What is Tanimoto similarity?

The Tanimoto coefficient (also called Jaccard similarity) is the most common way to measure how similar two molecules are. It compares their fingerprints.

The formula

For two bit fingerprints A and B:

T(A, B) = c / (a + b − c)

where a is the number of bits set in A, b the number set in B, and c the number set in both. In words: shared features divided by all features present in either molecule.

Example: molecule A sets 40 bits, molecule B sets 50, and 30 are shared. T = 30 / (40 + 50 − 30) = 30 / 60 = 0.50.

For count fingerprints the same idea uses minimums and maximums: the sum of the smaller count for each feature divided by the sum of the larger count.

Interpreting the score

• 1.0 means the fingerprints are identical — usually, but not always, the same molecule. Different stereoisomers, for example, often share a fingerprint. Use exact structure matching to be sure.
• High values indicate close analogues. Thresholds depend on the fingerprint: with Morgan radius-2 fingerprints, pairs above roughly 0.6–0.7 are usually clearly related, while random pairs of drug-like molecules typically score well below 0.3.
• The same pair will give different scores with different fingerprint types, lengths or radii, so compare scores only within one method.

Similarity search in practice

A similarity search computes the Tanimoto score between a query and every molecule in a library, then returns the top hits. It is the workhorse of ligand-based virtual screening and of “analogue by catalogue” hit expansion, where you buy close analogues of a hit to explore its structure–activity relationship.

Limits

Tanimoto similarity is based on 2D structure. Two molecules can look different in 2D but present the same shape and features to a protein (the basis of scaffold hopping), and a small change such as one methyl group can sometimes destroy activity (an “activity cliff”). Similarity guides the search; experiments decide.