Learning protein fitness models from evolutionary and assay-labeled data

Area of Science:

• Computational biology
• Machine learning
• Protein engineering

Background:

• Protein fitness prediction models often rely on evolutionary data or experimental labels.
• Combining these data sources is beneficial, especially with limited experimental data.
• Existing methods for combining data can be complex.

Purpose of the Study:

• To propose a simple yet effective method for predicting protein fitness.
• To combine evolutionary sequence information with experimental fitness data.
• To evaluate the performance of this combined approach against existing methods.

Main Methods:

• Utilized ridge regression on site-specific amino acid features.
• Incorporated a probability density feature derived from evolutionary data modeling.
• Compared a variational autoencoder (VAE)-based density model with other evolutionary models.

Main Results:

• The proposed simple combination approach demonstrated competitive and often superior performance compared to more sophisticated methods.
• A variational autoencoder-based probability density model yielded the best results within the proposed framework.
• The study underscored the necessity of rigorous evaluations and robust baselines.

Conclusions:

• A straightforward combination of evolutionary and experimental data offers a powerful approach for protein fitness prediction.
• The choice of evolutionary density model impacts performance, with VAEs showing promise.
• Systematic evaluation is crucial for advancing machine learning models in protein science.