Interaction Fingerprint

An interaction fingerprint encodes protein-ligand contacts or interaction types as a vector. It is useful for comparing poses, clustering docking results, summarizing binding modes, and building interpretable features.

For residue or atom interaction indicators:

$$z_k=1[\text{interaction }k\text{ is present}]$$

The fingerprint is:

$$z=(z_1,\dots ,z_K)$$

Similarity between two interaction fingerprints can be computed with overlap or Tanimoto-style measures:

$$T(z,z^{\prime })=\frac{z\cdot z^{\prime }}{\Vert z{\Vert }_1+\Vert z^{\prime }{\Vert }_1-z\cdot z^{\prime }}$$

Key Ideas

• Interaction fingerprints summarize contacts, not full geometry.
• The feature definition depends on distance cutoffs, atom typing, residue selection, and interaction rules.
• Similar fingerprints can hide different poses if geometry is coarse.
• They are useful diagnostic features but should not replace pose quality checks.

Feature Contract

An interaction fingerprint should state:

Field	Question
receptor unit	residue, atom, pharmacophore point, pocket region, or interaction site
ligand unit	atom, fragment, pharmacophore feature, or whole-ligand contact
interaction types	hydrophobic, H-bond, ionic, aromatic, metal, water-mediated, clash
distance/angle rules	cutoffs, angular constraints, donor/acceptor typing
chemical state	hydrogens, protonation, tautomer, charge, metals, cofactors
aggregation	binary, count, weighted, residue-level, atom-level, or graph feature

Changing any of these fields changes the feature space:

$$z=\mathrm{IFP}(P,L,\hat{X};\rho )$$

where $\rho$ is the rule set.

Comparison Boundary

The same fingerprint can be used for different claims:

Use	Metric	Risk
compare poses for one complex	Tanimoto or overlap	coarse contacts may hide wrong geometry
cluster docking poses	fingerprint distance	clusters depend on interaction rules
screen molecules	similarity to reference interaction pattern	reference pose or ligand can leak target knowledge
model input	binary/count features	hand-crafted features may dominate learned model
interpret prediction	highlighted contacts	contact does not prove causal binding contribution

An interaction fingerprint similarity is not a replacement for pose RMSD, clash checks, affinity, or enrichment.

Leakage Risks

• Using a known native ligand interaction pattern to filter test poses can leak the answer.
• Residue selection based on the bound ligand can make the task easier than deployment.
• Protonation or hydrogen placement rules can create inconsistent H-bond features across splits.
• Comparing across targets without residue mapping can produce misleading similarity.

Practical Checks

• Which interaction types are encoded?
• Are cutoffs and atom typing rules documented?
• Are waters, metals, cofactors, and protonation states included or ignored?
• Does the fingerprint compare poses within one target or across targets?
• Is it used for interpretation, clustering, filtering, or model input?
• Is the reference interaction pattern available at inference time?
• Are geometry checks reported alongside fingerprint similarity?

Related

• Protein-ligand complex
• Protein-ligand interaction
• Pose quality
• Pose RMSD
• Scoring function
• Molecular fingerprint
• Binding site