Single cell Raman spectroscopic profiles predict treatment responses in patients with de novo acute myeloid leukemia

Leukemia is a clonal malignant proliferative disease originating from hematopoietic stem cells. Although its treatment strategy has gradually developed from traditional chemotherapy to a multimodal treatment system including novel targeted therapy and immunotherapy, primary drug resistance in particular remains the core clinical problem leading to poor patient prognosis. This clinical dilemma indicates that the traditional genotyping system based on genomics has not been able to fully resolve the molecular heterogeneity of acute myeloid leukemia (AML), and it is urgent to establish a precise stratified model that can dynamically reflect the functional status of tumor cells in the initial stage of treatment.

In this study, Raman spectroscopy (RS) combined with machine learning algorithm was used to construct a metabolic prognosis prediction model for AML chemotherapy response. Bone marrow single cell Raman spectroscopy data of newly diagnosed AML patients were collected, and the molecular fingerprint was analyzed by principal component analysis linear discriminant analysis (PCA-LDA) and multivariate curve resolute alternating least square method (MCR-ALS).

The results showed that the PCALDA model achieved complete remission or non-remission (CR/NR) classification through 24 principal components (cumulative variance contribution of 90.1%), the accuracy of external validation was 94.8% (sensitivity 97.9%, specificity 92.0%), and the AUC reached 96.27%. Protein, lipid, nucleic acid and mixed components were decomposed by MCR-ALS, and lipid and nucleic acid metabolic pathways were enriched in NR group (P < 0.001).

Studies have shown that RS single-cell metabolic fingerprint can decode the metabolic reprogramming features associated with chemotherapy resistance in AML, providing a new marker-free and highly sensitive tool for real-time prognostic stratification and targeted intervention.