Predicting and Interpreting Protein Developability Via Transfer of Convolutional Sequence Representation.

Area of Science:

• Protein Engineering
• Computational Biology
• Biophysics

Background:

• Engineered proteins are valuable tools but often suffer from poor developability (expression, solubility, stability).
• Predicting protein developability from amino acid sequence can streamline candidate selection and reduce experimental costs.
• High-throughput screening has generated large datasets for protein developability, enabling machine learning approaches.

Purpose of the Study:

• To evaluate neural networks' ability to learn protein developability representations from high-throughput data.
• To predict recombinant expression for unseen protein sequences.
• To visualize and understand the protein fitness landscape and key amino acid properties influencing developability.

Main Methods:

• Development of a neural network model using a high-throughput developability dataset for the Gp2 protein scaffold.
• Convolutional approach to learn amino acid properties and predict expression levels.
• Nonlinear dimensionality reduction and nested sampling for fitness landscape analysis.

Main Results:

• The neural network model predicted expression levels 44% closer to experimental variance than a control.
• Learned amino acid embeddings revealed the significance of cysteine, hydrophobicity, and charge, while aromaticity was less important for small protein developability.
• Direct visualization of the protein fitness landscape identified evolutionary bottlenecks and distinct subpopulations.

Conclusions:

• Neural networks can effectively predict protein expression and interpret developability from limited data.
• Understanding amino acid contributions and fitness landscapes aids in designing improved protein scaffolds.
• This work bridges applied protein engineering with foundational biophysics for characterizing protein developability.