Querying training label

[Clara-mm]

Hi Ludovic,

Thank you for this great resource! Just a couple questions. "Active compounds were defined as molecules with IC50, XC50, EC50, AC50, Ki, Kd or potency inferior or equal to 10 uM". Which of measured values has been selected in label generation?
Does your model solve a binary classification task? And what's your training label(1/0 or active/inactive)?

Thank you,

Clara

[ludovicchaput]

Dear Clara,

The procedure we used to generate the dataset of active compounds is described in the paper : https://academic.oup.com/bioinformatics/article-abstract/36/14/4225/5836501?redirectedFrom=fulltext

There is only active compounds in the dataset we provide. The spirit of the program is to identify putative targets for the query compounds, based on the similarity between the query compounds and the active compounds (with known targets) in the dataset.

We did not incorporated known "inactive" compounds in the dataset, because it raises the problem of the activity cliff concept, which is a complex subject. Especially in our case, where we deal with hundreds of thousands of active compounds and thousands of targets for which:

• the activity/inactivity matrix is largely incomplete,
• and the active/inactive ratio is highly imbalanced.

Ludovic