DEEPScreen Supporting Data for Output/results about deepscreen HOT 1 OPEN

Thank you for your interest in DEEPScreen.

The actual binding affinity is not included in the prediction since we binarize all training data points as active and inactive before the training procedure. As a result, the model only sees labels as active/inactive and the prediction output is like that as well. In the current version of our tool, we selected binding affinities less than 10 uM as active and the ones greater than 20 uM as inactive, as a result, binarized prediction output reflects these values as well (i.e., an active prediction means that the actual binding affinity between the compound and target on interest should be less than 10 uM). It is possible to change these activity thresholds in the dataset pre-processing step if required (though it requires a bit of coding, we do not have an automated procedure for this yet).

Regarding the second question, we do not have a virtual docking tool incorporated into DEEPScreen. DEEPScreen is designed as a fast tool to scan large-scale compound sets on selected targets before using more targeted/directed methods such as docking and/or MD simulations. DEEPScreen can detect potentially active compounds from very large datasets, on which running targeted methods would be unfeasible. These targeted methods can easily be run on the compounds in a small dataset which are predicted as active by DEEPScreen. In other words, it is possible to run virtual docking on selected compound-target pairs (via a web tool or on a local server) after you get the predictions results from DEEPScreen and select the ones that seem interesting to you. This is the way we imagined DEEPScreen will be used.

from deepscreen.