MAHOMES II: A webserver for predicting if a metal binding site is enzymatic

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Protein function prediction is crucial for understanding biological processes.
• Experimental structure determination is a bottleneck for many proteins.
• Computationally generated protein structures offer a potential solution for functional annotation.

Approach:

• Improved a machine learning (ML) method to distinguish metalloenzyme from non-enzyme sites.
• Enhanced physicochemical features to reduce the need for sub-angstrom precision structures.
• Utilized ML to minimize errors in training data labeling.

Key Points:

• The improved classifier achieves 94% precision and 92% recall in identifying protein-bound metal sites.
• The model demonstrates high accuracy (90-97.5%) on predicted metalloprotein structures lacking experimental data.
• Key predictive features include local protein density, residue burial, and pocket accessibility.

Conclusions:

• The enhanced ML model accurately predicts enzymatic and non-enzymatic metal sites on computationally derived protein structures.
• This advancement facilitates the discovery of new enzymatic mechanisms and improves de novo metalloenzyme design.
• A webserver is available for the scientific community to explore protein function in uncharted structural space.