ECFP

ECFP stands for Extended-Connectivity Fingerprints, a class of binary molecular fingerprints widely used in cheminformatics to encode structural information for similarity searching and machine learning. They are based on circular substructures centered on each atom and grown outward to a fixed radius, capturing local neighborhoods rather than entire molecular graphs.

Generation of ECFP fingerprints involves selecting a diameter or radius, often expressed by the ECFPn convention where n is the diameter. For example, ECFP4 uses a diameter of 4 (corresponding to a radius of 2). The algorithm assigns an initial identifier to each atom from its properties (atomic number, valence, degree, etc.), then iteratively updates these identifiers by hashing the atom’s current value with those of its neighboring atoms to form a set of circular substructures. The resulting identifiers are mapped into a fixed-length bit vector (or integer counts), with bits set to indicate observed features. In practice, ECFP fingerprints are often implemented as hashed fingerprints, where many different substructures may collide into the same bit.

ECFP fingerprints are a specific realization of circular fingerprints and are closely related to Morgan fingerprints; FCFP (functional-class fingerprints) are a related variant that encodes atom classes rather than exact atom identities. In many software packages, ECFP is provided under the Morgan fingerprint implementation.

Applications of ECFP include measuring molecular similarity with the Tanimoto coefficient, building QSAR models, and conducting virtual screening in drug discovery. They are favored for capturing substructure patterns and pharmacophore-like features while remaining computationally efficient. Limitations include potential hash collisions, sensitivity to the chosen radius and vector length, and limited direct interpretability of individual bits. Nonetheless, ECFP remains a widely used, versatile tool in cheminformatics.