Use of Genomic Information to Predict Treatment Response in Multiple Myeloma Patients By Computational Mapping of Protein Network Disturbances

Background: Multiple myeloma (MM) is an incurable heterogeneous hematological malignancy in which immune suppression and complex biology affect the disease and its response to treatment. Bortezomib (btz) and lenalidomide (len) alone or in combination with dexamethasone (dex) or other agents, are the predominant treatments for newly diagnosed and relapsed MM. Unfortunately, no precise method exists to predict disease response, making MM patient management difficult. Predicting treatment response would improve treatment effectiveness, and potentially reduce unnecessary treatment-related adverse events and health care costs.

Aim: To determine the application of a genomics-informed predictive simulation model in MM patients treated with btz or len in combination with dex.

Methods: Fourteen patients were selected from two datasets. Nine relapsed MM patients were identified from Washington University and 5 newly diagnosed MM patients were identified from the publicly accessible MMRF CoMMpass dataset. In all cases, whole exome sequencing and array CGH were performed. For each patient, every available genomic abnormality was entered into a computational biology program (Cellworks Group) that uses PubMed and other online resources to generate patient-specific protein network maps of activated and inactivated protein pathways (Doudican, et al, J Transl Med, 2015). Digital drug simulations with HMAs were conducted by quantitatively measuring drug effect on a composite MDS disease inhibition score (i.e., cell proliferation, viability, and apoptosis). Clinically, patients received standard of care treatment and clinical responses were recorded. Predictive values were calculated based on comparisons of the computer predictions and actual clinical outcomes.

Results: The models predicted that 9 patients would respond to combination treatment and 5 would not. All response predictions were properly matched to their clinical response, resulting in 100% PPV, NPV, sensitivity, specificity, and accuracy. Interestingly, the model predicted that 6 of the 9 responders would not have responded to btz or len alone; instead, response was predicted to combination therapy with dex.

Conclusions: Computational biology for MM demonstrated high predictive value for response to btz and len with dex. The model may be useful in uncovering the mechanisms for treatment failure and highlight additional pathways that could be targeted to increase chemosensitivity.