Machine learning on ligand-residue interaction profiles to significantly improve binding affinity prediction

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Structure-based virtual screening (SBVS) is crucial for drug discovery but faces challenges in accurately predicting binding affinity and prioritizing ligands.
• Current methods often struggle with precise scoring and ranking of potential drug candidates.

Purpose of the Study:

• To develop and evaluate a novel machine learning-based prediction model, termed an IP scoring function (IP-SF), to enhance SBVS performance.
• To systematically investigate improvements for IP-SFs by exploring sampling methods and machine learning algorithms.

Main Methods:

• Developed IP scoring functions (IP-SFs) using ligand-residue interaction profiles (IPs) and machine learning (ML) algorithms.
• Investigated various sampling protocols (e.g., MIN+GB) and ML algorithms (e.g., GBDT).
• Critically evaluated IP-SF performance against Glide SF using six drug targets with known ligands.

Main Results:

• The GBDT algorithm combined with the MIN+GB protocol achieved the best performance for IP-SFs.
• IP-SFs significantly outperformed Glide SF in scoring, ranking, and screening power.
• Observed average reductions of 38% in MAE and 36% in RMSE; average increases of 225% in SCC and 73% in PI.
• Achieved a superior average AUC of 0.87 compared to Glide's 0.71 for receiver operating characteristic curves.

Conclusions:

• IP-SFs, particularly those using GBDT/MIN+GB, offer a significant advancement in SBVS.
• The developed method shows promising potential for broad applications in accelerating drug discovery pipelines.
• This approach effectively addresses the challenge of accurate binding affinity prediction and ligand prioritization in virtual screening.