Target prediction utilising negative bioactivity data covering large chemical space

Area of Science:

• Computational chemistry
• Cheminformatics
• Pharmacology

Background:

• In silico analyses are crucial for drug discovery but often underutilize inactive compound data.
• Chemogenomic repositories contain vast amounts of bioactivity data, including inactive compounds, which can enhance predictive models.
• Target prediction for orphan compounds is challenging and can benefit from comprehensive data integration.

Purpose of the Study:

• To integrate large-scale inactive bioactivity data into in silico target prediction models.
• To develop a novel method for predicting the probability of activity and inactivity for a range of targets.
• To improve the prediction of drug targets for orphan compounds using comprehensive datasets.

Main Methods:

• Constructed a novel human bioactivity dataset from ChEMBL and PubChem, incorporating over 195 million data points.
• Applied a sphere-exclusion algorithm to oversample presumed inactive compounds.
• Trained and evaluated a Bernoulli Naïve Bayes algorithm using fivefold cross-validation and external datasets (WOMBAT).

Main Results:

• The model achieved high precision (99.7%) and recall (99.6%) for inactive compounds and good performance for active compounds (67.7% recall, 63.8% precision).
• Model performance was influenced by training data similarity, class size, and oversampling.
• External validation using WOMBAT data yielded improved precision-recall AUC (0.56) and BEDROC (0.85) scores compared to models trained solely on active data (0.45 AUC, 0.76 BEDROC).

Conclusions:

• Incorporating inactive data into in silico models significantly enhances target prediction accuracy, improving AUC and early recognition capabilities.
• Model performance varied based on internal and external validation, highlighting the importance of robust validation strategies.
• The developed target prediction protocol (PIDGIN) is publicly available, facilitating its use in drug discovery research.