PPB

PPB predicts protein targets for small molecules by combining ten molecular fingerprint types with ChEMBL bioactivity data to assess structural similarity and produce significance-ranked target lists.

Key Features:

• Multi-fingerprint similarity: Uses ten distinct fingerprint types covering composition, substructures, molecular shape, and pharmacophores to assess structural similarity.
• ChEMBL-derived database: Searches a dataset of 4,613 groups of bioactive molecules, each annotated with at least 10 identical targets from ChEMBL.
• Consensus voting: Aggregates predictions across fingerprints using a consensus voting scheme to produce a ranked target list.
• Statistical ranking: Reports statistical significance for predicted targets using p-values.
• Validation set: Validated on a set of 670 drugs with up to 20 known targets per drug.
• Performance metrics: Integration across all ten fingerprints identified approximately 50% of a drug's known targets on average and reported an overall hit rate of 25%.
• Case study — TRPV6 inhibitor: Profiled a novel TRPV6 calcium channel inhibitor against 24 safety screen targets, correctly predicting inhibition for all five predicted hits and observing additional activity in seven of the 18 non-predicted targets.
• Benchmarking: Achieved a correct prediction rate of 5/12 and an incorrect prediction rate of 0/12 for the TRPV6 inhibitor, comparable to other web-based tools.

Scientific Applications:

• Target prediction: Prediction of putative protein targets for small molecules using multi-fingerprint similarity to ChEMBL bioactivity annotations.
• Polypharmacology profiling: Profiling of compounds for multi-target (polypharmacological) activity across annotated target groups.
• Safety panel screening: Prioritization and interpretation of potential off-targets in safety screen panels, as demonstrated for a TRPV6 inhibitor against 24 targets.
• Method validation and benchmarking: Comparative assessment of prediction performance using validation sets of drugs with known target annotations.

Methodology:

Compute structural similarity using ten molecular fingerprints against a ChEMBL-derived set of 4,613 bioactive molecule groups, aggregate predictions via a consensus voting scheme, and rank predicted targets by p-value.