An attention mechanism-based LSTM network for cancer kinase activity prediction

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Kinase drug resistance remains a significant challenge in cancer therapy.
• Predictive models for compound bioactivity against cancer kinases are highly desirable.

Purpose of the Study:

• To develop a Long Short-Term Memory (LSTM) framework for predicting lead molecule activities against seven cancer kinases.
• To integrate an attention mechanism for interpreting chemical feature contributions to kinase activity.

Main Methods:

• Utilized 14,907 compounds from the ChEMBL database for model training.
• Employed two molecular representations: 2D descriptors and MACCS fingerprints with the LSTM method.
• Integrated an attention mechanism to identify significant chemical moieties influencing kinase activity.

Main Results:

• Achieved test set accuracies of 0.81 for 2D descriptors and 0.78 for MACCS fingerprints.
• Reported Receiver Operating Characteristic (ROC)-Area Under the Curve (AUC) scores ranging from 0.8 to 0.99 for both models.
• Demonstrated effective extraction of significant chemical moieties by the attention mechanism.

Conclusions:

• The proposed LSTM framework shows promise for predicting cancer kinase activities.
• This computational approach can serve as a foundation for developing novel anti-cancer kinase drugs.
• The integrated attention mechanism enhances the interpretability of the predictive model.