Prioritizing Virtual Screening with Interpretable Interaction Fingerprints.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Drug discovery relies heavily on effective molecular representations for machine learning models.
• Traditional molecular fingerprints lack protein interaction data and structural context, limiting model interpretability.

Purpose of the Study:

• To introduce LUNA, a Python toolkit for generating novel hashed fingerprints encoding protein-ligand interactions.
• To enhance machine learning model interpretability and performance in drug discovery through new interaction fingerprints.

Main Methods:

• Developed LUNA toolkit to compute Extended Interaction FingerPrint (EIFP), Functional Interaction FingerPrint (FIFP), and Hybrid Interaction FingerPrint (HIFP).
• Applied fingerprints to machine learning models for predicting DOCK3.7 scores using 1 million Dopamine D4 complexes.
• Evaluated fingerprint performance against traditional molecular and interaction fingerprints.

Main Results:

• EIFP-4,096 demonstrated superior performance (R² = 0.61) in predicting DOCK3.7 scores compared to other fingerprints.
• LUNA facilitated the development of interpretable machine learning models.
• Interaction fingerprints identified molecular complex similarities missed by conventional methods.

Conclusions:

• LUNA and its interaction fingerprints offer a promising approach for machine learning-based virtual screening.
• The toolkit enhances interpretability and predictive power in computational drug discovery.
• LUNA is freely available, promoting wider adoption in the research community.