Neural networks to learn protein sequence-function relationships from deep mutational scanning data.

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Predicting protein function from amino acid sequence is a complex challenge.
• Deep mutational scanning (DMS) generates extensive data on sequence-function relationships.
• Current prediction methods have limitations in accuracy and scope.

Purpose of the Study:

• To develop a supervised deep learning framework for accurate protein sequence-function mapping.
• To evaluate different neural network architectures, including structure-aware models.
• To demonstrate the framework's ability to design novel proteins with improved properties.

Main Methods:

• Implemented a supervised deep learning framework using DMS data.
• Tested various neural network architectures, including graph convolutional networks (GCNs) incorporating protein structure.
• Compared deep learning performance against physics-based and unsupervised methods.

Main Results:

• The supervised deep learning approach significantly outperformed existing prediction methods.
• Networks capturing nonlinear interactions and sharing parameters across sequence positions were crucial for learning.
• Trained models learned biologically relevant information about protein structure and mechanism.
• Successfully designed a protein G B1 domain variant with substantially higher affinity for immunoglobulin G.

Conclusions:

• Supervised deep learning provides a powerful tool for understanding and predicting protein sequence-function relationships.
• The framework can effectively guide the design of proteins with novel or enhanced functions.
• This approach has broad implications for protein engineering and drug discovery.