Selection of Informative Examples in Chemogenomic Datasets.

Area of Science:

• Chemogenomics
• Cheminformatics
• Machine Learning in Pharmacology

Background:

• Large chemogenomic datasets are crucial for developing ligand-target interaction models in drug discovery.
• Current machine learning approaches often prioritize complex models, sacrificing training efficiency and interpretability.

Purpose of the Study:

• To investigate the efficacy of automated adaptive selection for training chemogenomic models.
• To demonstrate that smaller, curated datasets can achieve comparable predictive performance to large datasets.

Main Methods:

• Implementing an iterative chemogenomic selection technique for training data.
• Developing protocols for model construction, updating, and evaluation.
• Analyzing the iterative process of model building.

Main Results:

• Automated selection yields significantly smaller training sets compared to using all available data.
• Models built on selected data retain prediction performance comparable to those trained on large datasets.
• This approach enhances training efficiency and potentially interpretability.

Conclusions:

• "Big data" is not always necessary for effective chemogenomic modeling.
• Automated adaptive data selection offers a more efficient alternative for building predictive ligand-target models.
• This method supports robust pharmacological research with reduced computational burden.