Correction in cancer tissues
Correction in the article by Horvath S (2013) Genome Biol. 2013 Oct 21;14(10):R115 PMID: 24138928
By Steve Horvath, October 27, 2013
All of my results involving healthy, non-cancerous tissues are unaffected but I have to report some corrections for the cancer tissue data. In particular, I have to retract the statement that cancer is associated with an increased DNA methylation age (i.e. positive age acceleration) in most cancer types. While some cancer types show positive age acceleration, others show the opposite behavior (negative age acceleration). I deeply regret this software coding error. Fortunately, all of the other statements about cancer remain largely intact since the coding error effectively added an offset term to predicted age. The effect of the coding error on the age estimate is illustrated in Figure 1. I am comforted by the fact that most of the reported results for cancer become even more significant and meaningful including the following.First, the results for cancer tissues are now more congruent with those obtained for cancer cell lines (which remain unchanged). Second, the age predictor leads to a much lower error in cancer tissues (now 16 years). Third, the results for TP53 become more significant, e.g. TP53 mutations turn out to be associated with lower age acceleration in colorectal cancer. The error arose from me using the wrong age calibration function for the cancer tissue data sets, which led to a systematic over-estimation of DNA methylation age (Figure 1).
Figure 1. Evaluating the effect of the coding error on the DNAm age estimate in the cancer samples.
The old, incorrect estimate of DNAm age (y-axis) versus the correct estimate (x-axis). Note that the two estimates are highly correlated (r=0.98), which explains why most results are unaffected, but the old estimate is poorly calibrated. The coding error effectively added an "offset term" which led to an average bias of 42 years. After using the correct version of DNAm age, not all cancer types show positive age acceleration. While some cancer types are associated with high positive age acceleration (e.g. glioblastoma multiforme), others are associated with strong negative age acceleration (e.g. uterine corpus endometriod carcinoma), see Additional file 13.
As a result of this error, the following Figures and Additional files changed (since they reported age acceleration effects in cancer tissues).
1. Figure7MutationCountAll.pdf Figure7MutationCountAll.pdf
2. Figure8BreastCancer.pdf Figure8BreastCancer.pdf
3. Figure9ColorectalGBMCancers.pdf Figure9ColorectalGBMCancers.pdf
4. AdditionalFile12DataDescriptionCancer.xlsx AdditionalFile12DataDescriptionCancer.xlsx
5. AdditionalFile13PredictionInCancerTissues.pdf AdditionalFile13PredictionInCancerTissues.pdf
6. AdditionalFile14TumorGrade.pdf AdditionalFile14TumorGrade.pdf
7. AdditionalFile15MutationCountBreastCancer.pdf AdditionalFile15MutationCountBreastCancer.pdf
8. AdditionalFile16SelectedImportantMutations.pdf AdditionalFile16SelectedImportantMutations.pdf
9. AdditionalFile17TP53Mutation.pdf AdditionalFile17TP53Mutation.pdf
A word file with the above description can be found here: CorrectionCancerDescription.docx