A signed response to comments on: Interpretation of genome-wide Infinium methylation data from ligated DNA in formalin-fixed paraffin-embedded paired tumor and normal tissue.
Farzana Jasmine1, Habibul Ahsan1, Mohammed Kamal2, and Muhammad G Kibriya1*
E-mails: farzana@uchicago.edu; habib@uchicago.edu; kamalzsr@yahoo.com and kibriya@uchicago.edu
Address:
1 Department of Health Studies, University of Chicago, Chicago, IL 60637, USA
2 Department of Pathology, Bangabandhu Sheikh Mujib Medical University (BSMMU), Dhaka 1000, Bangladesh.
*Corresponding Author: kibriya@uchicago.edu
In our recently published paper [1], we have shown genome-wide methylation data derived from modified Infinium protocol using ligated FFPE DNA as was originally shown by Thirlwell et al.[2]. It may be noted that Thirlwell et al.[2] did not compare tumor and normal tissue to identify tumor specific DML from FF and FFPE tissue. Our data suggested that the tumor specific DML detected in FFPE samples are not quite the same as the tumor-specific DML detected from FF tissue (gold standard) from exactly the same patients. Based on our findings we concluded that Infinium methylation data from FFPE should be interpreted cautiously. We thank Thirlwell et al. for their interest in our paper and commenting on the following issues:
Batch effect: We acknowledge that in the presence of batch effect there is potential for finding some false tissue specific DML (FF vs. FFPE). However, the fact that all the FF samples from tumor and normal were analyzed in a single batch and paired samples were put on same chip (same for FFPE samples) there should not be any batch effect while detecting the tumor-specific DML. Unfortunately, we found largely different sets of tumor specific DML while using FF and FFPE samples. It may be mentioned that in our experience Infinium methylation assay is very reproducible and inter-batch replicate (done in two different labs at different time) gave us very strong correlation (r = 0.96 to 0.98) [7].
DML analysis: we have shown the actual numbers and/or percentage of the overlapping tumor specific DML derived from FF and FFPE samples. We agree that statistically speaking 7 loci out of top 50 DML (14% overlap) being shared among the two lists is unlikely to be due to chance, however we leave the importance of this overlapping proportion to the readers and researchers who would like to clinically/biologically interpret the results. We also looked at the performance of modified Infinium protocol on FFPE samples for the usual suspects (frequently found DML in tumor) - see Additional file 6: Table S2 in the original manuscript) [1]. Results were not satisfactory for FFPE. Regarding sample size and power of the study, analysis of our own methylation data suggested that we had >90% power to detect a 1.25 fold change between tumor and normal groups (in other words a difference of methylation beta value of 0.1 and 0.125 or between 0.5 and 0.625) both in FF and FFPE samples.
Methodology for calling DML: It may be noted that Thirlwell et al. have somehow overlooked the fact that we used multivariate ANOVA to detect the DML. We also showed results from paired t-test and Bootstrapping (see Additional file 3: Figure S3 of original paper) [1].Therefore their comments regarding the variance/noise estimates does not apply to our data analysis. We acknowledge that we did not use Bayesian framework, but all three statistical tests that were applied to identify DML in our paper are pretty standard and statistically appropriate for the type of research question and there was substantial overlap between them (see Additional file 3: Figure S3 of original paper [1]) indicating that we picked up the true DML [1]. Previously we also have seen very good overlap between results from Illumina’s DiffScore and results based on the multivariate ANOVA [8] and therefore could not completely agree with their criticism of DiffScore analysis. Vast majority of true DML should survive any valid statistical tests applied to the data [8].
Concordance: We agree with Thirlwell et al. that p-value is often unstable. In fact we have tested the research question from different angles. The scatterplot from –log p-value was shown in additional material, but in the main body we also showed the loci specific correlation between FF and FFPE (see Figure six in the original manuscript [1]) for all the 27 K loci [1]. In our data, we could not see meaningful strong correlation. As per the suggestion by Thirlwell et al., we have now added the scatterplot of t-statistics from paired t-test between tumor and normal in FFPE (y-axis) over FF (x-axis) in Figure 1, which also does not show strong concordance between FF and FFPE results. In the original paper we went further to examine the effect of ligation on FFPE DNA by testing the performance on SNP genotyping platform. Although call rate was not very low, the performance was significantly poor than FF samples [1]. Performance of FFPE samples on DASL assay was also not optimal [9]. Although somewhat pessimistic, our words of caution for interpretation of genome-wide data from Illumina Infinium platform are based on our microarray data.
In fact we, like other cancer researchers, would be very pleased if one could get identical biological information regarding tumor specific DML or gene from widely available FFPE samples as one would expect from FF samples from same individuals. Unfortunately so far no published paper has clearly demonstrated that. It would have been better if Thirlwell et al. presented some real data showing tumor specific DML from FF and FFPE samples in their signed response.
Lastly, our entire data is submitted to GEO and will be released in due course. So, researchers are welcome to apply any statistical tests that they feel appropriate and make their own conclusions. Recently Illumina has further modified the assay by adding another step to restore FFPE DNA before ligation step for methylation assay. We are yet to evaluate that kit to investigate if FFPE samples can provide similar information as FF samples.