The brain tumor classifications based on machine learning models trained with older version of methylation microarray chip compatible with the new EPIC v2 illumina’s chip.

Methylation profiling is a critical tool for brain tumor classification, and the Illumina methylation EPIC microarrays have been the dominant platform used for this purpose. The version 1 of the EPIC beadchips released in 2015 covers approximately 850k CpG sites, while version 2, released this year, has coverage over 950k CpG sites. Most probes on both chips overlap. However, the vendor has replaced over 15% of probes in v2 to address issues such as underlying SNPs, non-unique mapping, and off-target hybridization [1]. Moreover, EPIC v2 has over 200K additional new probes designed to enhance coverage of poorly covered genomic regions like enhancers and open chromatin regions. The EPICv2 lacks about 144k old probes present in the original content, comprising 16.5% of the old array content (Figure 1). If some of these missing probes were heavily used in predictors, they could potentially affect classifier results. The impact of changes in chip content on classifiers' predictions is uncertain. Herein, we validated the new methylation array to show that the new chip design does not affect tumor classification, although we found it may affect the classification score. Two approaches were used for the validation: 1. We performed technical DNA replicates on both chip types. This involved reprocessing and hybridizing DNA from 16 FFPE samples of various brain tumors previously run on the Illumina EPICv1 methylation chips. For the missing probes in new design (EPICv2), we replaced beta values or intensities with sample mean. We conducted prediction tests using four available classifiers and found no differences in tumor class calls in all samples and all tested classifiers. However, the new chip design had a significant effect on prediction scores in some classifiers. The figure 2 illustrates the correlation of class prediction scores between the chips and a calibrated scores between chip types on box plots. Furthermore, we examined the effect of probe imputation on the prediction of tumor content using three previously published algorithms: RFpredict.ABSOLUTE, RFpredict.ESIMATE, LUMP, and the prediction of MGMT promoter methylation using an R package. Our analysis revealed that replicates on both array types are highly correlated and produce compatible results. 2. For the second approach we conducted an in silico assay. Here we modified the original results of EPICv1 and 450k chips where we replaced methylation beta values for the 485 probes with median beta values from the entire sample. These 485 cpg probes are used amongst 10k in CNS classifier v12b6 developed by DKFZ in Heidelberg. Simulation done on 22,504 real brain tumor samples which were re-run on the v11b6 classifier and compared with the original results: the unmodified beta values. We found no differences in the predicted brain tumor classification between the original and simulated data. The calibrated prediction scores were highly correlated (R>0.98). However, there was a small but highly significant drop in the prediction calibrated score (Paired t-test p-value <2.2E-16).