Exploring conformational dynamics of the HER2 DFG-flip using machine-learning-guided metadynamics for type II...

Area of Science:

• Computational Biology and Biochemistry
• Structural Biology
• Drug Discovery

Background:

• Protein kinases, like Human Epidermal Growth Factor Receptor 2 (HER2), are crucial for cell regulation, and their malfunction drives cancer.
• Current HER2 drug development primarily targets the active DFG-in conformation, overlooking the inactive DFG-out state due to a lack of structural data.
• Developing inhibitors for the inactive HER2 state is hindered by the absence of experimentally resolved structures.

Purpose of the Study:

• To characterize the conformational landscape of HER2, specifically the transition between the DFG-in and DFG-out states.
• To generate a reliable structural model of the inactive DFG-out conformation of HER2 for structure-based drug design.
• To explore the utility of deep learning in sampling complex conformational transitions in kinases.

Main Methods:

• Employed molecular dynamics (MD) simulations combined with machine-learning-guided well-tempered metadynamics (MetaD) to sample HER2 conformations.
• Utilized Deep Targeted Discriminant Analysis (DeepTDA) to construct a deep learning-based collective variable for efficient sampling of the DFG-flip pathway.
• Performed covalent docking and MD simulations to assess the ligand-binding reliability of the predicted DFG-out structure.

Main Results:

• Successfully sampled the DFG-in to DFG-out transition, revealing the inactive DFG-out conformation is energetically favored by ~25 kJ/mol.
• Identified an alternative DFG-in conformer and a DFG-up intermediate conformer, offering insights into HER2's basal activity and kinase dynamics.
• Validated the predicted DFG-out structure through docking, showing comparable binding affinity to experimental data.

Conclusions:

• Developed a structurally grounded model of inactive HER2, providing a mechanistic framework for designing type II inhibitors targeting this conformation.
• Demonstrated the effectiveness of deep learning-based collective variables (CVs) in efficiently characterizing complex free energy landscapes of kinases.
• The findings facilitate structure-based design of novel HER2 inhibitors targeting inactive conformations, potentially overcoming resistance mechanisms.