PON-tstab: Protein Variant Stability Predictor. Importance of Training Data Quality

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Predicting the impact of amino acid substitutions on protein stability is crucial for understanding protein function and disease.
• Existing machine learning (ML) methods for variant stability prediction rely on benchmark datasets, primarily the ProTherm database.
• Issues with the quality, content, and relevance of the ProTherm database have been identified, potentially biasing ML model training.

Purpose of the Study:

• To address the limitations of existing benchmark datasets for protein stability prediction.
• To develop a more accurate and broadly applicable tool for predicting the effects of amino acid substitutions on protein stability.
• To emphasize the critical role of benchmark data quality in the development of predictive models.

Main Methods:

• A corrected dataset was obtained to overcome the identified issues with the ProTherm database.
• A random forests-based machine learning model was trained on the curated dataset.
• The developed tool, PON-tstab, was designed for applicability to variants in any organism.

Main Results:

• Identified significant errors and uncommunicated features in the widely used ProTherm database.
• Demonstrated that previous ML variant stability predictors were trained on biased and incorrect data.
• Developed PON-tstab, a novel tool providing predictions for stability-decreasing, stability-increasing, and neutral amino acid substitutions.

Conclusions:

• The quality, suitability, and appropriateness of benchmark datasets are paramount for developing reliable predictive models.
• PON-tstab offers a more accurate approach to predicting variant effects on protein stability due to the use of a corrected dataset.
• This work underscores the need for rigorous data curation in bioinformatics for robust ML model development.