Kinome-wide activity modeling from diverse public high-quality data sets

Area of Science:

• Biochemistry
• Pharmacology
• Computational Biology

Background:

• Large volumes of kinase small molecule inhibitor data exist in public literature.
• Data is heterogeneous due to varied methodologies and experimental procedures across laboratories.

Purpose of the Study:

• To assess the applicability of heterogeneous public datasets for predicting kinase activities.
• To identify data characteristics that contribute to predictive utility.

Main Methods:

• Accessed ~500,000 molecules from the Kinase Knowledge Base (KKB).
• Generated over 180 distinct datasets covering the human kinome after aggregation and standardization.
• Developed and cross-validated hundreds of classification and regression models.

Main Results:

• Highly predictive classification and quantitative models were generated for most kinase targets.
• Model accuracy was dependent on a minimum number of active compounds or structure-activity data points.
• Predicted activities showed good agreement with experimental profiling from the NIH LINCS program.

Conclusions:

• Heterogeneous public kinase inhibitor datasets are highly valuable for developing accurate predictors.
• These datasets are well-suited for Kinome-wide virtual screening applications.
• Public data can effectively complement experimental kinase profiling.