A machine learning protocol for predicting structural distributions of amyloid-forming proteins from 2D IR spectra

Protein misfolding plays a central role in diseases such as Alzheimer's disease, Parkinson's disease, type 2 diabetes, and transthyretin amyloidosis (ATTR), often driven by specific aggregation-prone segments such as Aβ_17-23 and Aβ_37-42 of amyloid-β42 (Aβ42), α-Syn_66-74 and α-Syn_71-82 of α-synuclein (α-syn), hIAPP_22-27 of human islet amyloid polypeptide (hIAPP), and TTR_105-115 of transthyretin (TTR). Capturing the atomic-level structural features of these transient and dynamically fluctuating regions remains challenging. Two-dimensional infrared (2DIR) spectroscopy provides rich vibrational fingerprints that are highly sensitive to protein conformational dynamics, but extracting atomic-resolution structural information from these complex signals is nontrivial. In this study, we present a machine learning framework that integrates 2DIR spectra with deep structural modeling to reconstruct the three-dimensional atomic structures of monomeric intrinsically disordered aggregation-prone segments of amyloidogenic proteins. Using this model, we were able to predict the conformational ensembles of aggregation-prone segments from amyloid-β42, α-synuclein, human islet amyloid polypeptide, and transthyretin, as well as the structural evolution of Aβ42 bound to a small-molecule inhibitor, directly from computationally derived 2DIR spectra. An attention module highlights the most informative spectral features associated with local structural variations, providing interpretable links between spectra and structure. This generalizable strategy paves the way for interpreting time-resolved spectroscopic studies and offers a promising computational framework for probing misfolding-related structural dynamics and therapeutic mechanisms.