Spectral-driven Machine Learning for Amyloid Core Conformational Prediction

📌 Introduction

Amyloid protein misfolding underlies a range of neurodegenerative diseases, including Alzheimer’s, Parkinson’s, and Type II diabetes. Predicting the structural evolution of amyloidogenic protein fragments remains a fundamental challenge due to their dynamic and heterogeneous conformations. This project introduces a spectral-driven machine learning framework that integrates two-dimensional infrared (2DIR) spectroscopy simulations with Transformer-in-Transformer (TNT-S) models to predict misfolding-associated amyloid core structures.

⚙️Workflow

The workflow consists of three stages (Fig. 1):

1. Dataset construction
2. Machine learning protocol
3. Model application

The dataset used in this project can be accessed here: Dataset Link.

Fig. 1

📊 Results

The model demonstrates robust predictive capability across multiple amyloidogenic fragments (Fig. 2):

Model-informed prediction of the structural evolution of Aβ42 during molecule inhibitor binding (Fig. 3):

Usage

To reproduce the experimental results or apply this method to your own spectral data, please use the provided train.py script together with the pre-trained model weights available on Google Drive.