Decoy Generation

Decoys must share key physicochemical properties with active ligands to prevent a model from discriminating based on trivial features. Critical properties to match include:

• Molecular weight (MW) distribution
• Calculated logP (cLogP) for lipophilicity
• Hydrogen bond donors and acceptors count
• Rotatable bonds and topological polar surface area (TPSA)

Tools like the DUDE-Z benchmark enforce strict property distributions using a maximum Tanimoto distance between property histograms, ensuring the decoy set is property-matched within a 0.5 standard deviation tolerance.

A fundamental requirement is that decoys are topologically distinct from active ligands to ensure they are unlikely to bind the same target. This is enforced through:

• Tanimoto coefficient thresholds on 2D fingerprints (e.g., ECFP4, MACCS keys)
• A maximum similarity cutoff, typically Tanimoto < 0.4 for ECFP4 fingerprints
• Avoidance of shared Murcko scaffolds or core ring systems

This dissimilarity guarantees that decoys do not share the critical pharmacophoric features required for target binding, making them true negatives for validation studies.

Decoys must be experimentally verified as inactive or, at minimum, presumed inactive with high confidence. The gold standard sources include:

• ChEMBL database entries with explicit inactivity annotations (IC50/EC50 > 30 μM)
• PubChem BioAssay records showing no dose-response activity
• BindingDB entries with confirmed non-binding status

When experimental data is unavailable, decoys are selected from ligands of phylogenetically unrelated proteins with dissimilar binding pockets, a method validated by the DEKOIS and DUDE benchmark frameworks.

A robust decoy set maintains a fixed ratio of decoys per active ligand to prevent class imbalance from skewing enrichment metrics. Standard ratios include:

• 36:1 or 50:1 decoys per active in DUDE and DUDE-Z benchmarks
• 30:1 in the DEKOIS 2.0 library
• Custom ratios based on the size of the screening database

This ratio ensures that enrichment factor (EF) and ROC AUC calculations are statistically meaningful. An insufficient number of decoys can inflate early enrichment metrics, while too many can make the evaluation computationally prohibitive.

Decoy sets must be rigorously filtered to exclude compounds that cause false positive artifacts in biochemical assays. Key filters include:

• PAINS (Pan-Assay Interference Compounds) substructure filters to remove frequent hitters
• Aggregator detection using logP and critical micelle concentration thresholds
• Redox cyclers and covalent modifiers that promiscuously react with proteins
• Colloidal aggregators that non-specifically inhibit enzymes at micromolar concentrations

Tools like Faf-Drugs4, SwissADME, and the Aggregator Advisor database are used to computationally screen and eliminate these nuisance compounds from decoy libraries.

The primary retrospective metric used to evaluate decoy set quality. Enrichment Factor quantifies how many more true actives are identified in the top x% of a ranked database compared to a random selection. A well-constructed decoy set should yield an EF₁% significantly greater than 1.0, demonstrating that the scoring function can discriminate actives from property-matched decoys. Poor decoy selection—such as using random compounds instead of property-matched decoys—artificially inflates EF values and provides a false sense of model performance.

The foundational principle of decoy generation: decoys must share physicochemical property distributions with active ligands while being topologically dissimilar. This prevents virtual screening models from exploiting trivial property biases. Key matching criteria include:

• Molecular weight (±25 Da tolerance)
• Calculated logP (±1.0 unit tolerance)
• Rotatable bond count (±1 bond tolerance)
• Hydrogen bond donors and acceptors (exact match)
• Net formal charge (exact match)

Failure to property-match produces decoys that are trivially separable by simple descriptors like size or lipophilicity.

The second critical constraint in decoy generation: decoys must be topologically distinct from active ligands to avoid including latent actives. This is typically enforced using Tanimoto similarity on extended-connectivity fingerprints (ECFP4) with a maximum threshold of 0.5–0.6. Decoys with higher similarity risk being weak binders that contaminate the negative set. Advanced methods use Murcko scaffold analysis to ensure no decoy shares the same core ring system as any active, providing an additional layer of topological filtering.

A critical quality control problem where decoy sets inadvertently contain true binders due to incomplete bioactivity annotation. LADS contamination artificially depresses enrichment metrics and leads to incorrect conclusions about model performance. Detection methods include:

• Cross-referencing against ChEMBL and PubChem BioAssay databases
• Consensus docking across multiple scoring functions to flag suspicious decoys
• PAINS filtering to remove frequent-hitter substructures

Studies suggest 5–15% of decoys in some benchmark sets may be latent actives, particularly for well-studied targets like COX-2 and HIV protease.