceeded 0.91 across all three partitions. The model classified 840 samples into TNBC and non-TNBC with high precision, recall, and F1 scores, confirming that combining all four molecular layers provides synergistic predictive power.
The MRI deep learning model achieved 89% accuracy on training data and 78% on the validation set, but dropped to 68% on transfer testing. AUC values followed a similar pattern (92%, 90%, 80%), confirming that the MRI model works well on the data it was trained on but does not transfer as cleanly to a different imaging dataset.
Analysis of misclassified MRI cases revealed a clear pattern: false positives tended to be large, irregularly shaped tumors, while false negatives were typically small tumors with shapes and edge diffusion characteristics that overlapped with normal tissue on vascular imaging. This insight points to specific imaging scenarios requiring improved feature extraction.
At the mRNA level, TNBC-positive samples showed 718 upregulated and 444 downregulated genes compared to non-TNBC samples. Gene ontology (GO) analysis showed enrichment in hormone regulation and immune cell-related pathways, while KEGG pathway analysis highlighted cytokine-related pathways -- biological signals consistent with TNBC's known inflammatory and immune-evasive characteristics.
The miRNA analysis identified significantly elevated levels of miR-21, miR-155, and miR-210 in TNBC-positive samples. These three miRNAs are established oncomiRs -- small non-coding RNAs that promote tumor growth, invasion, and immune evasion -- making their overexpression in TNBC biologically coherent.
The mutation analysis found higher mutation frequencies of key genes in TNBC-positive samples, notably TP53, PIK3CA, and CDH1, among others. TP53 mutations are particularly characteristic of TNBC, occurring in the vast majority of cases. The mutation model's high AUC (0.978 training, 0.909 validation) reflects this strong signal.
The DNA methylation model selected 20 methylation sites from 472 frequently methylated genes and achieved moderate performance (AUC 0.773/0.707). While methylation patterns contribute meaningful information to the integrated model, they appear to be less discriminative on their own than mutation or mRNA signatures for TNBC classification.
The study's central finding -- that multi-omics integration dramatically improves TNBC classification accuracy -- aligns with a fundamental principle of cancer biology: no single molecular layer fully captures tumor identity. Gene expression, epigenetic silencing, non-coding RNA regulation, and somatic mutations all contribute independently and interactively to a cancer's phenotype.
The MRI radiomics model's drop in performance during transfer testing (from 78% to 68% accuracy) highlights a known limitation of imaging-based AI: MRI features are sensitive to acquisition parameters, scanner hardware, and contrast agent protocols, which vary across institutions. Genomic data, being highly standardized at the measurement level, transfers more cleanly between datasets.
The authors note that DL models' inherent black-box nature is a clinical concern. While the models demonstrate impressive accuracy, clinicians and patients need to understand the reasoning behind predictions to exercise informed consent and to trust the recommendations. This represents the key barrier between high-performing research models and routine clinical deployment.
A practical strength of the multi-omics model is its potential integration with future biomarker discovery. As new molecular markers relevant to TNBC are identified, they could be incorporated into the feature set, and the Bayesian optimization framework would adapt the network structure accordingly. This extensibility makes the platform valuable beyond its current accuracy metrics.
This study establishes that multi-omics deep learning -- integrating mRNA, miRNA, mutation, and methylation data -- achieves superior and transferable TNBC subtype classification (91% external accuracy, AUC > 0.91) compared to any single molecular data type or MRI-based approach.
Accurate TNBC subtyping has direct clinical consequences: patients in different molecular subtypes respond differently to chemotherapy regimens, immunotherapy, and experimental targeted agents. A reliable subtyping model gives oncologists actionable information for treatment planning that is otherwise unavailable through standard pathology.
Key limitations include the use of retrospective public database data (which may not reflect all patient populations), the need for original image data sharing for full reproducibility, and the absence of an additional independent clinical validation cohort. The authors call for larger prospective multi-site studies to confirm generalizability.
Future directions include improving DL model interpretability through explainability tools (such as attention visualization and feature importance mapping), integrating other bioinformatics data types, and validating the models in prospective clinical settings where subtype predictions can be linked to actual treatment outcomes.