Molecules clustering code (PCA and K-means) for training datasets:

[Khrystofor19]

Could you please add the script/code for PCA and K-means clustering on the training dataset?

From the paper: "For all public datasets, compounds were clustered into five clusters using K-means based on a PCA-reduced 2048-bit Morgan circular fingerprint (radius 2), and one of the clusters is selected as a test set with the remaining four used as a training set."

I just wanted to make sure the process is as reproducible as possible.

[sveccham]

Hello @Khrystofor19,

We have shared both the public datasets including the train/val/test splits on Figshare (link to datasets). I hope this makes it easier for your to reproduce our results.

Let me know if this does not address your issue.

[Khrystofor19]

Thanks @sveccham ! I meant just PCA parameters like variance and component count; I was just trying to make sure that correct molecules split is maintained for fine-tuning (i.e. for train/val/test splits).

[sveccham]

Hi @Khrystofor19,
So the molecules in each of the splits (train, val, and test) are already in different files in the Figshare link. So you would not need the code that actually performed the splitting. Are you looking to use this procedure to split other datasets?

[Khrystofor19]

@sveccham yes! I was hoping to try benchmarking KERMT against other molecular foundation models but I wanted to make sure the scaffold-cluster molecule split is consistent.