Protein NMR assignment by isotope pattern recognition

Area of Science:

• Structural Biology
• Biophysics
• Computational Chemistry

Background:

• Sequential assignment in protein NMR spectra traditionally relies on manual, labor-intensive triple-resonance experiments.
• Existing software and heuristics aid the process but do not eliminate the manual effort required.
• Machine learning approaches are hindered by the need for massive training datasets encompassing all physical nuances and instrumental artifacts.

Purpose of the Study:

• To develop an automated and efficient method for amino acid signal identification in protein NMR spectra.
• To overcome the limitations of manual assignment and large, complex training databases in machine learning for NMR.
• To integrate a novel computational approach with existing industry-standard software for practical application.

Main Methods:

• Utilized polyadic decompositions for storing millions of simulated three-dimensional NMR spectra.
• Implemented on-the-fly generation of instrumental artifacts during the training of neural networks.
• Incorporated probabilistic methods for prior and posterior information and integrated with the CcpNmr software framework.

Main Results:

• Developed neural networks that process [1H, 13C] slices of HNCA spectra, accounting for varying CA signal shapes.
• The neural networks output an amino acid probability table, enabling rapid assignment.
• Successfully assigned backbones of common proteins (GB1, MBP, and INMT) rapidly using only the HNCA spectrum combined with primary sequence information.

Conclusions:

• The proposed method offers a significant advancement in automating amino acid signal identification in protein NMR.
• This approach drastically reduces the time and manual effort required for protein backbone assignment.
• The integration with CcpNmr and efficient data handling pave the way for broader adoption in structural biology.