DeepKinome: Quantitative Prediction of Kinase Binding Affinity by a Compound Using Deep Learning Based Regression Model

Yeeun Lee¹Jisu Eun^2,3Jinhyuk Lee^2,3*Seungyoon Nam^1,4*

• ¹Department of Genome Medicine and Science, Gachon University Gil Medical Center, Incheon, Republic of Korea
• ²Bio-design Editing Research Center, Korea Research Institute of Bioscience & Biotechnology, Yuseong-gu, Republic of Korea
• ³Department of Bioinformatics, KRIBB School of Bioscience, University of Science and Technology (UST), Daejeon, Republic of Korea
• ⁴Department of Health Sciences and Technology, gachon advanced institute for Health Sciences and Technology (GAIHST), Incheon, Republic of Korea

Abstract Introduction Kinases are essential for cellular regulation and drug development. Predicting the quantitative binding affinity between small-molecule compounds and kinases remains a challenge because of data complexity. Method We developed DeepKinome, a 20-layer convolutional neural network-based deep learning (DL) regression model, to predict quantitative binding affinity. Given the continuous nature of binding affinity, the Pearson's correlation coefficient (PCC) between actual and predicted values, the coefficient of determination (R2), the acceptance interval ratio (AIR), and the root mean square error (RMSE) were evaluated. Trained using data from 234 kinases and 163 compounds from the L1000 database. Results DeepKinome outperformed five DL and four machine learning models, achieving a PCC of 0.743, an R2 of 0.535, an AIR of 0.570, and an RMSE of 1.157. Explainable artificial intelligence analysis revealed key amino acid sequences that influenced the predictions aligned with known kinase phosphorylation sites. Conclusion DeepKinome offers a promising approach for understanding kinase inhibition and compound binding.