Combining machine learning and pharmacophore-based interaction fingerprint for in silico screening

Area of Science:

• Computational chemistry
• cheminformatics
• Machine learning in drug discovery

Background:

• Traditional in silico screening methods often rely on similarity-based ranking or scoring functions.
• Developing more efficient and accurate virtual screening techniques is crucial for accelerating drug discovery.

Purpose of the Study:

• To introduce and evaluate a novel pharmacophore-based interaction fingerprint (Pharm-IF) for virtual screening.
• To compare the performance of Pharm-IF models against established methods like GLIDE score and PLIF using machine learning algorithms.

Main Methods:

• Development of the pharmacophore-based interaction fingerprint (Pharm-IF).
• Application of machine learning algorithms, including Support Vector Machine (SVM) and Random Forest (RF), for virtual screening.
• Validation using docking results from multiple protein targets (PKA, SRC, cathepsin K, carbonic anhydrase II, HIV-1 protease).
• Comparison of screening efficiencies based on enrichment factors at 10%.

Main Results:

• The combination of SVM and Pharm-IF achieved an average enrichment factor of 5.7 at 10%, outperforming GLIDE score (4.2) and PLIF (4.3).
• Machine learning models using Pharm-IF demonstrated stable and superior performance compared to GLIDE score when trained on more than five crystal structures.
• Utilizing docking poses of known active compounds as positive training samples significantly boosted the enrichment factors for RF models.

Conclusions:

• Pharm-IF, when integrated with machine learning, offers a powerful and effective approach for in silico screening.
• The Pharm-IF method shows significant potential to enhance the accuracy and efficiency of virtual screening in drug discovery pipelines.
• Incorporating diverse training data, such as docking poses, further improves the predictive power of machine learning-based screening models.