Integrative machine learning predicts activating kinase mutations for precision oncology.

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology and Bioinformatics
• Genomics and Genetics

Background:

• Protein kinases are vital enzymes regulating cellular processes; aberrant activation by mutations drives cancer progression.
• Understanding kinase missense mutations is crucial for personalized cancer therapy and predicting drug efficacy.
• Current methods struggle to accurately predict the functional impact of kinase mutations.

Purpose of the Study:

• To develop an accurate machine learning framework, Kinome-AI, for classifying kinase missense mutations as activating or non-activating.
• To integrate multi-modal data, including sequence and structural features, for enhanced predictive power.
• To enable precise identification of cancer-driving kinase mutations for targeted therapeutic strategies.

Main Methods:

• Developed Kinome-AI, an integrative machine learning framework utilizing multi-modal features.
• Incorporated residue-level biochemical changes, protein language model sequence embeddings, and molecular modeling structural descriptors.
• Employed a teacher-student learning strategy to impute missing structural data, leveraging available structural information.

Main Results:

• Kinome-AI achieved an AUROC of 0.85 and BACC of 0.76 across 1,003 mutations in 110 kinases.
• The model significantly outperformed existing bioinformatics and general-purpose variant effect predictors.
• The imputation strategy improved performance without necessitating structural inputs for novel mutations.

Conclusions:

• Kinome-AI provides a robust method for predicting kinase mutation activation status.
• This framework quantifies sequence-structure-function relationships in cancer-related kinase mutations.
• Kinome-AI holds promise for advancing personalized cancer treatment by informing targeted therapy decisions.