Learning the shape of protein microenvironments with a holographic convolutional neural network

Area of Science:

• Computational biology
• Machine learning
• Structural biology

Background:

• Predicting protein function from sequence or structure remains a significant challenge despite advances in protein structure prediction.
• Proteins are crucial for numerous biological processes, including immune recognition and brain activity.

Purpose of the Study:

• To introduce a physically motivated machine learning approach, holographic convolutional neural network (H-CNN), for modeling amino acid preferences in protein structures.
• To develop an interpretable computational model for protein structure-function relationships.

Main Methods:

• Developed holographic convolutional neural network (H-CNN) for proteins.
• H-CNN models physical interactions within protein structures.
• The model incorporates evolutionary data to capture functional information.

Main Results:

• H-CNN accurately models amino acid preferences in protein structures.
• The method successfully predicts the impact of mutations on protein stability.
• H-CNN accurately predicts the binding of protein complexes.

Conclusions:

• H-CNN provides a physically grounded and interpretable method for understanding protein structure-function relationships.
• This approach can accurately predict functional impacts of mutations and complex binding.
• The model holds potential for guiding the design of novel proteins with specific functions.