Machine Learning Scoring Functions for Drug Discovery from Experimental and Computer-Generated Protein-Ligand...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Machine learning (ML) is increasingly used for developing accurate scoring functions in computational drug discovery.
• Concerns exist regarding over-optimistic ML results due to database correlations in training and testing sets.
• Investigating ML performance with diverse datasets is crucial for reliable drug discovery tools.

Purpose of the Study:

• To evaluate the performance of an artificial neural network (ANN) for predicting protein-ligand binding affinity.
• To compare ANN performance using experimental versus computer-generated protein-ligand structures.
• To explore the potential of per-target scoring functions trained on computer-generated data.

Main Methods:

• An artificial neural network was trained and tested for binding affinity prediction.
• Two distinct datasets were used: experimental protein-ligand structures and computer-generated structures.
• Performance was assessed using both random (horizontal) and target-specific (vertical) cross-validation tests.

Main Results:

• The ANN demonstrated similar performance regardless of whether experimental or computer-generated structures were used for training and testing.
• A significant performance decrease was observed in vertical tests (unseen target proteins) compared to horizontal tests.
• Training ANNs on computer-generated databases enabled the development of effective per-target scoring functions.

Conclusions:

• Computer-generated structural databases can be effectively utilized for training machine learning models in drug discovery.
• Per-target scoring functions, trained on specialized datasets, show promising results for specific protein targets.
• This approach offers a viable strategy to mitigate database correlation issues and improve the reliability of computational drug discovery.