Multiple machine learning methods aided virtual screening of Na_V 1.5 inhibitors.

Area of Science:

• Cardiovascular Science
• Computational Chemistry
• Pharmacology

Background:

• Nav1.5 sodium channels are crucial for cardiac action potential and myocardial excitability.
• The patch clamp method, while standard for inhibitor screening, is technically demanding, costly, and slow.
• Developing faster, more efficient screening methods is essential for drug discovery.

Purpose of the Study:

• To develop novel machine learning (ML) models for screening chemical blockers of Nav1.5 sodium channels.
• To identify privileged substructures associated with Nav1.5 inhibition.
• To overcome the limitations of conventional patch clamp techniques.

Main Methods:

• Utilized data from the ChEMBL Database to build 30 classification models using six molecular fingerprints and five ML algorithms.
• Employed validation and test sets to evaluate model performance.
• Extracted privileged substructures using the bioalerts Python package.

Main Results:

• The RF-Graph ML model demonstrated superior performance, achieving a Prediction Accuracy (Q) of 0.9309 and a Matthew's correlation coefficient of 0.8627 on the test set.
• Identified sulfa structures and sterically hindered fragments as key substructures that tend to block Nav1.5.
• Unsupervised learning highlighted MACCS and Graph fingerprints for identifying sulfa drugs.

Conclusions:

• Effective ML models were successfully developed for screening potential Nav1.5 inhibitors.
• Key privileged substructures with high affinity for Nav1.5 inhibition were identified.
• This ML-driven approach offers a more efficient alternative to traditional screening methods for ion channel modulators.