P-357: Individualized Risk in Newly Diagnosed Multiple Myeloma

Introduction: Outcomes for newly diagnosed multiple myeloma (NDMM) patients are heterogenous, with overall survival (OS) ranging from months to over 10 years. Though genomic events strongly contribute to clinical outcomes, yet the “one-size-fits-all” treatment paradigm remains dominant for NDMM.

Methods: To decipher the molecular and clinical heterogeneity at unprecedented depth and develop the first artificial intelligence individualized prediction model for risk in NDMM (IRMMa), we assembled a large training (n=1933) and validation (n=256) set of patients with NDMM with available clinical-, genomic-, and therapeutic data. The training set was composed by the following cohorts: MMRF CoMMpass (n=1062), Myeloma Genome Project (n=492), Moffit AVATAR (n=177), and MSKCC MyTYPE (n=109). The validation set included patients enrolled in the GMMG-HD6 (NCT02495922) clinical trial with available whole genome sequencing (WGS).

Results: To integrate clinical, demographic, genomic, and treatment data we compared different methods of predicting OS and event-free survival (EFS). Overall the Neural Cox Non-proportional-hazards (NCNPH)-based model emerged as most accurate in predicting EFS and OS, (c-index 0.69 and 0.73, respectively) and was used as the core-engine of IRMMa. IRMMa’s accuracy was significantly higher than all existing prognostic models: ISS (EFS: 0.56; OS: 0.61), R-ISS (EFS: 0.54, OS: 0.57), R2-ISS (EFS 0.56; OS: 0.63). Among all 132 genomic features tested, only twenty significantly improved model accuracy for OS, including deletions on 1p, 1q21 gain/amp, TP53 loss, t(4;14)(NSD2;IGH), a high contribution of the APOBEC mutational signature and copy number variation signatures indicating chromothripsis. Importantly, while the inclusion of each feature improved the model accuracy, IRMMa has been developed as a flexible tool able to predict outcomes with incomplete data. Specifically, because genomic profiling is only rarely performed in current clinical practice, IRMMa performance was tested without genomic data. Despite this, IRMMa still outperformed ISS, R-ISS, and R2-ISS with EFS and OS c-index of 0.69 and 0.71, respectively. IRMMa accuracy and superiority compared to other prognostic models were validated on 256 patients enrolled in the GMMG-HD6 (NCT02495922) clinical trial.
As a key innovation, IRMMa is able to predict differences in outcomes across different treatment combinations and strategies. For instance, using IRMMa we were able to predict NDMM patients, in whom high-dose melphalan followed by autologous stem cell transplant (HDM-ASCT) provides a significant advantage, and those in whom HDM-ASCT does not impact outcome.

Conclusions: Integrating clinical, demographic, genomic, and therapeutic data, we have developed the first individualized risk-prediction model enabling personally tailored therapeutic decisions for NDMM patients.