Quantitative and Predictive Folding Models from Limited Single-Molecule Data Using Simulation-Based Inference

Single-molecule force spectroscopy resolves folding dynamics one molecule at a time, but extracting quantitative free-energy landscapes typically requires extensive datasets and careful instrument calibration to disentangle the molecule from linker and apparatus artifacts. We introduce a simulation-based inference framework that combines physics-based modeling with deep learning to recover the Bayesian posterior of a folding model directly from a single short trajectory. A 2-s constant-force measurement of a DNA hairpin is sufficient to reconstruct the folding landscape, matching deconvolution baselines that require 20-100 times more data and eliminating separate linker or instrument characterization. The same approach recovers four metastable states of a riboswitch aptamer from a single 5-s trajectory. Every parameter, including diffusion coefficients and linker stiffness, is included in the calibrated posterior. The current implementation assumes one-dimensional Markovian dynamics, but more complex models can be substituted within the same framework, opening single-molecule force spectroscopy to high-throughput parallel experiments and to systems where extensive data collection is impractical.