A High Density Polymorphism Analysis of Factors Correlated with Neurotoxicity to Bortezomib in Patients with Multiple Myeloma.

Abstract 4816

Bortezomib, a proteasome inhibitor, has become an essential compound in the treatment of multiple myeloma (MM). Clinically relevant (i.e. grade 2 or greater) peripheral neurotoxicity is a significant side-effect of bortezomib. As the occurrence of neurotoxicity is unforeseeable and apparently dose independent, we think that it might be associated with individual genetic characteristics such as Single Nucleotide Polymorphisms (SNP) or Copy Number Variants (CNV). In order to identify the genetic characteristics (SNP and CNV) correlated with the occurrence of neurotoxicity induced by bortezomib, we performed high density SNP analysis in 300 patients using Affymetrix Genome-Wide Human SNP Arrays 6.0. All patients were included in the prospective randomized trial IFM 99-01 and had received bortezomib and dexamethasone as induction therapy for previously untreated MM. Neurotoxicity was graded as severe (ECOG score of 2 or greater) or non severe (ECOG score of 0 or 1).

Quality controls included Contrast Quality Control (CQC), the call rate and the percentage of heterozygotes. CQC estimates the ability to separate the intensities corresponding to the different alleles into three clusters. In poor quality samples, CQC is excessively low and a large difference between CQC calculated on SNP that reside on Nsp and Sty fragments can indicate a single-enzyme target preparation failure. Only one sample was considered to have an excessively low CQC value (<0.4) and two samples were considered to have an excessive difference between CQC calculated on SNP that reside on Nsp and Sty fragments ( CQC_Sty − CQC_Nsp   > 2), according to Affymetrix recommendations for quality control on SNP6.0 chips. Then the birdseed algorithm (Korn et al., 2008) was used to genotype the samples.. The median call rate was 99.57%. Eight samples were excluded to a call rate < 97%. Percentage of heterozygotes is also evaluated as an excessive percentage can be a consequence of contamination. We thus excluded 3 additional samples for which the percentage of heterozygotes was greater than m + 3 sd (m: mean of heterozygotes in the entire series, sd: standard deviation). Further analysis was performed on 281 samples that meet these quality criteria and for which we know the neurotoxicity grade of the treatment. 833,192 SNP were considered on these samples.

For the CNV analysis, the Median of the Absolute Values of all Pairwise Differences (MAPD) metric was considered for quality control. According to Affymetrix recommendations, 22 samples with a MAPD>0.35 were excluded. Further analysis was performed on 270 samples that meet this quality criterion and for which we know the neurotoxicity grade of the treatment. 1,720,978 CNV markers were considered on these samples.

To identify the SNP and CNV most associated with neurotoxicity in patients receiving bortezomib, we use of two efficient and highly scalable feature selection techniques. The aim is to identify the relevant SNP and CNV markers that are strictly necessary to construct an efficient classifier of the bortezomib neurotoxicity from data alone. One is based on Support Vector Machines (SVM-RFE), a common and well-studied method in the Machine Learning community. The other uses a probabilistic feature selection technique based on recent Bayesian networks advances to identify a Markov Blanket of the target variable T, that is, a subset of the variables that renders the the remaining ones independent of T. Statistical significance of the selected SNP and CNV markers are evaluated using bootstrapping. Additional data will be presented during the meeting.