Deep learning-based classifier for malignant plasma cell identification in myeloma.

Multiple myeloma (MM) displays significant genetic heterogeneity, making it challenging to distinguish malignant from non-malignant plasma cells in single-cell datasets. Existing marker-based and CNV detection methods require manual intervention or high computational resources, limiting their scope. We developed a supervised deep learning autoencoder to classify malignant cells across MM and its precursor stages and validated its performance and biological relevance. The model outperformed alternative approaches, achieving mean AUCs of 0.86 on internal and 0.80 on external datasets, with strong performance on unseen samples across single-cell platforms (mean AUC 0.92). Despite training solely on MM samples, our model distinguishes malignant cells in precursor stages, with predicted malignant cell proportions increasing from monoclonal gammopathy of undetermined significance (MGUS, 29.00%) to smoldering multiple myeloma (SMM, 77.83%) and MM (96.08%). In paired samples, the model also accurately captured a reduction in malignant plasma cell proportions following treatment, reflecting patient variability in treatment response. Differential expression analysis uncovered a 12-gene malignant plasma signature linked to poor survival (HR = 2.7, Log-rank P = 0.0023) and validated in the MMRF CoMMpass cohort. Overall, our resulting model enables accurate malignant cell identification and provides biologically and clinically relevant insights for MM research.