Abstract
Nuclear Magnetic Resonance (NMR) is used in structural biology to experimentally determine the structure of proteins, which is used in many areas of biology and is an important part of drug development. Unfortunately, NMR data can cost thousands of dollars per sample to collect and it can take a specialist weeks to assign the observed resonances to specific chemical groups. There has thus been growing interest in the NMR community to use deep learning to automate NMR data annotation. Due to similarities between NMR and audio data, we propose that methods used in acoustic signal processing can be applied to NMR as well. Using a simulated amino acid dataset, we show that by swapping out filter banks with a trainable convolutional encoder, acoustic signal embeddings from speaker verification models can be used for amino acid classification in 2D NMR spectra by treating each amino acid as a unique speaker. On an NMR dataset comparable in size with of 46 hours of audio, we achieve a classification performance of 97.7% on a 20-class problem. We also achieve a 23% relative improvement by using an acoustic embedding model compared to an existing NMR-based model.
Introduction
Nuclear Magnetic Resonance (NMR) spectroscopy is a vital tool for determining protein structures, but it faces significant bottlenecks: high data collection costs and the manual expert labor required for annotation. While deep learning offers automation potential, the field suffers from a "data desert" of oversimplified synthetic data and slower architectural innovation compared to fields like speech processing.
This research demonstrates that amino acids in 2D NMR spectra can be classified with high accuracy by treating them like human speakers in an audio recognition task. By repurposing state-of-the-art acoustic embedding models and a physics-based simulation pipeline, the researchers achieved 97.7% classification accuracy, outperforming existing NMR-specific models by 23%.
Methods
The researchers bridged the gap between chemistry and audio engineering by treating molecular "fingerprints" as "audio clips." The pipeline involves several key stages:
- Signal Stacking: 2D NMR interferograms (TOCSY experiments) are stacked into a single long 1D sequence.
- Acoustic Architectures: The team utilized the Speechbrain framework to implement speaker verification models, specifically x-vector and ECAPA-TDNN (Emphasized Channel Attention, Propagation, and Aggregation).
- Trainable Encoders: Instead of fixed Mel filter banks, a trainable 1D convolutional encoder was used to learn the specific "phonemes" of NMR signals.
- Transfer Learning: Weights pre-trained on human speech (VoxCeleb) were tested for fine-tuning on NMR data.
Datasets
A new, high-fidelity simulated dataset was introduced using the Spinach library for physics-based simulations of spin dynamics. This dataset includes 20,000 simulated proton-proton TOCSY experiments (1,000 for each of the 20 standard amino acids). This scale represents roughly 46 hours of dual-channel 8k audio, making it comparable to standard speech corpora like VCTK.
Key Insights and Ideas
- The "Speaker" Analogy: Treating each amino acid as a "unique speaker" and each experiment as an "utterance" allowed for the successful porting of speaker ID progress to structural biology.
- Visual Similarity: Mel spectrograms of NMR signals show high-energy bands that resemble human speech, despite lacking harmonic structures.
- Cross-Domain Transfer: A model with frozen weights trained only on human voices achieved 72.7% accuracy on NMR data, suggesting universal underlying patterns in time-domain oscillations.
Key Results
- Top Performance: The ECAPA-TDNN-L model with a 1D Conv encoder achieved 97.7% accuracy.
- Relative Gain: This is a 23% relative improvement over the existing NMR-based DEEP Picker model.
- Encoder Impact: Trainable encoders significantly outperformed fixed Mel filter banks, improving x-vector accuracy from 82.8% to 94.8%.
- Classification Challenges: The model struggled most with structural isomers like Isoleucine (ILE) and Leucine (LEU), which have nearly identical resonance patterns.
Limitations and Future Work
While the ECAPA-TDNN-L model is highly effective, it is significantly more complex (104M parameters) than previous models. Future work will focus on moving from simulated data to real-world experimental data, which includes instrument noise and solvent interference, and scaling the approach to classify sequences in larger, complex proteins.
Summary updated