Sparse integrative dictionary learning resolves conserved drug-response states in multiple myeloma (MM).

Multiple myeloma (MM) is a severe plasma-cell malignancy that continues to cause substantial morbidity and mortality despite major therapeutic advances. Although current therapies target distinct molecular processes, many patients eventually relapse, raising the question of whether diverse treatments impose different pressures that nevertheless converge on shared transcriptional adaptation programs. While many studies have profiled transcriptional changes before and after treatment, these responses are often analyzed within individual drugs, leaving recurrent programs across therapies insufficiently explored. Here, we developed PRISM-MM (Perturbation Response Inference by Sparse integrative Modeling for Multiple Myeloma), an interpretable sparse integrative dictionary-learning framework that decomposes drug-control transcriptional shifts into signed response programs while accounting for drug identity, study background and cell-source state. By incorporating paired scRNA-seq as a calibration layer, PRISM-MM preserved cell-state-level interpretability and outperformed representative perturbation-prediction models in recovering reproducible response structure. Applied to a curated multi-study MM perturbation compendium, PRISM-MM identified conserved treatment-associated programs validated in held-out bulk and external paired scRNA-seq datasets. These programs revealed a recurrent remodeling pattern characterized by depletion of plasma-cell-like secretory activity and enrichment of inflammatory, adhesion-associated and stress-tolerant immune-interface states. We further nominated state-maintaining genes and highlighted a RFXAP-centered immune-regulatory axis that may drive the Program 8 inflammatory adaptation state and shape MM drug response. Together, PRISM-MM provides an interpretable framework for discovering conserved drug-response programs and generating mechanistic hypotheses about treatment adaptation in MM.