Butina clustering

How to cluster molecules with Butina clustering?

Paste a SMILES library and choose a fingerprint and distance threshold. Butina clustering groups molecules around representative compounds using Tanimoto distances. The results show each compound's cluster, cluster size and representative, with downloadable assignments and a representative SDF for building a smaller screening library.

1. Paste the molecules to cluster, one SMILES per line.
2. Set the distance threshold. Keep 0.2 for a tight initial grouping.
3. Run the tool and compare cluster sizes.
4. Download representatives.sdf when one compound per cluster is sufficient.

Worked example

Input SMILES:

CCO ethanol
CCO ethanol_copy
c1ccccc1 benzene

At distance 0.2, the two ethanol records share a cluster and benzene forms another:

Name	SMILES	Cluster	Cluster size	Representative
ethanol	CCO	1	2	ethanol_copy
ethanol_copy	CCO	1	2	ethanol_copy
benzene	c1ccccc1	2	1	benzene

Input

Put the SMILES first, for example CCO ethanol, rather than a header row or name-first column. Blank lines are ignored. Unnamed molecules receive generated identifiers. Invalid SMILES fail the job instead of being silently dropped.

Settings

Results

Downloads are available from Files. The exported records retain returned row order.

Choosing a molecular clustering threshold

Butina uses Tanimoto distance, 1 - similarity. A distance threshold of 0.2 makes compounds with similarity of at least 0.8 neighbors of the representative. A larger distance threshold permits broader clusters; a smaller one restricts the neighborhood.

The threshold applies between the representative and its neighbors. Two members of one cluster need not meet it against each other. The native representative is a cluster center, not necessarily the molecule with the best potency, lowest weight or highest average similarity.

Changing the fingerprint can change the clusters. Keep fingerprint family, radius, size and chirality consistent across runs. Recalculate neighbors updates the remaining neighborhoods after a cluster is removed and can change the next representative.

Selecting representative compounds

representatives.sdf contains one center per cluster. Use it to start a diversity selection, then apply experiment-specific criteria. Cluster IDs belong to this submitted library and are not persistent chemical family names.

For a ranked list around one reference, use molecular similarity search. For a required motif, use chemical substructure search.

FAQ

Is Butina clustering hierarchical clustering?

No. It produces representative-centered clusters without a dendrogram or coordinate embedding.

Does molecular clustering delete duplicates?

No. Duplicate input records remain separate rows. The worked example puts both ethanol records in one cluster.

Why can two members of a cluster be less similar than the threshold suggests?

Each can be close to the center while being less similar to the other. Butina does not impose the threshold on every pair.