Modeling and Interpretability Study of the Structure-Activity Relationship for Multigeneration EGFR Inhibitors.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Fourth-generation EGFR inhibitors are primarily in clinical trials, highlighting the need for new drug development.
• EGFR mutations, including L858R, T790M, and C797S, confer resistance to existing therapies.
• Developing effective inhibitors against these resistant mutations is a significant challenge in cancer treatment.

Purpose of the Study:

• To establish a predictive structure-activity relationship (SAR) model for multigeneration EGFR inhibitors.
• To compare the performance of a multitask deep neural network (MT-DNN) against single-task models.
• To leverage interpretability analysis for understanding structural determinants of EGFR inhibitor activity.

Main Methods:

• Collected a dataset of 2302 multitarget EGFR inhibitors against wild-type and mutated EGFR.
• Developed a multitask deep neural network (MT-DNN) for predicting bioactivities.
• Constructed single-task models (SVM, RF, XGBoost, ST-DNN) for comparison.
• Utilized SHAP/delta-SHAP value analysis for model interpretability.

Main Results:

• The MT-DNN model significantly outperformed all single-task models on an external validation set of 304 fourth-generation EGFR inhibitors.
• The MT-DNN model demonstrated superior predictive accuracy for inhibitors targeting various EGFR mutations.
• SHAP/delta-SHAP analysis successfully identified core scaffolds and key fragments of effective EGFR inhibitors.

Conclusions:

• The MT-DNN approach provides a robust platform for predicting the bioactivity of multigeneration EGFR inhibitors.
• This study offers valuable structural insights to guide the design of novel EGFR inhibitors overcoming resistance mutations.
• The integration of MT-DNN with SHAP/delta-SHAP analysis is a powerful strategy for drug discovery in EGFR-targeted therapy.