Develop machine learning-based regression predictive models for engineering protein solubility.

Area of Science:

• Biotechnology
• Computational Biology
• Protein Engineering

Background:

• Protein activity is crucial for biocatalyst cost-effectiveness, but direct prediction from sequence is limited by data availability.
• Protein solubility is correlated with activity, making it a viable indirect predictor.
• Existing models for protein solubility prediction often yield binary outputs, insufficient for guiding experimental improvements.

Purpose of the Study:

• To develop novel machine learning models for predicting protein solubility from amino acid sequences.
• To provide a more effective tool for protein engineering by predicting solubility in continuous numerical values.
• To indirectly aid in the prediction and improvement of protein activity.

Main Methods:

• Implementation of machine learning algorithms to predict protein solubility.
• Utilizing a novel approach for continuous numerical solubility prediction, moving beyond binary classifications.
• Development and hosting of an ML workflow as IPython notebooks on GitHub.

Main Results:

• Achieved an R2 value of 0.4115 using a support vector machine algorithm for continuous solubility prediction.
• Demonstrated that continuous solubility values are more informative for protein engineering than binary values.
• Established a publicly available ML workflow for protein solubility analysis.

Conclusions:

• The developed ML models offer a significant advancement in predicting protein solubility.
• Continuous solubility predictions provide a more practical approach for guiding experimental protein design and improvement.
• The provided workflow serves as a valuable template for future analyses of protein expression and solubility datasets.