Interpretable Thermodynamic Score-based Classification of Relaxation Excursions.

Area of Science:

• Computational Biology
• Artificial Intelligence
• Statistical Mechanics

Background:

• Deep learning classifiers offer accurate predictions but lack uncertainty estimates and feature interpretability.
• Diffusion models excel at data generation but are underutilized for classification and prediction tasks.

Purpose of the Study:

• To develop a physics-inspired framework, Keeping SCORE, that enhances diffusion models for probabilistic classification and regression.
• To enable uncertainty quantification and feature attribution in AI predictions without altering existing models.

Main Methods:

• Transforming diffusion models into probabilistic engines by measuring dissipation along noising trajectories.
• Calculating exact class likelihoods and quantifying prediction confidence based on physical principles.
• Generating feature attributions to identify key input variables driving model decisions.

Main Results:

• Demonstrated accurate probability estimates and confidence quantification across diverse tasks including image recognition, single-cell genomics, and molecular biophysics.
• Provided interpretable feature attributions linked to physically meaningful coordinates.
• Showcased the framework's ability to enhance predictions without modifying pre-trained models.

Conclusions:

• Keeping SCORE bridges non-equilibrium statistical mechanics and AI, creating interpretable, uncertainty-aware predictions.
• This approach advances biological discovery by providing deeper insights into complex datasets.
• The framework offers a powerful tool for reliable and explainable AI applications in science.