Target Fishing

Unlike conventional docking that screens many ligands against one target, inverse docking docks a single query ligand into the binding pockets of thousands of protein structures. The algorithm systematically samples the ligand's conformational space within each pre-identified cavity, generating a ranked list of potential macromolecular targets based on computed binding scores. This reverse screening approach requires specialized scoring functions calibrated for cross-target comparability rather than single-target enrichment.

Effective target fishing demands comprehensive protein structure databases such as the Protein Data Bank (PDB) and AlphaFold-predicted structures. The screening panel must represent diverse protein families—GPCRs, kinases, nuclear receptors, ion channels, and enzymes—to capture the full polypharmacological landscape. Coverage gaps in the target database directly limit the method's sensitivity for detecting novel off-targets.

An alternative to structure-based inverse docking, ligand-based target fishing uses 3D pharmacophore models derived from known ligand-receptor interactions. The query molecule's pharmacophoric features—hydrogen bond donors/acceptors, hydrophobic centroids, aromatic rings—are matched against a library of target-specific pharmacophore models. This approach excels when high-quality protein structures are unavailable but requires extensive curated pharmacophore databases.

SEA relates proteins based on the chemical similarity of their known ligands. The method computes a statistical expectation value (E-value) comparing the query molecule's similarity to each target's ligand set against a random background model. Key advantages include:

• No protein structure required
• Captures cryptic off-targets through ligand set comparison
• Scalable to thousands of targets simultaneously
• Validated across ChEMBL and DrugBank databases

Modern target fishing increasingly employs proteochemometric models that learn joint representations of ligand chemical space and protein sequence space. Deep neural networks trained on interaction matrices (e.g., BindingDB, DrugBank) can predict novel compound-target pairs without explicit docking. Graph neural networks operating on molecular graphs and protein contact maps have demonstrated superior generalization to unseen targets compared to traditional similarity-based methods.

The primary output of target fishing is a ranked list of potential targets annotated with confidence scores, binding mode predictions, and known pharmacological implications. This profile enables researchers to:

• Explain observed phenotypic effects through polypharmacology
• Anticipate toxicity mechanisms before in vivo testing
• Identify repurposing opportunities for existing drugs
• Prioritize targets for experimental validation via SPR or thermal shift assays

The study of a single drug molecule interacting with multiple distinct targets. Target fishing is the primary computational tool to uncover this polypharmacological profile. While traditional drug discovery sought absolute specificity, modern systems biology recognizes that a drug's therapeutic efficacy and toxicity are often the result of a multi-target mechanism of action. Understanding this network of interactions is critical for predicting off-target side effects and identifying opportunities for drug repurposing.

A machine learning paradigm that simultaneously models the ligand space and the target space to predict bioactivity across a large interaction matrix. Unlike target fishing, which queries one ligand against many targets, PCM models are trained on known drug-target pairs to predict missing interactions. They use descriptors from both the drug's chemical structure and the protein's sequence, enabling predictions for novel ligands on novel targets—a key advantage over traditional QSAR.

A critical data preparation step for training machine learning models in drug-target interaction prediction. A model trained only on known binding pairs will overfit and predict everything as a binder. Negative sampling constructs a set of presumed non-interacting drug-target pairs to teach the classifier the difference. Strategies include:

• Random pairing: Assumes most random pairs do not bind.
• Hard negative mining: Selecting pairs with high similarity to known binders but no reported activity.
• Decoy-based: Using molecules with similar physical properties to actives but different topology.

A prerequisite step for structure-based target fishing. Before docking a ligand to a protein, the concave, solvent-accessible cavities on the protein surface must be identified. Algorithms like Fpocket and DeepSite use geometric criteria or deep learning to locate these pockets. In a target fishing campaign, a single ligand is docked into the detected pockets of thousands of proteins in a database, making accurate pocket detection essential to avoid false negatives.

A retrospective performance metric used to validate a target fishing protocol. It quantifies how many more known true targets are identified in the top-ranked fraction of a screened database compared to a random selection. For example, an EF₁% of 20 means that the top 1% of ranked targets contains 20 times more true binders than expected by chance. This metric is calculated using curated benchmarking datasets where the true targets of a ligand are already known.