JMIR Publications

ABSTRACT

Background:

Hormone receptor–positive (HR+) breast cancer exhibits limited and heterogeneous clinical benefit from immune checkpoint inhibitors. While peripheral blood single-cell profiling provides a minimally invasive approach to monitoring systemic immune dynamics, its utility in predicting treatment response remains to be fully established.

Objective:

This study aims to develop and evaluate a multimodal machine learning framework that integrates peripheral blood single-cell transcriptomes and T-cell receptor (TCR) encodings to predict response to chemo-immunotherapy in patients with early-stage HR+ breast cancer.

Methods:

I analyzed the GSE300475 cohort, comprising longitudinal samples from 4 patients (11 total samples; 100,067 cells). The feature set included principal components of gene expression, TCR k-mer and physicochemical encodings, and quality control covariates. I compared several classification algorithms, including logistic regression, tree-based baselines, and sequence-aware deep models, using leave-one-patient-out cross-validation for cell-level evaluation. Model interpretability was assessed via SHAP (SHapley Additive exPlanations) for tree models and gradient-based attributions for neural networks, with uncertainty quantified through nonparametric bootstrapping.

Results:

The multimodal models achieved high cell-level discrimination, with a peak area under the receiver operating characteristic curve of 0.97 and an accuracy of 91.7%. Transcriptomic signatures related to cytotoxicity and interferon response were the primary drivers of model predictions. The integration of TCR encodings provided complementary signals that improved model calibration. Sensitivity analyses confirmed the robustness of these findings to imputation and initialization variations, though the results remain exploratory due to the small cohort size.

Conclusions:

These proof-of-concept results suggest that combining peripheral single-cell multimodal profiling with interpretable machine learning can identify coherent predictive signatures of immunotherapy response. Future research in larger independent cohorts is necessary to validate these biomarkers for clinical use.