Deep Dive into Machine Learning Models for Protein Engineering

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Protein redesign is crucial for pharmaceutical R&D.
• Laboratory evolution mimics natural selection for protein engineering.
• The vast number of possible protein mutations makes exhaustive screening impractical.

Purpose of the Study:

• To benchmark machine learning (ML) models for protein redesign.
• To evaluate various protein sequence descriptors, including novel ones.
• To identify the most effective ML approaches for pharmaceutical protein engineering.

Main Methods:

• Benchmarking ML prediction models (e.g., deep learning).
• Utilizing diverse protein sequence descriptors (single amino acid and 3D structure-based).
• Evaluating model performance on public and proprietary datasets using multiple metrics.

Main Results:

• Convolution Neural Network (CNN) models demonstrated strong performance.
• Models utilizing amino acid property descriptors were particularly effective.
• Performance varied across different ML methods and descriptor types.

Conclusions:

• CNNs with amino acid property descriptors are highly applicable to pharmaceutical protein redesign.
• Further exploration of ML models and descriptors is warranted.
• This work provides a benchmark for selecting ML tools in protein engineering.