Bioinformatics methods for identification of amyloidogenic peptides show robustness to misannotated training data

Area of Science:

• Biochemistry
• Computational Biology
• Biophysics

Background:

• Protein misfolding and aggregation into amyloid structures are implicated in neurodegenerative diseases like Alzheimer's and Parkinson's.
• Oligomeric species are the most cytotoxic forms, but their detection and classification can be challenging.
• Experimental methods for amyloid detection (e.g., AFM, IR spectroscopy, Thioflavin T) vary in cost, speed, and reliability, especially for oligomers.

Purpose of the Study:

• To evaluate the robustness of bioinformatics tools for amyloid identification when trained on imperfectly annotated data.
• To assess the capacity of in-silico methods to correct misclassifications arising from experimental data limitations.

Main Methods:

• Application of the AmyloGram predictor and three other amyloid predictors to assess their performance with weak supervision.
• Utilizing infrared (IR) spectroscopy and atomic force microscopy (AFM) for experimental validation of computational findings.

Main Results:

• Bioinformatics tools demonstrated a capacity to mitigate the impact of misannotations in reference datasets, even those included in the training set.
• Computational predictions were corroborated by new experimental data obtained through IR and AFM.

Conclusions:

• In-silico amyloid predictors show resilience to imperfect training data, offering a valuable approach to enhance classification accuracy.
• These findings support the utility of bioinformatics tools in overcoming experimental limitations for reliable amyloid detection and classification.