This repository bridges empirical physical chemistry with interpretable artificial intelligence. It contains a headless data engineering pipeline designed to automate the extraction of covalent labeling (DEPC) footprints from MS2 fragmentation data (.mzML), formatting raw chemical mass shifts into standardized tensors.
These tensors feed a 1D Convolutional Neural Network (CNN) engineered to predict the kinetic aggregation rate of β-2-microglobulin (β2m) amyloid fibrils under varying Cu(II) concentrations.
Crucially, this architecture refuses the "black box" paradigm. Utilizing PyTorch's Captum library, the pipeline applies Integrated Gradients (Saliency Mapping) to the CNN's predictions, independently highlighting the specific amino acid residues driving the aggregation. This allows for the mathematical extraction of biological mechanisms, directly translating neural network gradients into 3D structural hotspots via py3Dmol.