Enhancing the accuracy of chemogenomic models with a three-dimensional binding site kernel

Area of Science:

• Computational chemistry
• Cheminformatics
• Structural biology

Background:

• Computational chemogenomic methods predict target-ligand interactions using machine learning.
• Previous studies questioned the benefit of 3-D structural target descriptors over simpler sequence-based or property-based information.

Purpose of the Study:

• To assess if 3-D structural information improves chemogenomic modeling accuracy.
• To compare 3-D target descriptors with sequence-based descriptors in predicting target-ligand pairs.

Main Methods:

• Utilized a target kernel based on 3-D binding site distances from X-ray structures.
• Employed a support vector machine (SVM) classifier with standard ligand kernels and 3-D or sequence-based target kernels.
• Trained and tested models on known target-ligand PDB complexes and false pairs.

Main Results:

• The 3-D target kernel significantly outperformed the sequence-based target kernel in discriminating true from false target-ligand pairs.
• SVM models achieved high recall (70%), precision (99%), and specificity (99%) in predicting external target-ligand associations.
• Global models, pooling all data, generally yielded better statistics than local models.

Conclusions:

• Chemogenomic models incorporating both ligand and 3-D target information outperform simpler ligand-based models.
• Recommended modeling practices include ligand-based models for abundant ligand data, structure-based chemogenomic models for known protein structures, and sequence-based chemogenomic models for other targets.