Machine learning methods can replace 3D profile method in classification of amyloidogenic hexapeptides

Area of Science:

• Biochemistry and computational biology
• Protein structure and aggregation
• Disease-associated protein studies

Background:

• Amyloids are protein aggregates linked to diseases like Alzheimer's.
• Identifying amyloidogenic peptides computationally is crucial but time-consuming.
• Current methods like ZipperDB are effective but computationally demanding.

Purpose of the Study:

• To accelerate the generation of amyloidogenic peptide datasets.
• To evaluate simplified computational methods and machine learning for amyloid prediction.
• To improve the efficiency of identifying disease-associated protein segments.

Main Methods:

• Developed a simplified 3D profile algorithm for faster dataset generation.
• Created a new dataset of 1779 hexapeptides, classifying 204 as amyloidogenic.
• Trained and tested machine learning models (Alternating Decision Tree, Multilayer Perceptron) on the dataset.

Main Results:

• Simplified 3D profile showed 93.5% overlap with ZipperDB.
• Machine learning models achieved high accuracy (91%) and ROC AUC (0.96).
• Computational time reduced from hours to seconds, significantly improving efficiency.

Conclusions:

• Simplified 3D profile generation is accurate and computationally efficient.
• Machine learning models effectively predict amyloidogenicity with reduced computational cost.
• This approach enables faster identification of disease-related amyloidogenic peptides.