Modeling Static and Dynamic Protein Structure from 2D Infrared Spectrum with Stochastic Interpolant

A protein's function is intrinsically linked to its dynamic structure, yet predicting these conformational changes in real-time remains a central challenge in molecular biology. While two-dimensional infrared (2DIR) spectroscopy provides a powerful experimental window into these dynamics, translating its complex, low-dimensional signals into high-resolution 3D structures is a formidable interpretive hurdle. Current machine learning approaches typically sidestep this challenge by predicting auxiliary information for other computationally expensive tools, creating a slow and indirect workflow. In this paper, we introduce a direct, end-to-end solution: an equivariant generative pipeline based on stochastic interpolants augmented with a length prediction head. The model pretrained on a dataset of static protein bypasses intermediate steps, learning to generate 3D protein structures directly from their corresponding 2DIR spectra, outperforming previous methods. We further show that the model can be finetuned on correlated samples drawn from simulated reversible folding trajectories to improve its predictive power for dynamic protein structure prediction.