Uncertainty Exploration of Deep Learning Enabled Fast Multidimensional NMR Spectroscopy of Proteins

Area of Science:

• Biochemistry
• Spectroscopy
• Artificial Intelligence

Background:

• Deep learning (DL) accelerates multidimensional NMR spectroscopy but requires reliability assessments to prevent erroneous biological interpretations.
• Current DL methods for NMR lack robust quality control, hindering practical application in structural biology.

Purpose of the Study:

• To establish a reference-free quality assessment benchmark for reliable DL reconstruction in biological NMR spectroscopy.
• To integrate and evaluate uncertainty quantification frameworks for DL-based NMR data processing.

Main Methods:

• Integrated Deep Ensemble, Monte Carlo (MC) Dropout, and Evidential Deep Learning into DL reconstruction backbones.
• Developed and validated new metrics for assessing prediction reliability on 2D and 3D protein NMR spectra.
• Compared the performance of different uncertainty quantification frameworks.

Main Results:

• The established metrics enable global and frequency-level reliability assessment without reference data.
• Deep Ensemble showed superior reconstruction accuracy and uncertainty-reconstruction consistency.
• Evidential Learning provided real-time, computationally inexpensive uncertainty estimates consistent with traditional metrics.

Conclusions:

• The developed benchmark enhances the trustworthiness of DL in NMR spectroscopy.
• Findings empower structural chemists and biochemists to identify and focus on high-uncertainty regions in NMR data.
• This work facilitates more secure and reliable biomolecular structure determination using AI.