Dataset Raw File

[li-ziang]

Hi,

I found this work super interesting! I tried to reproduce the results in Figure3b, but I am unfamiliar with R so it's a little hard for me to run and understand the code. Is there a way for me to get the csv raw file of the dataset in this paper like the files in the ThermoMPNN repo, so that I can use the dataset in a python version code like ThermoMPNN?

For example, the s669 filr provided by ThermoMPNN looks like this.

head -n 5 s669.csv
pdbid,chainid,variant,score
1A0F,A,S11A,1.8
1A7V,A,A104H,2.69
1A7V,A,A66H,1.98
1A7V,A,A91H,1.7

Thanks!

[belmaran]

Hi, thanks for the interest. For ease of use I have added a new Supplementary Table 5 to the zenodo repository with all the variant effect/stability predictions and the aPCA fitness scores. You can download it here: https://zenodo.org/records/11260616/files/Supplementary_Table_5_aPCA_vs_variant_effect_predictors.txt?download=1

[li-ziang]

Hi, thanks so much for your prompt reply! I found this file is exactly what I needed, and will follow up if I have further questions.

[belmaran]

Here I leave a new version of the file with more complete EVE, popEVE and Tranception predictions: https://zenodo.org/records/11493742/files/Supplementary_Table_5_aPCA_vs_variant_effect_predictors.txt?download=1

[li-ziang]

That helps a lot, thanks!

[li-ziang]

Hi,

I have tried the new data of the models you provided and attempted to reproduce the results in Fig. 3C. I found that some models' predictions for the fitness score are NaN. Could you explain why these values appear? Currently, my method for calculating rho is to ignore the NaN values, essentially treating them as if they do not exist. This approach has yielded results similar to those reported in the paper. However, if the model fails to give a value, wouldn't it be better to penalize the NaNs in some way, for example, by setting them to a particularly large value, like 1000?

Thanks!