Data set and fitting dependencies when estimating protein mutant stability: Toward simple, balanced, and...

Area of Science:

• Computational biology
• Biophysics
• Bioinformatics

Background:

• Accurate prediction of protein stability changes upon mutation (ΔΔG) is vital for understanding evolution, protein engineering, and identifying disease-causing gene variants.
• Existing prediction methods are often challenged by biases present in training datasets, particularly destabilization and mutation-type biases.

Purpose of the Study:

• To investigate the impact of training data biases on linear regression models for predicting ΔΔG.
• To develop and validate novel models that account for destabilization and mutation-type biases.
• To provide accessible and interpretable tools for ΔΔG prediction.

Main Methods:

• Systematic investigation of 45 linear regression models trained on datasets addressing destabilization and mutation-type biases (B_M).
• External validation of models on three diverse test datasets representing different pathologies.
• Internal consistency checks including symmetry and neutrality assessments.
• Development of two specialized models: SimBa-IB for natural mutations and SimBa-SYM for balanced mutation scenarios.

Main Results:

• Model performance was significantly influenced by training data composition and fitting methodologies.
• Two refined models, SimBa-IB and SimBa-SYM, were developed.
• SimBa-SYM demonstrated minimal bias on the S^sym dataset and robust performance across all test sets (R ≈ 0.46-0.54, MAE = 1.16-1.24 kcal/mol).

Conclusions:

• The developed SimBa models effectively address common biases in ΔΔG prediction.
• SimBa-SYM offers a simple yet powerful, non-biased approach for specific mutation scenarios.
• These models enhance interpretability, usability, and future development in protein stability prediction and are available on GitHub.