Integrating Molecular Dynamics and Machine Learning Algorithms to Predict the Functional Profile of Kinase Ligands

Area of Science:

• Chemical biology and medicinal chemistry
• Computational drug discovery
• Protein-ligand interactions

Background:

• Traditional drug development often targets active sites.
• Allosteric modulation offers a way to fine-tune protein activity.
• Predicting ligand binding modes is crucial for lead discovery.

Purpose of the Study:

• To develop machine learning models for distinguishing orthosteric and allosteric binders.
• To accelerate the identification and selection of drug candidates.
• To predict the functional impact of novel ligands.

Main Methods:

• Utilizing machine learning classifiers to differentiate ligand binding modes.
• Integrating ligand chemical fingerprints with protein dynamics data from molecular dynamics (MD) simulations.
• Training and testing various ML architectures on cyclin-dependent kinases (CDKs) and other kinases.

Main Results:

• Machine learning models successfully classified orthosteric versus allosteric ligands.
• Models demonstrated applicability beyond the initial training dataset, including FDA-approved drugs.
• Integrating dynamic cross-talk information improved model performance and applicability.

Conclusions:

• Machine learning approaches can effectively predict ligand binding modes.
• Incorporating protein dynamics enhances the predictive power of these models.
• This strategy aids in the rational design and discovery of small molecule modulators.