Applying computational modelling to a high-risk multiple myeloma data set to create novel risk stratification groupings

Abstract Introduction: Multiple Myeloma (MM) is a heterogeneous disease making it difficult to accurately predict the disease course in individual patients. Staging systems in MM have evolved, firstly the International Staging System (ISS), then incorporation of genomic aberrations with revised-ISS. Recently, the IMWG agreed on consensus genomic staging of high-risk MM. However, no scoring system captures the nuances of the genetic landscape of each MM patient. In the era of genomic profiling, this information should be utilised to predict disease behaviour, which can be achieved through computational modelling. We recently developed and validated a model signalling network within B cells, which predicts pathway activation and protein abundance in B cells. We hypothesised that inclusion of genetic mutations from MM as parameter changes in this model could enable the creation of patient specific virtual cells that could improve disease stratification. Methods: Genetic data for 53 newly diagnosed MM patients, enrolled to the NCRI Myeloma XI was collected. The cohort were phenotypically high risk, relapsing within 30 months of maintenance randomisation. All patients achieved at least a partial response prior to relapse. Whole exome sequencing was conducted at presentation and relapse. All genetic mutations (SNPs, copy number variants, translocations) were verified for their oncogenic potential with OncoKB. Each mutation in each patient was mapped to a model parameter to create a set of 53 unique patient models in silico (e.g. gain of one copy of BCL2 increased the parameter representing BCL2 expression by 50%). To determine the apoptotic, and proliferative signalling state of the patient models 6-hour simulation was performed and the predicted abundance of cytoplasmic Smac, cytochrome C and cadherin-1 was stored. Patients were grouped according to their predicted signalling state, namely anti-apoptotic (AA, n=16), pro-proliferative (PP, n=11), anti-apoptotic and pro-proliferative (AAPP, n=15) and non-proliferative or apoptotic (NAP, n=11). We compared progression-free survival (PFS) in days between different groups using Cox regression for Hazard ratios (HR) and Log-Rank test for p-values. Results: As a benchmark we first evaluated the ISS in our cohort. ISS II versus ISS I was associated with a significantly worse PFS (HR 2.27 95% CI 1.07-4.8, p=0.032) as well as ISS III (HR 2.45 CI 1.13-5.31 p=0.024). However, ISS II was not significantly different from ISS III, likely due to the phenotypic high-risk cohort assessed. High-risk lesions del(17p), gain(1q), del(1p), (t(4;14), t(14;16), t(14;20) and neutral lesions, del(13q), HRD, t(6;14), t(11;14) and t(MYC), were all non-significant predictors of progression (p>0.05). Maintenance strategy (lenalidomide vs observation) was also not a significant predictor of outcome, consistent with the phenotypic behaviour. We then applied the model-driven signalling state stratification (AA, PP, AAPP, NAP) and found model stratification alone significantly predicted PFS (p=0.047). The PP group was associated with the longest PFS and the NAP group had the shortest (HR 3.33 CI 1.37-8.11, p=0.008). No HR genetic lesions were noted in the PP group, compared to 2/11 (18%), 4/16 (25%) and 7/15 (47%) for the NAP, AA and AAPP groups respectively. To incorporate modelling with existing prognostic tools we combined all good prognosis patient groups (ISS I [n=13], and PP patients [n=11]) into one group and compared it to the rest. This stratification method identified 9/53 (17%) patients with high ISS (II/III), but good prognosis predicted through modelling, and assigned them to the low-risk group. The remaining high risk patients (n=29) had a significantly higher HR of 2.21 for shorter PFS (CI 1.24-3.94, p=0.007). Conclusion:We show that computational modelling can be utilised to stratify MM patients into both good and poor prognostic groups in ways that improve on previous staging systems. Furthermore, this has been illustrated within a phenotypical high-risk cohort, emphasising its ability to uncover distinct subgroups in previously hard to classify patients. This approach provides the foundation for a novel approach to improving treatment decisions, developing personalised approaches such as treatment intensification based on the full genetic profile of an individual patient. Characterising the signalling state in a larger cohort and predicting response to treatment is underway.