Transcriptomic Classifier for Predicting Neoadjuvant Immunotherapy Response in Triple-Negative Breast Cancer

Presenting their work at the inaugural European Society for Medical Oncology (ESMO) Artificial Intelligence (AI) & Digital Oncology Congress (Abstract 48P), David Kvaratskhelia, MD, and Tamar Melkadze, MD, of Todua Clinic, Tbilisi, Georgia, reported in their poster that integrating baseline tumor immune gene expression signatures with machine learning yields an accurate and interpretable prediction of pathologic complete response to neoadjuvant immunotherapy in patients with triple-negative breast cancer.

As the investigators explained, with few actionable genomic targets, triple-negative breast cancer carries a high risk of recurrence. They noted that, although neoadjuvant immune checkpoint inhibition has improved pathologic complete response rates, reliable biomarkers of benefit remain limited, but that recent advances in AI and transcriptomics have enabled the present study to explore new avenues for individualized patient stratification and precision oncology.

“[Our] findings support clinical integration of transcriptomic biomarkers and AI for patient stratification,” the investigators commented.

Methods

The investigators analyzed the publicly available The Cancer Genome Atlas Breast Invasive Carcinoma (TCGA-BRCA), Molecular Taxonomy of Breast Cancer International Consortium (METABRIC), and the National Center for Biotechnology Information’s Gene Expression Omnibus (GEO; GSE25066, GSE76124) datasets for data from patients with triple-negative breast cancer, focusing on pretreatment samples to profile baseline immune activity. Microarray and RNA sequencing data were normalized (transcripts per million [TPM]/counts per million [CPM]), log₂-transformed, and batch-corrected, and low-variance genes (< 0.1 TPM²) were filtered.

Immune-related transcripts from the tumor inflammation signature and interferon-γ (IFN-γ) pathways (~100 genes) defined the feature set. Surrogate pathologic complete response labels were generated using validated immune-activity gene expression scores and 5-year distant relapse–free survival.

Machine-learning models, including XGBoost (learning rate: 0.05; depth: 4; 500 boosting iterations) and L1 logistic regression (C = 0.2), were trained and evaluated using five-fold stratified cross-validation and independent test sets. Model performance metrics included the area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), sensitivity, specificity, decision-curve analysis, and calibration using Platt or isotonic scaling, and SHapley Additive exPlanations (SHAP) interpretability values were used to assess feature importance. Kaplan-Meier and Cox analyses were also applied to assess survival.

Key Findings

According to the investigators, the classifier demonstrated reproducible performance across independent cohorts. The XGBoost model achieved an AUROC, AUPRC, and accuracy of 0.83 (95% confidence interval [CI] = 0.78–0.87), 0.79 (95% CI = 0.74–0.84), and 0.78 (± 0.03), respectively, outperforming L1 logistic regression and random-forest baselines. At the optimal cutoff, the sensitivity was 76%, and the specificity was 80%.

Based on SHAP analysis, CXCL13 (B/T-cell chemokine), CD8A (cytotoxic T-cell marker), GZMB (granzyme B effector), IFN-γ, and antigen presentation genes (HLA-DRA, TAP1) were dominant predictors for response. This indicates a T-cell–inflamed phenotype, the investigators noted.

Relapse-free survival appeared to be significantly prolonged in predicted responders vs nonresponders (51 vs 31 months; hazard ratio = 0.46, 95% CI = 0.29–0.74; P = .001).

Receiver operating characteristic and precision–recall curves confirmed robust discrimination and calibration, the investigators reported.

Insights and Opportunities

The investigators concluded, “Baseline immune gene signatures achieved reproducible discrimination and predicted improved survival across public triple-negative breast cancer cohorts. SHAP-based interpretability enhanced clinical trust by revealing gene-level determinants of therapeutic response. Validated across TCGA, METABRIC, and GEO datasets, the model provides a transparent, reproducible framework for future trial-grade integration in triple-negative breast cancer immunotherapy.”

Nevertheless, they identified the need for prospective, trial-grade validation as a limitation. Other limitations included the use of retrospective public datasets, with surrogate pathologic complete response labels lacking uniform pathology, and the potential for residual platform effects even after correction.

In practice, the model may help triage patients most likely to achieve a pathologic complete response, informing treatment decisions and trial design. Future work will focus on prospective validation using both pathologic complete response and relapse-free survival endpoints, integrating spatial/immunohistochemistry and radiomic features, and deploying a calibrated web-based calculator and automated reporting tool.

DISCLOSURE: The study authors reported no conflicts of interest.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.