What is a molecular fingerprint?

Comparing two molecule graphs atom by atom is slow. A molecular fingerprint turns a structure into a fixed-length vector of numbers that records which structural features are present. Comparing two vectors takes nanoseconds, so a computer can compare a query with millions of molecules in seconds.

Main families

• Substructure-key fingerprints (such as MACCS keys) use a predefined dictionary: bit 42 might mean “contains a carbonyl”. They are easy to interpret but coarse.
• Path-based fingerprints enumerate linear paths of atoms and bonds up to a certain length and hash each path into a bit.
• Circular fingerprints, the most widely used today, describe the neighbourhood around every atom.

How Morgan / ECFP fingerprints work

Extended-connectivity fingerprints (ECFP), computed in RDKit with the Morgan algorithm, are built in rounds:

1. Each atom gets an identifier from its own properties (element, connections, charge, ring membership).
2. In each round, every atom’s identifier is updated by combining it with its neighbours’ identifiers, so it describes a larger circle around the atom.
3. After the chosen number of rounds (the radius), all identifiers are hashed into a vector, typically 1,024 or 2,048 positions long.

Radius 2 corresponds to ECFP4 (the “4” is the diameter) and is the standard choice for similarity searching.

Bits or counts?

A bit fingerprint only records whether a feature is present. A count fingerprint records how many times it occurs, so it can distinguish, for example, a molecule with one phenyl ring from one with three. Count fingerprints often give more meaningful similarity rankings for larger molecules such as macrocycles, which contain many repeated groups.

What fingerprints are used for

• Similarity search and clustering.
• Fast pre-filtering for substructure search.
• Input features for machine-learning models of activity or ADMET properties.

One caution: because features are hashed, different features can share a bit (“collisions”). Fingerprints are approximate descriptions, not unique identifiers — two different molecules can have identical fingerprints.