Assessing the prognostic value of transcriptomic data in multiple myeloma through machine learning methodologies.

Abstract

7513 Background: Multiple myeloma (MM) is a hematological malignancy characterized by the clonal proliferation of plasma cells in the bone marrow. Identifying the characteristics within transcriptomic data of MM patients could be imperative for elucidating long-term cancer prognosis. Prediction and bioinformatic evidence could be valuable for therapeutic intervention. Methods: To discover a precise collection of biomarkers, a large-scale disproportionality analysis of MM patients using the FDA Spontaneous Reporting System database is conducted. This analysis aims to provide pharmacovigilance evidence regarding severe adverse events (AEs) due to medications. Additionally, pharmacogenomic insights from OMIM and MM drug targets are integrated into the analysis to carefully select 84 essential biomarkers. Seven different machine learning (ML) algorithms, namely AdaBoost, K-Nearest Neighbors (KNN), Decision Tree (DT), Gaussian Naïve Bayes (GNB), Support Vector Machine (SVM), Random Forest (RF), and Multilayer Perceptron (MLP) were then employed to assess the classification of disease outcomes. Specifically, the focus was on distinguishing deceased patients (due to MM) from surviving patients in the MM Research Foundation (MMRF) dataset based on the identified essential biomarker. Results: The classification involved a total of 787 patients from MMRF (624 alive and 163 deceased). Four sampling strategies (original raw data, undersampling, oversampling, and simultaneous sampling) were included in the training and testing process on ML models to alleviate the influence of imbalanced data. SVM achieved the highest accuracy while RF achieved highest mean across four sampling methods (Table). We then performed the Kaplan-Meier survival analysis of the prognostic values of essential genes. The expression of two genes (BRAC1 and CTLA4) from MMRF data when comparing deceased to survival patients was associated with patient outcomes. Additionally, seven genes were potentially prognostic factors in MM (Positive prognostic: ATM, CYBA, NR3C1, PIK3CA, and PIK3CG; Negative prognostic: IFNG and NTRK2). Conclusions: This study successfully showcased a computational methodology that holds potential for optimizing future clinical trials in MM. The approach includes identifying the most suitable patient population for a particular MM regimen, thereby contributing to improved patient care. It also offers strategies for reducing AEs, lowering mortality rates, and ensuring the optimal allocation of healthcare resources. [Table: see text]