Simulation-based inference for non-parametric statistical comparison of biomolecule dynamics.

Area of Science:

• Biophysics
• Computational Biology
• Statistical Mechanics

Background:

• Biomolecule random walks are complex, influenced by environmental factors and molecular changes.
• Existing models often fail due to experimental data limitations, preventing accurate analysis of dynamics.
• Multi-scale dynamics and non-canonical random walk behaviors complicate trajectory interpretation.

Purpose of the Study:

• To develop a model-agnostic statistical framework for comparing biomolecule dynamics across different experimental conditions.
• To enable robust analysis of random walk trajectories without requiring prior assumptions about underlying models.
• To provide a method for interpreting detected differences in terms of individual trajectory characteristics.

Main Methods:

• A two-step statistical testing scheme combining graph neural networks (GNNs) and maximum mean discrepancy (MMD) tests.
• GNNs are trained for simulation-based inference to generate rich summary statistics for individual trajectories.
• Non-parametric MMD tests are applied to compare summary statistics from trajectories under different conditions.

Main Results:

• The framework successfully validates on simulated trajectories, demonstrating statistical power and descriptive capability of learned statistics.
• The method detects changes in alpha-synuclein dynamics at neuronal synapses in response to membrane depolarization.
• Detected differences are attributed to increased protein mobility in the depolarized state, consistent with prior research.

Conclusions:

• The developed method offers a powerful, model-agnostic approach to analyze and compare biomolecule dynamics from experimental data.
• This framework facilitates the characterization of random walks and probing dataset heterogeneity at various granularities.
• The approach enhances our ability to interpret complex biomolecular motion and its environmental influences.