SCORING
DockScore 2.0
Our scoring function, DockScore 2.0 (DS2), uses physico-chemical interface features and machine
learning to distinguish native-like poses from decoys. The features are first computed using a
Python re-implementation of the scoring function from the
DOCKSCORE paper. The features are
then passed through an XGBoost filter that distinguishes between incorrect and correct
(acceptable, medium and high quality - based on CAPRI criteria) poses. The correct poses are
then ranked using a graph neural network.
This pipeline has been trained, validated and tested on the BM5.5 dataset with complexes
generated using LightDock. Additionally, out-of-distribution performance has been evaluated on
the CAPRI Scoreset v2022.
The source code is available on
GitHub.