My aim is to cluster a set of molecules coming from a virtual screening (in the order of thousands of compounds). I want to cluster them by chemical similarity using hierarchical clustering. The distance metric used is the Tanimoto index. Ideally, I want to identify a limited number of clusters (10-30) that can be conveniently represented by the molecule corresponding to the centroid of each cluster.
My problem is that the results of hierarchical clustering seem to be strongly dependent on the linkage algorithm chosen. The optimal number of clusters predicted (in terms of Kelley criterion) seem to vary a lot according to different algorithms tested. Eventually, I am afraid that the results of my clustering will depend too strongly on the linkage algorithm chosen.
Is there a way to perform clustering without this kind of bias? Can consensus clustering help removing this problem? If so, is there any software that performs consensus clustering on molecular libraries?
Hierarchical clustering does not make much sense when applied to chemical compounds. Quantitatively, only similarities above a certain threshold have chemical meaning. I would rather recommend using some partitioning technique, e.g., k-means clustering (https://en.wikipedia.org/wiki/K-means_clustering).
I do suggest using hierarchical cluster analysis (HCA). When this technique is used, it is true, one has to decide which linkage algorithm (LA) to use and also which similarity function (SF) to apply. A nice book helping in those selections is this one by MacCuish & MacCuish: https://www.amazon.com/John-David-MacCuish-Norah-Bioinformatics/dp/B0088RAKTS
But the subtle problem is that even if one knows which LA and SF is going to use, the problem of many ties in proximity is always present. This is the problem of having many equidistances among objects (in this case molecules), which ends up in a combinatorial explosion of possible dendrograms. A paper where we have addressed the problem bringing statistical support to HCA results is this one:
Article How frequently do clusters occur in hierarchical clustering ...
So, if you know which LA and SF to apply, I would strongly suggest using the algorithm presented in our paper.
Besides clustering algorithm, the molecular fingerprint used to calculate the similarity is another (maybe more) critical factor to affect your result. There are many different types of molecular fingerprints available, each of them focuses on different aspect of the compounds. One of the commonly used fingerprint is MACCS (http://www.dalkescientific.com/writings/NBN/fingerprints.html). Openbabel can deal with path-based fingerprint FP2 and structure-based fingerprints FP3 and FP4 as well. (https://openbabel.org/docs/dev/Fingerprints/intro.html) With the same clustering algorithm, I suggest to try different types of molecular fingerprints, or to define your own fingerprint to calculate the similarity.
- Badges
- Science topic
- Similar topics
- Higher Education
- Universities