Classifying metal-binding sites with neural networks.

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Predicting protein scaffold impacts on catalysis requires understanding metal-binding features.
• Metalloenzymes utilize metal cofactors for critical catalytic functions.
• Robust classification schemes are needed to define protein features influencing reactivity.

Purpose of the Study:

• To develop and evaluate convolutional neural networks (CNNs) for classifying metal cofactor binding pockets in proteins.
• To assess the ability of CNNs to identify metal-binding sites, even when the metal cofactor is not visible.
• To advance the prediction of protein scaffold effects on catalysis.

Main Methods:

• Fine-tuning six CNN models for classifying 20 standard amino acids to select an optimal model.
• Training the selected CNN model to classify 2D images of the environment around metal binding sites (Fe ion or [2Fe-2S] cofactor).
• Evaluating CNN performance with visible and hidden metal cofactors, including sub-classifications of [2Fe-2S] cofactors (standard and Rieske).

Main Results:

• The CNN model achieved >95% accuracy in correctly identifying all three defined features (Fe ion, standard [2Fe-2S], Rieske [2Fe-2S]).
• The model demonstrated high performance in classifying metal-binding sites, even when the metal cofactor was not visible.
• CNNs proved effective in distinguishing between similar metal-binding sites.

Conclusions:

• Machine learning, particularly CNNs, offers a viable tool for identifying and classifying metal-binding sites in proteins.
• This methodology can predict impacts of protein scaffolds on catalysis by defining metal-binding features.
• The ability to classify sites without visible cofactors enhances metalloenzyme identification and characterization.