Large-scale design and refinement of stable proteins using sequence-only models

Area of Science:

• Protein Engineering
• Computational Biology
• Biophysics

Background:

• Engineered proteins require stable structures for function, but stable designs are rare among all possible amino acid sequences.
• Finding stable protein designs is resource-intensive due to extensive computational and experimental testing.
• A need exists for efficient methods to identify and generate stable protein sequences.

Purpose of the Study:

• To experimentally evaluate the stability of a large set of novel proteins using a high-throughput assay.
• To develop and validate a neural network model for predicting protein stability from amino acid sequences.
• To create a generative model for designing novel stable protein sequences.

Main Methods:

• Utilized a high-throughput, low-fidelity assay to assess the stability of approximately 200,000 novel proteins.
• Constructed a neural network model to predict protein stability based on amino acid sequences.
• Developed a second neural network for generating amino acid sequences of stable proteins.

Main Results:

• Experimentally evaluated stability for a large dataset of novel protein sequences.
• Developed a predictive neural network model correlating sequence with stability.
• Created a generative neural network capable of designing novel stable proteins.
• Demonstrated that the predictive model can enhance protein stability in both expert-designed and model-generated proteins.

Conclusions:

• High-throughput screening and neural network modeling can accelerate the discovery of stable engineered proteins.
• Predictive and generative models offer powerful tools for overcoming the rarity of stable protein designs.
• The developed models show potential for substantially improving protein stability in future designs.