Toward more realistic drug-target interaction predictions.

Area of Science:

• Pharmacology
• Computational Biology
• Machine Learning

Background:

• Supervised machine learning models are increasingly used for predicting drug-target interactions (DTI) using chemical and genomic data.
• Current DTI prediction models are often evaluated under simplified conditions that do not reflect real-world applications.
• Network pharmacology applications, such as drug repositioning, could benefit from more robust DTI prediction methods.

Purpose of the Study:

• To investigate the impact of evaluation settings on the performance of supervised machine learning models for DTI prediction.
• To identify factors that contribute to overoptimistic DTI prediction results.
• To propose guidelines for more realistic DTI prediction model formulation and evaluation.

Main Methods:

• Utilized quantitative drug-target bioactivity assays for kinase inhibitors.
• Employed a popular benchmarking dataset for binary drug-target interactions across various target classes (enzyme, ion channel, nuclear receptor, GPCR).
• Analyzed the effects of four key factors: problem formulation (classification vs. regression), evaluation dataset selection, cross-validation procedure, and experimental setting (shared vs. distinct drugs/targets).

Main Results:

• Problem formulation significantly impacts prediction outcomes, with regression offering a more realistic approach than binary classification.
• The choice of evaluation dataset, specifically drug and target families, critically influences prediction performance.
• Evaluation procedures (simple vs. nested cross-validation) and experimental settings (common or distinct drugs/targets) lead to substantial differences in reported prediction results.

Conclusions:

• Current DTI prediction models may report overoptimistic results due to simplified evaluation settings.
• Realistic DTI prediction requires careful consideration of problem formulation, dataset relevance, and rigorous evaluation procedures.
• Development of novel benchmarking datasets capturing continuous drug-target interaction data is needed for more accurate kinase inhibitor studies.