Curated and Structure-Based Drug-Target Interactions Improve Underprediction of Drug Side Effects in Network Models.

Area of Science:

• Pharmacology and Drug Development
• Computational Biology
• Toxicology

Background:

• Accurate prediction of drug-induced side effects is a major hurdle in pharmaceutical development, often leading to late-stage failures.
• Traditional methods like animal testing and in vitro assays have limitations in cost, ethics, and human translatability.
• Existing computational models for predicting adverse drug effects, such as protein-protein interaction networks, often suffer from underprediction and data inconsistency.

Purpose of the Study:

• To evaluate the impact of integrating drug-binding targets from multiple sources into the PathFX platform for improved drug side effect prediction.
• To analyze how different data integration strategies affect the sensitivity and specificity of predicting adverse drug reactions.
• To provide a foundation for enhancing machine learning approaches in drug safety assessment.

Main Methods:

• Integrated drug-binding targets from six distinct databases (DrugBank, ChEMBL, PubChem, STITCH, TTD, PocketFEATURE) into the PathFX platform.
• Analyzed unique drug-target interactions, protein classes, and functions across the integrated data sources.
• Quantitatively assessed the trade-offs between prediction sensitivity and specificity based on data source characteristics.

Main Results:

• Integration of new drug targets led to the prediction of previously unrecognized side effects.
• A clear trade-off between sensitivity and specificity was observed; larger, exploratory databases improved sensitivity at the cost of specificity.
• Smaller, curated, or structurally predicted target databases enhanced specificity, suitable for explainable platforms like PathFX.

Conclusions:

• Integrating diverse drug-target data sources significantly impacts drug side effect prediction accuracy.
• The choice of data source and integration strategy is critical for balancing sensitivity and specificity in predictive models.
• This work supports the development of sophisticated machine learning models for large-scale data and simpler, explainable platforms for hypothesis generation in drug safety.