Prediction of amyloid aggregation rates by machine learning and feature selection

Area of Science:

• Biochemistry
• Computational Biology
• Machine Learning

Background:

• Amyloid aggregation is a complex process implicated in various diseases.
• Predicting protein aggregation rates is crucial for understanding and treating amyloid-related disorders.
• Existing predictive models often lack sufficient accuracy and comprehensive feature analysis.

Purpose of the Study:

• To develop a novel, data-based machine learning algorithm for predicting amyloid aggregation rates.
• To compare the performance of the novel algorithm against traditional predictive methods.
• To identify key features that characterize amyloid aggregation kinetics.

Main Methods:

• A feedforward fully connected neural network (FCN) with one hidden layer was employed.
• The FCN was trained on a dataset of 21 amyloid proteins and tested on 4 additional proteins.
• Feature importance was assessed using correlation analysis and principal component analysis.

Main Results:

• The FCN achieved an average accuracy higher than 90%, outperforming multivariable linear regression and support vector regression.
• Seven key features were identified as crucial for characterizing amyloid aggregation kinetics: folding energy, HP patterns (helix, sheet, helices cross membrane), pH, ionic strength, and protein concentration.
• A minimum feature set was established for predicting protein aggregation.

Conclusions:

• The developed FCN model provides a highly accurate method for predicting amyloid aggregation rates.
• The identified key features offer insights into the fundamental drivers of amyloid aggregation kinetics.
• This data-driven approach advances the field of computational biology and protein aggregation research.