Objective  Cellular adaptation to treatment, particularly in malignant cells, causes acquired resistance in breast cancer. The exploration of biomarkers related to malignant adaptation in breast cancer is limited. This study aimed to identify key genes involved in therapeutic adaptation and resistance in breast cancer by integrating single-cell atlases with machine learning (ML) modeling.

Methods A large-scale single-cell atlas of breast cancer was curated. Comparative analyses between pre-treat and post-treat identified common features of malignant cells, and random forest (RF) was utilized to select genes and model the malignant cell adaptation program across multiple longitudinal cohorts. Functional validation of MORF4L2 was performed on breast cancer cell lines in vitro.

Results A universal adaptation program was revealed during breast cancer treatment. An adaptation score using 15 genes was modeled and the prognostic and therapeutic potential of adaptation was validated in multiple validation datasets and experiments.

Conclusions Integrated single-cell transcriptomics and ML modeling uncovers a tumor-intrinsic transcriptional program governing adaptation and driving resistance to treatment in breast cancer. This provides novel biomarkers and strategies for breast cancer treatment.