How We Did On D3R Grand Challenge 4
It was late September in Barcelona and, at Acellera, we had only a few days left to make the first submission for the D3R challenge.
The D3R challenge is an annual, international competition which offers several tasks considered to be of pharmaceutical interest. These tasks usually involve predicting the binding mode and affinity of a series of molecules in a given protein. We thought we could use several Apps of PlayMolecule, e.g. Kdeep and DeltaDelta to make the affinity predictions and a new docking tool, SkeleDock, to generate the poses.
If you use our Playmolecule platform, you might already know about Kdeep and DeltaDelta. However, SkeleDock is still an in-house project which we have started developing last summer. It is a docking protocol that takes advantage of previous structural knowledge to generate predictions. If the ligand-protein system for which you have to make a pose prediction is similar to some other, already crystallized complex, the tool searches for equivalent chemical moeites between the two and model the position of those moieites in your system so that they match their corresponding place in the cocrystal. Ideally, with some further minimization process, we can produce high quality poses, which one could then submit to KDeep or DeltaDelta in order to get affinity predictions.
Internal tests at the end of August showed promising results. However, when the D3R challenge started on the 4th of September, we still had not considered the modeling of macrocyles, which were present in the vast majority of the ligands in the challenge. Modeling them was not an easy task. Our initial attempts were not what we expected: while the non-cyclic parts of the molecules did match the template that we provided, the cycles did not fit well; hence, the resulting poses were far away from the pose quality that we wanted. We worked on a temporary hack on the code to solve these issues and we were finally able to generate poses which, to our eyes, looked sensible. These poses were then used as input for KDeep and DeltaDelta to predict binding affinities.
Today, the D3R organizers have announced the results for the different subchallenges of the competition.
We are happy to report that we have ranked first in two of them (BACE free energy prediction and BACE scoring, both in stage 2), as well as achieving 6th position (depending on which metric is used) in the pose prediction challenge among more than 80 participants.
The median RMSD was 0.87 Å, where the template we used for most of the ligands [PDB code: 3K5C] has a resolution of 2.12 Å.
There were other tasks in which we participated where we did not perform as well; namely, those involving Cathepsin S. In this target, waters seem to play a very important role which, at this time, not SkeleDock nor KDeep account for. We will work on these aspects to improve our tools and, hopefully, we might perform better next time.
We want to congratulate all the winners and thank the D3R team for organizing this fantastic event. It has been a great exercise from which we have learned a lot. SkeleDock will be soon available on Playmolecule platform and an article will be published with further details on the protocol we used. All results can be seen here, under the tab: Evaluation Results.