Should regression targets be log scaled? about tdc HOT 7 CLOSED

Hey Cyrus, that's a good point! That is also why we provided a log transformation function. You can try:

from tdc.single_pred import ADME
data = ADME(name = 'VDss_Lombardo')
data.convert_to_log()
data.label_distribution()

We want to keep the raw data as well, but we will add a note on the website about this important information.

I think in our current ADMET benchmark, we are not using that. So, it may be a good idea to use a log-scale for VDss and Half life. I will try that out to see if the performance makes sense. What other datasets have you identified this issue? Thanks!

from tdc.

Hi Kexin, that's for your quick reply! I went ahead and ran this for all the TDC endpoints that I'm working with. I only see benefits for VD, clearance, and half life. Note for the log transform, I use a robust version so it works for non-positive numbers...

def robust_log(x):
    return np.sign(x) * np.log(abs(x) + 1)

from tdc.

Thanks so much for making this table! It looks like you are using the old version of TDC. We have found that eDrug3D is very noisy, so we replaced them with higher quality ones. You can check them out in our website.

Would it be fast to generate these numbers for these new datasets? If not, let me know, i can also run some codes to test the difference.

Also, regarding the log transformation, i made a PR to incorporate your point: #56 one difference is instead of using +1 for numeric stability, i use 1e-10. Since it looks like some raw values are pretty small and +1 would make a difference.

from tdc.

Hi Kexin, good catch adding a smaller number! I got called off to do covid variant work, but I should be able to get back to this next week. Once it's ready, I'll add the updated table here.

from tdc.

Sounds good, thanks a lot!!!

from tdc.

Here you go! I updated the robust log computation and added the latest datasets:

from tdc.

Thanks so much! This looks good, it seems the currently supported dataset does not require log transformation. Closing for now! Feel free to reopen if you have any question!

from tdc.