GrimAge not matching what I get from Clock Foundation

[audreylk]

I've tried translating the model for GrimAge from python into R and calculating GrimAge on my own. I'm comparing it to GrimAge that I get from submitting the same data to the Clock Foundation online calculator. I'd love to be able to just use the code and not the black box that is the clock foundation website. Has anyone else tried this comparison? I'm wondering if the clock foundation provides values for CpGs that are missing from the dataset without telling you, but then I would expect every sample to be shifted equally in the same direction, if the same beta is being supplied to everyone. But my estimates and component scores are not shifted reliably like that. Thoughts?

[sarudak]

Imputation may be happening not with CpGs for every sample but for missing values that only some samples have. It's also likely that the imputation they use on Clock Foundation is different from what is default in Biolearn. I don't know anything about the clock foundation calculator or how it works. Biolearn implementations were adapted from GrimAge code provided by the Horvath lab and it is verified to produce the same results as the code they provided.

[audreylk]

Thanks for the response. I am also thinking they probably do some normalization on the betas without telling you that they're doing it... I was worried that I had done something wrong, because I tried translating the BioLearn code for GrimAge from the repo to R, where I'm more comfortable and work more often, so I was comparing to my results from the Clock Foundation. The more fair comparison is to run the BioLearn python code and compare that, because I can "see" everything that is happening to the data. It's still entirely possible I did something wrong with my R translation, but it seemed like a pretty basic application of the weights for the component scores, then weighting those component scores and scaling, so I'm leaning towards normalization behind the scenes on the part of the clock foundation.

[sarudak]

From Discord
"Qiming — 2/22/25, 8:09 AM
Hi Seth. Thanks for the great package! I found if I used my dataset to calculate the GrimAge V2 results differed from results from the website DNA Methylation Age Calculator (https://dnamage.clockfoundation.org/). Therefore, I checked the coefficients of GrimAgeV2 and I found that the Cox coefficient of DNAmlogCRP (0.403587277157754 in your files) should be 1.90266 in the original paper (https://doi.org/10.18632/aging.204434, DNA methylation GrimAge version 2), while DNAmlogA1C (1.90266039640722 in your files) should be 0.40359. Shall I keep it as your package or exchange it using the original coefficients?"

[DrChablis]

I tried both coefficients from Biolearn and papers, and the results are totally different. If I using Biolearn coefficients, the GrimAgeV2 were younger than the results from the Clock Foundation online calculator and the mean value were less 20 years. While when I using the papers coefficients, the value of GrimAgeV2 were younger than both results using Biolearn and online calculator and the mean value were lower about 40 years than online calculator. It is very odd if the code doesn't match the paper or it provided by developer, the results should not have such big differences between 3 coefficients.

[sarudak]

I got in touch with Ake Lu and she confirmed that the coeffecients we are using in Biolearn are correct as they are from the R code that was used for all the calculations. I do not know what the clock foundation might be doing to change the values and I do not have access to submit any data to investigate.