Large-Scale Modeling of Sparse Protein Kinase Activity Data.

Area of Science:

• Biochemistry and Cheminformatics
• Computational Biology
• Drug Discovery

Background:

• Protein kinases are crucial in complex diseases like cancer, with conserved ATP binding sites enabling multitarget drug development.
• Balancing drug efficacy and minimizing toxicity requires understanding kinase inhibitor selectivity.
• Publicly available protein kinase activity data offers potential for advanced computational modeling.

Purpose of the Study:

• To construct a robust benchmark dataset for protein kinase activity prediction models.
• To address challenges in multitask learning with sparse data, specifically data leakage and missing values.
• To evaluate the performance of multitask deep learning models against other machine learning approaches.

Main Methods:

• Development of a protein kinase benchmark dataset with two balanced splits using random and dissimilarity-driven cluster-based sampling.
• Implementation and comparison of multitask deep learning, single-task deep learning, and tree-based models.
• Assessment of model performance and generalizability on the constructed benchmark set.
• Evaluation of the impact of data imputation on model performance.

Main Results:

• The dissimilarity-driven cluster-based split demonstrated lower model performance than the random split, indicating challenges in generalizability.
• Multitask deep learning models outperformed single-task deep learning and tree-based models on the sparse benchmark dataset.
• Data imputation strategies did not yield performance improvements for either single-task or multitask models.

Conclusions:

• The developed benchmark dataset is valuable for evaluating and advancing protein kinase activity prediction models.
• Multitask deep learning models show potential for effectively modeling sparse protein kinase activity data.
• Careful consideration of dataset splitting and model generalizability is essential for reliable kinase inhibitor drug discovery.