iScore: A ML-Based Scoring Function for De Novo Drug Discovery.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in bioinformatics

Background:

• Developing efficient scoring functions is crucial for accelerating de novo drug discovery.
• Current methods often rely on explicit protein-ligand interactions and atomic contacts, which can be slow and computationally expensive.
• The vastness of chemical space necessitates rapid screening of molecular libraries.

Purpose of the Study:

• To introduce iScore, a novel machine learning-based scoring function for predicting protein-ligand binding affinity.
• To develop a fast and accurate method that bypasses traditional conformational sampling.
• To enable rapid screening of ultrahuge molecular libraries for drug discovery.

Main Methods:

• iScore utilizes ligand and binding pocket descriptors to directly predict binding affinity, avoiding explicit interaction knowledge.
• Three machine learning models (Deep Neural Network, Random Forest, eXtreme Gradient Boosting) were trained and validated.
• A hybrid model (iScore-Hybrid) was developed by combining the strengths of individual models.

Main Results:

• The iScore-Hybrid model achieved a Pearson correlation coefficient (R) of 0.78 and RMSE of 1.23 in cross-validation.
• iScore-Hybrid set new benchmarks for scoring power (R = 0.814), ranking power (ρ = 0.705), and screening power (73.7% success rate).
• The model demonstrated strong performance in target fishing benchmarking studies.

Conclusions:

• iScore offers a significant advancement in computational drug discovery by enabling rapid and accurate binding affinity prediction.
• The approach circumvents slow conformational sampling, facilitating efficient screening of large chemical libraries.
• iScore-Hybrid establishes new performance standards for scoring functions in drug discovery research.