TocoDecoy: A New Approach to Design Unbiased Datasets for Training and Benchmarking Machine-Learning Scoring...

Area of Science:

• Computational Chemistry
• Machine Learning in Drug Discovery
• Bioinformatics

Background:

• Accurate machine-learning-based scoring functions (MLSFs) are vital for structure-based virtual screening.
• Existing datasets for MLSF development often lack structural diversity and may contain hidden biases.
• This limits the reliability and generalizability of trained MLSFs.

Purpose of the Study:

• To develop a novel decoy generation method, TocoDecoy, for creating unbiased and expandable datasets.
• To generate datasets suitable for training and benchmarking MLSFs.
• To evaluate the impact of decoy dataset bias on MLSF performance.

Main Methods:

• Developed TocoDecoy, integrating topology-based and conformation-based strategies to generate decoys by modifying active molecules.
• Generated datasets using TocoDecoy for training and benchmarking machine-learning models.
• Assessed the performance of InteractionGraphNet (IGN) trained on TocoDecoy, LIT-PCBA, and DUD-E-like datasets.

Main Results:

• The IGN model trained on the TocoDecoy dataset demonstrated competitive performance compared to the LIT-PCBA dataset.
• IGN trained on TocoDecoy significantly outperformed the model trained on the DUD-E dataset.
• These findings suggest that TocoDecoy generates unbiased decoys suitable for MLSF development.

Conclusions:

• TocoDecoy provides an effective strategy for generating unbiased and expandable decoy datasets.
• The developed datasets are suitable for training and benchmarking MLSFs.
• This approach helps mitigate hidden biases in datasets, leading to more reliable MLSF models.