Molecular Fingerprint

These fingerprints encode the presence or absence of a predefined dictionary of structural fragments. The most famous example is the MACCS (Molecular ACCess System) keys, which use a set of 166 or 960 expertly curated substructures.

• Mechanism: Each bit in the bit-string corresponds to a specific pattern (e.g., 'is a carbonyl group present?').
• Key Property: Highly interpretable, as each bit has a fixed chemical meaning.
• Limitation: Cannot encode novel fragments not present in the dictionary, leading to sparse representations for unusual chemotypes.

Also known as Daylight-style fingerprints, these encode all possible linear atom-bond paths up to a specified length within the molecular graph. They are generated by exhaustively enumerating every connected subgraph path.

• Hashing: Each unique path is hashed into a dense bit-string, typically 1024 or 2048 bits long.
• Branching: Standard implementations do not capture branching patterns, only linear sequences.
• Use Case: Excellent for rapid substructure screening and exact match queries in large chemical databases due to their exhaustive nature.

The Extended-Connectivity Fingerprint (ECFP) is the de facto standard for modern drug discovery. It encodes circular atom neighborhoods through iterative Morgan algorithm expansion.

• Iterative Process: For each heavy atom, the algorithm aggregates neighbor information up to a specified diameter (e.g., ECFP4 for diameter 4, capturing up to 2 bonds away).
• Invariance: Designed to be invariant to atom numbering and molecular orientation.
• Pharmacophoric Variant: FCFP replaces atom type identifiers with pharmacophoric features (donor, acceptor, hydrophobic), enabling scaffold hopping.

These fingerprints abstract away from specific atom connectivity to encode the spatial arrangement of pharmacophoric features. Instead of carbon or nitrogen, they record hydrogen bond donors, acceptors, hydrophobic regions, and charged groups.

• Triplet Encoding: Often encode all possible triangles of three pharmacophoric points and the distances between them.
• Fuzzy Matching: Enables the identification of structurally dissimilar molecules (scaffold hopping) that share a similar 3D interaction potential.
• Application: Critical for lead optimization when the core scaffold must be replaced to improve ADMET properties.

A paradigm shift from discrete bit-strings to dense, real-valued vectors learned end-to-end by a neural network. These are generated by graph neural networks or sequence models operating on SMILES strings.

• Differentiable: Unlike traditional fingerprints, these can be optimized via backpropagation for a specific task.
• Autoencoder Latent Space: Variational autoencoders (VAEs) can learn a continuous latent space where interpolation between points generates chemically valid molecules.
• Representative Model: The encoder of a Message Passing Neural Network (MPNN) acts as a learned fingerprint generator, capturing task-relevant features automatically.

A specialized 3D conformational fingerprint that encodes the torsion angles of rotatable bonds. It captures the shape of a molecule in a specific low-energy conformation.

• Conformer-Dependent: Unlike 2D fingerprints, the output changes with the molecule's 3D geometry, making it sensitive to stereochemistry and ring conformations.
• Application: Used in 3D-QSAR and shape-based virtual screening where the bioactive conformation is known or hypothesized.
• Representation: Often binned into discrete angle ranges to create a bit-string representing the conformational ensemble.

A metric for quantifying the structural overlap between two molecules by comparing their binary fingerprint vectors. Defined as the ratio of the intersection to the union of set bits.

• Ranges from 0 (no shared features) to 1 (identical fingerprints)
• A threshold of 0.7–0.8 is commonly used for analog identification
• Computationally efficient for screening billion-compound libraries

A widely used class of circular topological fingerprints that encode molecular neighborhoods up to a specified diameter. ECFP4 (diameter 4) captures all substructures within a radius of two bonds from each heavy atom.

• Iteratively assigns integer identifiers to atom environments
• Invariant to atom numbering and molecular orientation
• Forms the basis for many QSAR and virtual screening workflows

A set of 166 predefined structural keys, each corresponding to a specific chemical feature such as the presence of a carboxyl group or a ring system. Unlike hashed fingerprints, each bit has an explicit, interpretable meaning.

• Developed by MDL Information Systems for substructure searching
• Fixed-length binary vector: 1 indicates feature presence
• Useful when interpretability of the feature space is required

A computational modeling approach that establishes a mathematical correlation between molecular descriptors—often fingerprints—and measured biological activity. Fingerprints serve as the fixed-length feature vectors for training regression or classification models.

• Enables prediction of activity for untested compounds
• Requires careful validation to avoid overfitting
• Governed by OECD principles for regulatory acceptance

The identification of novel chemotypes that retain biological activity while possessing a fundamentally different core molecular scaffold. Fingerprint-based similarity searches can be tuned to prioritize scaffold diversity.

• Escapes existing intellectual property constraints
• Improves ADMET profiles while maintaining potency
• Pharmacophoric fingerprints are particularly effective for this task