Paper Details

ABSTRACT

Introduction:

Multiple Myeloma (MM) is a genetically heterogeneous disease, making individualised and accurate risk stratification difficult. Although staging systems have evolved to include some genetic data, much is still discounted. However, in the age of genomic profiling it is pertinent to incorporate more genetic information for prediction of disease trajectory. To achieve this, computational modelling can be utilised.
 

We have developed and validated a virtual B cell model, consisting of ordinary differential equations representing key reactions in B cells. The model can predict the impact of mutations on cellular networks in plasma cells, and how this affects survival and proliferation. We have previously shown that this approach can identify high-risk patients with greater accuracy than existing risk stratification systems. Here, we present an expanded model, including 54 genes with established or putative roles in the biology and pathogenesis of MM. Therefore, it has enhanced ability to predict disease behaviour based on a patient’s mutation profile.

Methods:

Whole exome sequencing data were acquired from 55 newly diagnosed MM patients who were enrolled onto the NCRI Myeloma XI trial. All patients were phenotypically high risk, defined as progression within 30 months of maintenance randomisation. All genetic mutations were graded on their oncogenic potential with the validated mutation database, OncoKB. All mutations deemed oncogenic were included in the model, along with copy number variations of genes (e.g. a gain of one copy of Cyclin D increased representation of Cyclin D expression by 50%).

55 virtual B cells were created, each representing a patient and their mutational profile – an in silico representation of their genetic landscape. These were simulated to run for 6 hours, and the predicted abundance over 280 molecules were obtained, with Cox Regression utilised to determine predictive potential.

Results:

Two pathways were found to be significant for prediction, NF-kB and Ras. In addition, TAK1 was significant, which is known to activate NF-kB.  We applied thresholds and dichotomised each group to stratify patients. Four computational signatures were identified: low NF-kB (n=14), low Ras (n=16), low Ras and NF-kB (lowNR (n=11)), and high Ras and NF-kB, (highNR (n=14)). Statistical analysis was performed on progression-free survival (PFS) with log-rank tests for survival curves and Cox Regression for hazard ratios (HR).

The computational signatures were highly significant in combination, with a p value of 0.0013. Patients with lowNR had the longest PFS and highNR performed the worst (HR 5.84 CI 2.26-15.1 p=0.00027). Having a high risk cytogenetic abnormality (HRCA) namely gain(1q), del(1p), del(17p), t(4;14), t(14;16) or t(14;20), trended towards poor PFS, but was non-significant for this cohort (p=0.07). There was no association between HRCA and computational signature. However, highNR, could successfully identify poor performing patients within the negative HRCA group, with a HR 4.587 (CI 1.51- 14.1) p =0.007, illustrating the added value of the computational model. Maintenance or induction treatment strategies were also non-significant predictors of outcome.

Combining computational signatures with ISS was highly effective. Combining highNR with ISS II/III revealed a poor performing subgroup, with much worse PFS when compared to high ISS alone - HR 2.93 (CI 1.43-6.02) p=0.003. Conversely, when combining the best prognosis groups, lowNR and ISS I, into one, they still performed much better than the remainder of the cohort - HR 0.444 (CI 0.244-0.808) p=0.008. Crucially, 8/55 (14.5%) patients with the lowNR signature and high ISS moved into the better performing group.

Conclusion:

In conclusion, we show that computational modelling can identify patients who are typically misclassified with current staging systems and assign more accurate predictions. The model has also exposed critical pathways, and genes that are important in myeloma pathophysiology, potentially revealing therapeutic targets. The model stratifies into four groups, which may be more representative of MM than dichotomous stratification, due to its unequivocal heterogeneity. It is acknowledged that validation is required in larger data sets, but the potential of computational modelling genomic profiles to ascertain risk and impact therapeutic strategy is clear.