Improving the generalization of protein expression models with mechanistic sequence information

Area of Science:

• Molecular Biology
• Bioinformatics
• Machine Learning

Background:

• Growing demand for biological products necessitates maximizing heterologous protein expression.
• High-throughput sequencing data enables machine learning models for predicting protein expression from nucleotide sequences.
• Current models often use one-hot encodings, achieving high local accuracy but limited generalization.

Purpose of the Study:

• To investigate if mechanistic sequence features can improve the generalization of sequence-to-expression models.
• To compare model performance across different datasets (Escherichia coli and Saccharomyces cerevisiae).
• To explore strategies for integrating one-hot encodings and mechanistic features.

Main Methods:

• Comparative study across datasets in Escherichia coli and Saccharomyces cerevisiae.
• Exploration of feature stacking, ensemble model stacking, and geometric stacking (a novel graph convolutional neural network architecture).
• Integration of mechanism-agnostic (one-hot encodings) and mechanism-specific sequence features.

Main Results:

• Mechanistic sequence features significantly enhance model generalization capabilities.
• Integration strategies, including geometric stacking, improve predictive accuracy.
• Domain knowledge and feature engineering are crucial for accurate protein expression prediction.

Conclusions:

• Mechanistic sequence features are vital for improving the generalizability of protein expression prediction models.
• Combining different feature types and advanced architectures like geometric stacking offers a promising direction for predictive sequence design.
• This research highlights the importance of integrating biological domain knowledge into machine learning for biological applications.