Got negative values for predicted activities

[boaty]

Dear authur,

I used your demo run on my dataset, but the result made me confused, because i got negative predicted values. how is this possible.

the model applied is 'model_chembl_noFEPligands" in your actfound_demo. i also tried another models from your google drive check_points results folder, the results were pretty much the same.

is there some important parameters to better fit my data? thanks

yz

[BFeng14]

Hi, thank you very much for your interest on Actfound, Actfound is trained to predict the relative activity difference between compounds, and then reconstruct the activity value of test compounds using the predicted activity difference. There is no limit to the predicted activity value ranges, so it is normal to find predicted activity lower than 0. BTW, I find that most of the real activity is near to zero, can you tell me the unit and meaning of your activity? which might be helpful for me you give futher suggestions.

[boaty]

thank you BFeng14

it is IC50, on microMol. I tested actFound with a collection of compounds from 3 patent papers, which compounds share similar scaffold but differ by R-group. i did reach first_step_loss < 0.8 but the corelation between real and pred is not good.

thanks

[BFeng14]

Can you tell me the unit? did you use pIC50=-logIC50, or directly using the IC50 values? In my understanding, pIC50 is a value usually ranges between 6 and 10, but I found that there are many values near zero.

[boaty]

sorry for the confusion. it is directly IC50

[BFeng14]

as it is mentioned in the "Bioactivity definition" part of our paper, we use pIC50=-logIC50 for training and testing. you can try to run actfound using pIC50 and see the results.

[boaty]

thanks.

and I got R=0.6 for correlaiton, impressive.

thanks