Infer global, predict local: Quantity-relevance trade-off in protein fitness predictions from sequence data

Area of Science:

• Computational biology
• Protein engineering
• Evolutionary biology

Background:

• Predicting mutation effects on protein function is crucial for evolutionary studies and biomedicine.
• Computational methods, including deep learning, analyze sequence data to predict protein fitness landscapes.
• The interplay between model complexity and data characteristics in prediction accuracy is not well understood.

Purpose of the Study:

• To develop a theoretical framework for understanding prediction error in computational models of protein function.
• To identify key data descriptors that influence predictive performance.
• To determine optimal data selection strategies for improving prediction accuracy.

Main Methods:

• Theoretical analysis of prediction error based on sequence data descriptors.
• Proposing quantity and relevance as key data characteristics.
• Analyzing trade-offs between data quantity, relevance, and model complexity.
• Using repeated subsampling to assess model capture of epistasis.

Main Results:

• A theoretical framework was established to describe prediction error.
• Descriptors for sequence data quantity and relevance were proposed.
• A trade-off between data quantity and relevance was identified, guiding optimal data subset selection.
• Simple models can outperform complex ones when trained on carefully selected data.
• Repeated subsampling reveals the extent to which computational models capture fitness landscape epistasis.

Conclusions:

• Optimal prediction of mutation effects relies on a balance between data quantity and relevance, not solely on model complexity.
• Careful data selection can enhance the performance of simpler computational models.
• The proposed framework provides insights into model limitations regarding epistasis in protein fitness landscapes.