Leveraging multiple data types for improved compound-kinase bioactivity prediction.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Machine learning applications in pharmacology

Background:

• Machine learning (ML) models are crucial for predicting compound-kinase interactions.
• Existing ML models often neglect valuable information from single-dose bioactivity data, relying solely on dose-response data.
• This limitation hinders the comprehensive utilization of available experimental results.

Purpose of the Study:

• To develop and validate a novel ML methodology for compound-kinase activity prediction that integrates both single-dose and dose-response bioactivity data.
• To enhance the accuracy and efficiency of predicting compound-kinase interactions.
• To improve the cost-effectiveness of generating training datasets for predictive models.

Main Methods:

• A two-stage machine learning approach was developed to leverage diverse bioactivity data types.
• The methodology was evaluated across five different machine learning algorithms.
• Experimental validation was performed on 347 selected compound-kinase pairs using the best-performing model.

Main Results:

• The proposed two-stage approach significantly improved model performance compared to models trained exclusively on dose-response data.
• The best performing model achieved a 40% hit rate and a 78% negative predictive value in experimental profiling.
• Incorporating model uncertainty estimates further enhanced prediction rates in compound selection.

Conclusions:

• Integrating multiple bioactivity data types, including single-dose measurements, leads to more accurate compound-kinase activity predictions.
• The developed ML methodology offers a more efficient and cost-effective strategy for building training datasets.
• This approach holds significant promise for accelerating drug discovery and development pipelines.