Enzyme structure correlates with variant effect predictability.

Area of Science:

• Computational biology
• Protein engineering
• Machine learning

Background:

• Machine learning models are increasingly used for pre-screening protein variants.
• Current models show variable prediction accuracy for different protein variants.
• The characteristics of variants with large prediction errors are not well understood.

Purpose of the Study:

• To investigate whether structural characteristics influence the predictability of protein variant effects.
• To identify specific structural features associated with prediction errors in machine learning models.
• To guide improvements in machine learning-guided protein engineering.

Main Methods:

• Created a high-order combinatorial dataset of 3,706 enzyme variants.
• Partitioned the dataset based on structural characteristics (buriedness, contact residues, active site proximity, secondary structure).
• Trained and evaluated four supervised variant effect prediction (VEP) models on these partitioned subsets.

Main Results:

• Predictability strongly depended on all four tested structural characteristics.
• These dependencies were consistent across all four tested VEP models.
• The findings indicate that current algorithms do not fully account for structure-function determinants.

Conclusions:

• Structural characteristics significantly influence the predictability of protein variant effects.
• Improvements in VEP models can be achieved by incorporating new inductive biases and data modalities.
• Stratified dataset design is crucial for identifying areas for improvement in machine learning-guided protein engineering.