SPLDExtraTrees: robust machine learning approach for predicting kinase inhibitor resistance.

Area of Science:

• Computational biology
• Biochemistry
• Machine learning

Background:

• Drug resistance, driven by protein mutations, poses a significant global health threat.
• Accurately predicting the impact of mutations on drug-target interactions is crucial for drug development and clinical practice.
• Existing computational methods face challenges with limited data and overfitting for machine learning-based drug resistance prediction.

Purpose of the Study:

• To develop a robust machine learning method for predicting ligand binding affinity changes upon protein mutation.
• To identify mutations that confer drug resistance.
• To overcome limitations of sample size and noise in machine learning for drug resistance studies.

Main Methods:

• Proposed a novel machine learning method, SPLDExtraTrees.
• Implemented a specific data ranking scheme, starting with easy samples and progressing to harder, diverse ones.
• Incorporated physics-based structural features to provide domain knowledge for data-limited tasks.
• Iteratively updated sample weights and the model.

Main Results:

• SPLDExtraTrees accurately predicts ligand binding affinity changes and identifies resistance-causing mutations.
• The method demonstrated predictive capability in three kinase inhibitor resistance scenarios.
• Achieved predictive accuracy comparable to molecular dynamics and Rosetta methods with significantly reduced computational cost.

Conclusions:

• SPLDExtraTrees offers a robust and computationally efficient solution for predicting drug resistance mutations.
• The method's data-ranking strategy and inclusion of physics-based features enhance its performance in data-limited scenarios.
• This approach holds significant potential for accelerating drug development and improving clinical treatment strategies.