Developing a SNP Classifier for Predicting Peripheral Neuropathy by Bortezomib in Multiple Myeloma Patients.

The prevalence of peripheral neuropathy (PNP) during the treatment of MM with Bortezomib is high. About 20% of patients develop a grade 3-4 PNP due to this treatment, and as a result Bortezomib treatment is stopped or a reduced dose is given. Therefore, there is a strong need to find markers which predict the susceptibility of a patient to develop Bortezomib related PNP. Materials and methods: Bortezomib treated patients from the Dutch/German Hovon 65 GMMG-HD4 trial and the French IFM-2005/01 trial were used for this analysis. In both trials, the efficacy of Bortezomib as induction treatment prior to high-dose therapy is evaluated and PNP status was recorded. Samples were genotyped using a custom-built molecular inversion probe (MIP)-based single nucleotide polymorphism (SNP) chip containing 3404 SNPs (Bank on a Cure program; Van Ness et al., 2008). In total, 232 patients who did not develop PNP were compared to 210 PNP cases (grade 1, n=82; grade 2 n=86, grade 3, n=31, grade 4, n=11).

The data were processed on the basis of the following criteria. First, SNPs genotyped in less than 75% of the samples were removed (n=155). This resulted in elimination of 59% of the data with unknown genotype while only 1% of the genotyped data were lost. The remaining 41% of the missing data were imputed using BIMBAM (Guan et al., PLoS Genet. 4:e1000279, 2008). As reference panels, the data sets of the BOAC chips from this study, 500 random samples from the Rotterdam ERGO study (Köttgen et al., Nat. Genet. 41, 712–717, 2009) and 60 phased CEU HAPMAP samples were used. Secondly, SNPs were excluded which did not show any genotype variance and which were not in Hardy Weinberg equilibrium. As a last step the data was adjusted for stratification using Eigenstrat (Price et al., Nat. Genet. 38: 904–909, 2006). By removing 21 SNPs and 14 samples the variance between the IFM and Hovon was reduced to an acceptable level (p = 0.011). The resulting combined IFM/Hovon dataset now contained 2764 SNP and 428 samples. The data set was divided in 6/7 (n=367) part as a learning set and 1/7 (n=61) as a validation set. Possibly informative SNPs were selected using information gain as a feature selection method (Cover et al., Elements of information theory. New York, John Wiley, 1991). 66 SNPs with an information gain in allele and genotype frequency were selected (p value < 0.05 after permutation test (n=10000)). Classifiers generated by Partial C4.5 decision tree (PART), support vector machine (SVM) and Random forest learned on this set reached a better than random performance. Sensitivity, specificity, positive predictive value and negative predictive value were respectively 55%, 70%, 60%, and 66% for the PART classifier.

Preliminary classifiers generated by this dataset suggest that building a classifier with clinically relevant performance may be within reach. To this end, we will report on the outcome of different combinations of existing classifier methods and feature selection methods.

Van Ness, B, Ramos, C, Haznadar, M, Hoering, A,Haessler, J, Crowley, J, Jacobus, S, Oken, M, Rajkumar, V, Greipp, P, Barlogie, B, Durie, B, Katz, M, Atluri, G, Ganf, G, Gupta, R, Steinbach, M, Kumar, V, Mushlin, R, Johnson, D, and Morgan, G. (2008). Genomic Variation in Myeloma: Design, content, and initial application of the Bank On A Cure SNP Panel to analysis of survival. BMC Medicine. 6:26.

Hanifi-Moghaddam:Skyline Diagnostics: Employment.