Optimizing clustering

[agitter]

At the UW2020 meeting, there was discussion about how many cluster there should be. 60k for the LC library seemed like a lot. We should follow up with Scott and Spence to decide what types of clustering to use. It could impact the iterative pipeline as much as the next batch selector.

[Malnammi]

LC1234 has 94857 cpds with 133 actives. The following is the cluster information:

• Murcko Scaffold ID has 23582 clusters
• Cluster_0.3 has 68243 clusters. This was generated using: 1024-rdkit-MorganFPs, tanimoto dissimilarity <= 0.3, rdkit's Butina-Taylor clustering method.

[Malnammi]

More discussions on these in emails:

• We have a custom implementation of BT clustering here: code
• As a result, the current dataset at the dropbox, contains 5 clustering methods: Murcko Scaffold ID, rdkit's BT clustering, Clustering_0.2, Clustering 0.3, Clustering_0.4. Clustering_* are the custom implementation.
• Due to large number of true singletons in this dataset; i.e. Clustering_0.4 has over 15k+ true singletons out of ~22k unique clusters. Discussion led to suggesting using an adaptive/fuzzy/weighted clustering of boundary cpds. Spencer had this to say which I think is important to this:

I was thinking that when compounds are prioritized within a cluster, cpds that also belong to another previously explored cluster could be down-weighted. Structural ambiguity in the boundary cpds is unavoidable--some are likely to be truly chimeric and have scaffolds from 2 clusters.

Everyone agrees that this will add complexity to the prioritization, so for the first implementation, we plan on using fixed clustering methods.