A Machine Learning Model to Predict Survival and Therapeutic Responses in Multiple Myeloma

Abstract

Multiple myeloma (MM) is a highly heterogeneous hematologic tumor. Ubiquitin proteasome pathways (UPP) play a vital role in its initiation and development. We used cox regression analysis and least absolute shrinkage and selector operation (LASSO) to select ubiquitin proteasome pathway associated genes (UPPGs) correlated with the overall survival (OS) of MM patients in a Gene Expression Omnibus (GEO) dataset, and we formed this into ubiquitin proteasome pathway risk score (UPPRS). The association between clinical outcomes and responses triggered by proteasome inhibitors (PIs) and UPPRS were evaluated. MMRF CoMMpass was used for validation. We applied machine learning algorithms to MM clinical and UPPRS in the whole cohort to make a prognostic nomogram. Single-cell data and vitro experiments were performed to unravel the mechanism and functions of UPPRS. UPPRS consisting of 9 genes showed a strong ability to predict OS in MM patients. Additionally, UPPRS can be used to sort out the patients who would gain more benefits from PIs. A machine learning model incorporating UPPRS and International Staging System (ISS) improved survival prediction in both datasets compared to the revisions of ISS. At the single-cell level, high-risk UPPRS myeloma cells exhibited increased cell adhesion. Targeted UPPGs effectively inhibited myeloma cells in vitro. The UPP genes risk score is a helpful tool for risk stratification in MM patients, particularly those treated with PIs.

Keywords

• multiple myeloma
• ubiquitin proteasome pathway
• proteasome inhibitors
• single-cell RNA-sequencing
• machine learning