Generalizable Cysteine Quantification in Pea Cultivars from SERS Spectra Using AI

This study demonstrates that a one-dimensional convolutional neural network (1D-CNN) can accurately and generally predict cysteine concentrations across diverse pea cultivars using surface-enhanced Raman spectroscopy (SERS) spectra, offering a rapid alternative to traditional high-performance liquid chromatography (HPLC) methods.

The Gist

The Big Picture: Finding the "Golden Needle" in a Haystack

Imagine you are a chef trying to make the world's best plant-based burger. You know that peas are a great source of protein, but there's a catch: they are missing a specific, crucial ingredient called cysteine. Without enough cysteine, the protein isn't "complete," and the nutritional value drops.

To fix this, plant breeders need to test thousands of different pea varieties to find the ones that naturally have high cysteine levels.

The Problem: The traditional way to test for cysteine is like sending a letter to a distant country and waiting three weeks for a reply. It involves complex chemistry, expensive machines, and takes a long time. You can't do this quickly enough to screen thousands of peas.

The Solution: This paper is about teaching a computer to "sniff out" the cysteine in peas instantly using a special kind of light (called SERS) and a super-smart brain (called AI).

---

The Tools: The "Super-Microscope" and the "Smart Brain"

1. The Super-Microscope (SERS)

Think of a normal flashlight. If you shine it on a tiny speck of dust, you might not see it. But if you use a magnifying glass that focuses all the light onto that speck, it glows brightly.

• Raman Spectroscopy is like a flashlight that bounces off molecules to tell you what they are.
• SERS (Surface-Enhanced Raman Spectroscopy) is that flashlight with a super-powerful magnifying glass. It uses tiny silver nanostructures (like a microscopic trampoline) to bounce the light off the cysteine molecules so brightly that the computer can see them clearly, even when they are hiding in a complex soup of pea juice.

2. The Smart Brain (AI)

Once the "flashlight" captures the data, it looks like a messy scribble of lines and bumps. A human can't read this quickly. That's where the AI comes in.

The researchers tried five different types of "brains" to read these scribbles:

• The Linear Thinkers (LR, PLSR): These are like students who only know how to draw straight lines. They try to connect the dots with a ruler. They work okay if the dots are neat, but they get confused if the dots are messy.
• The Flexible Thinkers (SVR, Random Forest): These are smarter. They can draw curves and handle some messiness.
• The Deep Learning Brain (1D-CNN): This is the champion. Imagine a detective who doesn't just look at individual dots, but looks at the shape of the whole pattern. It understands that a "hill" in the data might look slightly different depending on the weather, but the shape of the hill tells the real story.

---

The Experiment: The "New Student" Test

The researchers tested these brains on 20 different types of peas grown in three different locations.

Test 1: The Familiar Classroom (Within-Cultivar)
They trained the AI on Pea Type A and tested it on Pea Type A.

• Result: All the brains did pretty well. Even the "Linear Thinkers" could guess the cysteine levels because they were just memorizing the specific look of Pea Type A.

Test 2: The New Student (Leave-One-Cultivar-Out)
This is the real challenge. They trained the AI on 19 types of peas, then threw a brand new, unseen type of pea at it.

• The Linear Thinkers: They failed miserably. They were so used to the specific "look" of the old peas that when the new pea showed up (which looked slightly different due to soil or weather), they got completely lost.
• The Deep Learning Brain (1D-CNN): It didn't panic. It recognized the underlying pattern of cysteine, regardless of which pea variety it was. It successfully predicted the cysteine levels in the new pea.

The Analogy:
Imagine trying to recognize a friend's voice.

• The Linear Thinkers are like someone who only knows your friend when they are wearing a specific hat. If your friend takes off the hat, they don't recognize them.
• The Deep Learning Brain is like a parent. They recognize their child's voice whether the child is wearing a hat, a scarf, or shouting from across the room. They learned the essence of the voice, not just the costume.

---

The "Secret Sauce": What Did the AI Learn?

The researchers asked the AI, "How did you know that?" using a tool called SHAP (which is like a highlighter pen).

• The Linear Thinkers highlighted random parts of the data, often getting distracted by background noise (like the color of the cup the pea juice was in).
• The Deep Learning Brain highlighted a very specific range of the light spectrum (between 630 and 760). This range corresponds to the specific chemical "fingerprint" of the sulfur bond in cysteine.
• Why this matters: It proves the AI isn't just guessing; it actually learned the chemistry. It found the "golden needle" in the haystack.

The Practical Win: Speeding Up the Process

Finally, the researchers asked, "How many times do we need to scan the pea to get a good answer?"

• Scanning takes time. If you scan 64 times, it's very accurate but slow. If you scan once, it's fast but noisy.
• They simulated "noisy" scans and found that the AI could still give a great answer with just 8 scans.
• The Result: This cuts the testing time by a huge margin, making it possible to screen thousands of peas in a day instead of a month.

The Bottom Line

This paper shows that we can replace slow, expensive, chemical lab tests with a fast, cheap, light-based scanner powered by a smart AI.

By using a "Deep Learning" brain, we can now quickly find the best pea varieties for breeding programs, ensuring that the plant-based proteins of the future are not only tasty but also nutritionally complete. It's a giant leap forward for making healthy food faster and cheaper to produce.

Technical Summary

1. Problem Statement

• Nutritional Context: Legumes are a vital plant-based protein source, but their protein quality is often limited by low levels of sulfur-containing amino acids (SCAAs), specifically cysteine and methionine. Accurate quantification of cysteine is critical for breeding programs and quality control in the food industry.
• Limitations of Conventional Methods: Standard analytical techniques like High-Performance Liquid Chromatography (HPLC) and Gas Chromatography–Mass Spectrometry (GC–MS) are accurate but time-consuming, resource-intensive, and require complex sample preparation (hydrolysis, derivatization). They are unsuitable for high-throughput screening of large breeding populations.
• Challenges in Spectroscopy: While Surface-Enhanced Raman Spectroscopy (SERS) offers rapid, label-free detection, it suffers from significant variability due to:
  • Intra-cultivar variability: Instrumental noise, fluorescence background, and substrate heterogeneity (variations in plasmonic nanostructures).
  • Inter-cultivar variability: Genotype-by-environment (G×E) interactions that alter the biochemical composition of pea extracts, changing peak intensities and positions across different cultivars.
• The Gap: Existing machine learning (ML) models often fail to generalize across unseen cultivars because they rely on absolute peak intensities, which fluctuate due to the factors above. There is a need for a robust AI framework that can predict cysteine concentrations in new, unseen pea cultivars without retraining.

2. Methodology

Data Acquisition and Preparation

• Dataset: 20 distinct pea cultivars grown at three geographically distinct locations (Saskatchewan, Canada), resulting in 60 biological samples.
• Ground Truth: Cysteine concentrations were measured using a performic acid oxidation–acid hydrolysis HPLC method.
• SERS Measurement:
  • Substrate: Paper-based SERS (P-SERS) substrates with plasmonic surfaces.
  • Protocol: Pea flour extracts were treated with TCEP to reduce disulfide bonds, liberating free thiols for chemisorption to the substrate.
  • Acquisition: 785 nm excitation laser; 3 spots per substrate, 36 spectra per spot, resulting in 108 spectra per sample (Total: 6,480 spectra).
• Preprocessing: A workflow involving Savitzky–Golay (SG) smoothing, modified polynomial baseline correction (ModPoly), and min–max normalization was applied. Notably, the Deep Learning model utilized internal normalization (batch normalization) rather than external scaling.

AI Modeling Framework

Five models were evaluated to predict cysteine concentration from SERS spectra:

1. Linear Regression (LR)
2. Partial Least Squares Regression (PLSR)
3. Support Vector Regression (SVR) with RBF kernel
4. Random Forest Regression (RFR)
5. 1D Convolutional Neural Network (1D-CNN): A deep learning architecture with four convolutional blocks (kernel size 5, increasing filters 16→128), batch normalization, ReLU activation, max-pooling, and fully connected layers.

Evaluation Strategy

To rigorously test generalization, two distinct evaluation protocols were used:

1. Within-Cultivar Split: 80% training / 20% testing from the same cultivar. This tests robustness against instrumental noise and substrate heterogeneity (intra-cultivar variability).
2. Leave-One-Cultivar-Out (LOCO): The model is trained on 19 cultivars and tested on the 20th (unseen) cultivar. This tests the ability to generalize across genetic and environmental differences (inter-cultivar variability).

Post-Hoc Analysis

• Interpretability: Shapley Additive Explanations (SHAP) were used to identify which Raman shift regions drove the predictions.
• Noise Modeling: A synthetic noise augmentation study simulated varying scan counts (from 64 down to 1) to determine the minimum acquisition time required for accurate prediction.

3. Key Results

Performance Comparison

• Within-Cultivar Performance: All models performed reasonably well. The 1D-CNN achieved the best results (RMSE = 0.008 g/100 g, $R^2$ = 0.862) and was notably robust, performing equally well on raw and preprocessed data. Traditional ML models (LR, PLSR, SVR, RFR) required preprocessing to mitigate noise and baseline drift.
• LOCO (Generalization) Performance:
  • Traditional ML Models: Performance collapsed when applied to unseen cultivars. $R^2$ values dropped drastically to 0.03–0.12, and RMSE increased by an order of magnitude. This indicates these models learned cultivar-specific intensity patterns rather than the underlying chemical signature of cysteine.
  • 1D-CNN: Demonstrated exceptional generalization. It maintained an RMSE of 0.011 g/100 g and an $R^2$ of 0.795 on unseen cultivars. The convolutional architecture successfully learned local spectral features (peak shapes, widths, relative shifts) that are invariant to absolute intensity fluctuations caused by G×E interactions.

Interpretability (SHAP Analysis)

• Within-Cultivar: Feature importance was distributed across multiple Raman regions.
• LOCO: Feature importance concentrated in the 630–760 cm⁻¹ range. This region corresponds to Carbon-Sulfur (C–S) vibrational modes, confirming the model learned chemically relevant features for cysteine rather than substrate artifacts. A low-frequency feature (~200 cm⁻¹) related to substrate-metal interactions was also noted but was less dominant in the cross-cultivar context.

Noise Modeling & Acquisition Optimization

• The study simulated reducing the number of scans (co-additions) to optimize acquisition time.
• Finding: Performance remained stable down to 8 scans (RMSE ~0.011, $R^2$ ~0.77). Below 4 scans, performance degraded significantly. This suggests that acquisition time can be reduced by ~75% (from the experimental 36 scans/spot to ~8) without sacrificing predictive accuracy.

4. Key Contributions

1. First Quantitative DL Application for Amino Acids in Legumes: This is the first study to apply deep learning to quantify a specific amino acid (cysteine) in complex legume matrices using SERS.
2. Demonstration of Cross-Cultivar Generalization: The study proves that while traditional chemometric methods fail on unseen genotypes, 1D-CNNs can generalize effectively across diverse pea cultivars and growing environments, a critical requirement for breeding applications.
3. Chemical Interpretability: By linking SHAP importance to known C–S vibrational bands (630–760 cm⁻¹), the study validates that the AI model relies on chemically meaningful signals rather than spurious correlations.
4. Operational Efficiency: The noise modeling provides practical guidelines for reducing SERS acquisition time (to ~8 scans) while maintaining high accuracy, facilitating high-throughput deployment.

5. Significance

This research establishes a scalable, rapid, and non-destructive framework for high-throughput phenotyping of nutritional quality in legumes. By overcoming the "generalization gap" that plagues traditional spectroscopic methods, the proposed 1D-CNN approach enables breeders to screen thousands of new cultivars for high cysteine content without the need for slow, expensive HPLC analysis. This directly supports the development of high-quality plant-based protein ingredients and sustainable food systems. The methodology is also transferable to other complex food matrices and amino acid profiling tasks.