Synthesis: ML for TNBC subtype classification from omics

Evidence grades (GRADE-adapted): A high — multiple well-conducted RCTs or systematic reviews converge. B moderate — single pivotal RCT or consistent observational evidence. C limited — single observational study, mechanistic, or expert consensus. D preclinical / hypothesis-generating.

The classification problem

Molecular subtype classification of TNBC is a multiclass classification problem: given a tumor's molecular profile (transcriptome, methylome, mutation calls, copy-number, or multi-omic combination), assign it to one of several predefined subtypes. The exact taxonomy depends on the framework — intrinsic subtypes (5 classes), Lehmann/Pietenpol (6 then 4 classes), Burstein (4 classes), FUSCC (4 classes). All approaches share the same supervised-learning structure: labeled training data, feature selection, model training, validation, deployment.

The PAM50 era (2009-2015)

The PAM50 classifier (Parker 2009) was the breakthrough ML application that brought intrinsic-subtype classification from research to clinical use^[1]A. The methodology was straightforward by modern standards:

• Feature selection: prediction-strength scoring on 1900+ genes reduced to 50
• Classifier: centroid-based nearest-centroid classification (a simple supervised method)
• Validation: cross-cohort testing across multiple labeled training sets
• Deployment: NanoString-based readout from FFPE tissue, FDA cleared as Prosigna (2013)

PAM50 succeeded clinically for several reasons: (1) the intrinsic-subtype labels had strong biological coherence rooted in the Perou/Sørlie foundational work, (2) the 50-gene panel was practical for FFPE-compatible assays, (3) the centroid classifier was simple enough to validate across cohorts, and (4) the clinical-utility case (risk-of-recurrence in HR+ disease) was well-defined. In TNBC, PAM50 has been more research-relevant than clinical because ~80% of TNBC is basal-like and the subtype doesn't currently change treatment.

TNBC-specific classifier development

Lehmann/Pietenpol 2011 used k-means clustering on aggregated gene expression data from 587 TNBC tumors, producing the original six-subtype TNBCtype framework^[2]A. The web tool implementation of the classifier remains in use as a research instrument.

Burstein 2015 used integrated multi-omic clustering (gene expression + methylation + copy-number + mutation) on 198 TNBC tumors, producing the BLIA/BLIS/MES/LAR four-subtype framework^[3]B. The methodology was non-standard in the ML sense (integrative analysis with bespoke clustering) but produced a clinically coherent taxonomy.

Both frameworks have inspired derivative classifiers attempting to operationalize the subtypes for clinical or trial use. Inter-classifier concordance on individual patient samples is typically 60–75%.

Deep learning approaches

Deep learning entered TNBC subtype classification in the late 2010s. Several approaches:

• Autoencoders for representation learning. Unsupervised deep models that learn compressed representations of tumor expression profiles. Variational autoencoders and similar architectures have been applied to TNBC transcriptomic data with the goal of identifying latent biology that explicit features may miss. Several papers have demonstrated improved cluster stability and identification of finer subtypes.
• Supervised deep classifiers. Convolutional or fully connected neural networks trained on labeled subtype data. In the bulk-RNA-seq setting, simple supervised models (random forest, support vector machine, gradient boosting) often perform comparably to deep models because the feature dimension is relatively small (~20,000 genes) and the training set sizes are small (hundreds of tumors). Deep approaches show advantages mostly when augmented with additional data modalities or transfer learning from large pretraining datasets.
• Transfer learning from pan-cancer atlases. Models pre-trained on TCGA pan-cancer data and fine-tuned on TNBC-specific data leverage broader transcriptomic information; this approach has shown modest improvements over TNBC-only training in cross-cohort generalization.

Single-cell ML approaches

Single-cell RNA-seq data has driven new ML approaches:

• Cell-type identification. Methods such as scVI, scanpy, Seurat with reference-based annotation, and CellTypist classify individual cells into tumor cells, immune cell types, stromal cells, etc. In TNBC, this enables disentangling tumor-intrinsic biology from microenvironment composition that confounded bulk frameworks.
• Subtype mapping at single-cell resolution. Cells within a tumor can be assigned to multiple bulk subtype categories; intratumoral heterogeneity is quantified.
• Trajectory inference. Tools such as Monocle, Slingshot, and PAGA identify developmental or evolutionary trajectories from single-cell data, with applications to characterizing tumor evolution and stem-like-to-differentiated transitions.

See the single-cell heterogeneity synthesis for biology findings; this page focuses on the ML methods.

Multi-omic integration

The Burstein framework's biological success motivated continued multi-omic integration research. ML approaches include:

• Canonical correlation analysis (CCA) and related methods that find correlated patterns across data modalities
• Multi-omic factor analysis (MOFA) that decomposes multi-modal data into shared and modality-specific factors
• Deep multi-modal architectures with separate input encoders per modality and a shared latent representation
• Graph neural networks that model relationships between features (genes, pathways) explicitly

In TNBC, multi-omic ML approaches have proposed refined subtype taxonomies and identified candidate biomarkers. Clinical translation has been limited by sample-size constraints (multi-omic profiling of large TNBC cohorts is expensive) and validation challenges (multi-omic cohorts large enough to support held-out validation are rare).

The clinical-deployment gap

Despite over 15 years of ML-based subtype classifier development, clinical deployment of subtype calls in TNBC remains limited to research use. Reasons:

• Lack of actionable subtype-specific therapy. The LAR / AR antagonist combination is the closest to a clinically validated subtype-targeted therapy, but phase III confirmation hasn't occurred (see LAR synthesis).
• Regulatory validation requirements. A clinical-grade subtype classifier requires validation across slide preparation methods, RNA-extraction protocols, sequencing platforms, and patient populations. This validation infrastructure is substantial.
• Cost-effectiveness uncertainty. A subtype classifier that doesn't change treatment is unlikely to be reimbursed by insurance.
• Inter-framework discordance. The 60–75% concordance between Lehmann and Burstein systems on the same patient cohort is problematic for clinical-decision support — which classification is "correct"?

Evidence table — key ML subtype classifier publications

Open questions and active investigation

• Foundation models for transcriptomic profiles. Large pre-trained models for transcriptomic data, analogous to language foundation models, are being developed. Whether they will produce better TNBC subtype calls than dedicated classifiers is an open question.
• Targeted-panel TNBC classifiers. The PAM50 example shows that small targeted panels can be clinical-grade. A TNBC-specific 50-100 gene panel with NanoString or similar readout could be developed if a clinical-utility case were established.
• H&E-based subtype prediction. Deep learning on H&E images can predict molecular subtype at moderate accuracy (see ai-pathology synthesis). If H&E-based subtype calls reach clinical accuracy, the cost-and-workflow advantages over transcriptomic profiling could change adoption dynamics.
• Subtype-stability under treatment. Whether subtype calls remain stable through chemotherapy + IO treatment is unclear. Serial-biopsy studies with subtype-call comparison would inform whether subtypes can drive sequential treatment decisions.
• Inter-framework harmonization. A consensus integrative classifier that incorporates Lehmann, Burstein, and FUSCC frameworks could resolve concordance issues and enable cleaner clinical-utility studies.

For the underlying subtype frameworks that ML classifiers operationalize, see the intrinsic subtypes synthesis, the Lehmann/Pietenpol synthesis, and the Burstein synthesis. For complementary computational pathology approaches, see the AI pathology synthesis.

References

Each citation links to the original publication via DOI. The same records are searchable in the evidence library by title or DOI.

1. Parker JS, Mullins M, Cheang MCU, et al. Supervised risk predictor of breast cancer based on intrinsic subtypes (PAM50). J Clin Oncol. 2009;27(8):1160–1167. doi:10.1200/JCO.2008.18.1370. ↩
2. Lehmann BD, Bauer JA, Chen X, et al. Identification of human triple-negative breast cancer subtypes and preclinical models for selection of targeted therapies. J Clin Invest. 2011;121(7):2750–2767. doi:10.1172/JCI45014. ↩
3. Burstein MD, Tsimelzon A, Poage GM, et al. Comprehensive genomic analysis identifies novel subtypes and targets of triple-negative breast cancer. Clin Cancer Res. 2015;21(7):1688–1698. doi:10.1158/1078-0432.CCR-14-0432. ↩

Last reviewed: 2026-06-04. Researcher-layer synthesis page. Evidence grades follow the GRADE-adapted rubric defined at the top of this page.