Using Descriptor Counts in Clustering

BOOST: Close-Up of Leaf Nodes

1. Even with a small (2,000-compound), fairly diverse dataset, we get most high-similarity groupings. The leaf nodes capture

• 80% of all pairings > 90% similarity
• 94% of all pairings > 95% similarity
• 96% of all pairings > 98% similarity

BOOST Leaf Node for a 2,000-Compound Dataset
(1024-bit Daylight Fingerprints)

16 compounds in node. 3 compounds have 1 hi-sim neighbor (>90%) in the dataset, and also in node. 1 compound has 2 hi-sim neighbors in the dataset but only 1 in node. The 12 unmatched compounds in the node have no hi-sim neighbors in the dataset.	Group 1
	Group 2
	12 Unmatched Compounds

2. In a larger dataset, condensation into small, high-similarity groups is very evident:

BOOST Leaf Node for a 10,000-Compound Dataset
(1024-bit Daylight Fingerprints)

Group 1	Group 2	Group 3 1 Unmatched Compound
15 compounds in node. 7 compounds have 6 hi-sim neighbors in the dataset, and all are in the node. 5 compounds have 4 hi-sim neighbors in the dataset, and all are in the node. 2 compounds have 2 hi-sim neighbors in the dataset, but only 1 in node. The 1 unmatched compound has no hi-sim neighbors in the dataset.

3. With really large datasets, high-similarity groupings can be larger than the leaf nodes, so we do several iterations.