The study, published in Smart Agricultural Technology, was led by Elham Gorgannejad of the University of Manitoba together with Qian Liu, Catherine Rui Jin Findlay, Mohammad Nadimi, Alex Chun-Te Ko, Pankaj Bhowmik and Jitendra Paliwal. The researchers tackled a deceptively difficult problem: predicting the multi-environment mean cysteine concentration of twenty pea cultivars grown at three Saskatchewan field locations, using nothing but surface-enhanced Raman spectroscopy, or SERS, measurements of their alkaline extracts. SERS works by adsorbing molecules onto plasmonic nanostructures that dramatically amplify the otherwise weak Raman scattering signal, and under controlled conditions the scattered intensity scales with the number of molecules on the surface, a property that makes quantitative analysis theoretically possible.

The experimental pipeline was substantial. Flours from twenty cultivars of the CDC breeding program at the University of Saskatchewan, grown at Limerick, Rosthern and Sutherland, were homogenized in water and extracted under alkaline conditions at roughly pH 9. Just before measurement, each extract was mixed with tris(2-carboxyethyl)phosphine, a reducing agent that liberates free thiol groups so they can chemisorb to the sensor surface. Commercial paper-based SERS substrates from Metrohm were then immersed in each mixture for 45 minutes and interrogated with a 785-nanometer excitation laser. The team collected spectra from three spots per substrate, 36 spectra per spot, producing 108 spectra per sample and a staggering 6,480 total spectra across the study, spanning two separate substrate manufacturing batches.

Reference values came from conventional high-performance liquid chromatography using performic acid oxidation and acid hydrolysis, which converts cysteine into stable cysteic acid for accurate quantification. Because growing environment shapes amino acid content, the researchers averaged HPLC values across the three field locations to produce a single cultivar-level mean cysteine reference for each of the twenty cultivars. These means spanned a narrow window, from 0.3120 to 0.3732 grams per 100 grams, a total spread of barely 0.06 grams, a detail that would prove critical to the study’s outcome. With the dataset assembled, the team pitted five algorithms against the spectra: linear regression, partial least squares regression, support vector regression, random forest regression, and a one-dimensional convolutional neural network built in PyTorch.

The 1D-CNN was designed specifically for the structure of spectral data. Where classical regressors treat each of the 1,496 Raman shift bins as an independent feature, the convolutional network learns hierarchical local patterns such as peak shapes, widths and relative shifts, building them through four convolutional blocks with 16 to 128 filters, batch normalization, ReLU activation and max pooling, followed by fully connected layers with dropout regularization. Training used the AdamW optimizer with a Huber loss and a OneCycle learning rate schedule. Preprocessing was tailored per model, combining Savitzky–Golay smoothing, modified polynomial baseline correction to strip fluorescence background, and, for some models, min–max normalization. The source code and a sample dataset were released openly on GitHub to support reproducibility.

Evaluation was split into two deliberately different regimes. Within-cultivar testing assigned 80 percent of each cultivar’s spectra to model development and held out 20 percent for testing, a controlled assessment of how well models cope with technical variability from substrate heterogeneity, fluorescence drift and stochastic noise. The stricter test was leave-one-cultivar-out cross-validation, in which an entire cultivar was withheld and the model trained on the remaining nineteen was asked to predict it, repeated twenty times. Under within-cultivar evaluation, the deep learning model excelled, achieving an R-squared of 0.903 and a root mean squared error of 0.005 grams per 100 grams on preprocessed spectra, comfortably ahead of random forest at 0.822 and the linear baselines. The effect of preprocessing proved model-dependent, helping most models but actually hurting partial least squares regression.

The leave-one-cultivar-out results told a more sobering and scientifically revealing story. Every model declined sharply when facing an unseen cultivar, with R-squared values collapsing to between 0.045 and 0.168 for the classical methods, while the 1D-CNN held at 0.451 with an RMSE of 0.013 grams per 100 grams. The authors attribute this drop to three converging factors: only twenty independent cultivar-level reference targets existed despite the thousands of technical spectra, the cysteine range was extremely narrow, and cultivar-specific matrix characteristics beyond cysteine itself leak into the spectral signatures. Repeated measurements of the same flour sample, however numerous, cannot substitute for genuinely independent biological and environmental samples, a lesson with broad implications for anyone applying machine learning to spectroscopic food data.

Interpretability analysis using Shapley Additive Explanations shed light on what the network was actually learning. In the within-cultivar setting, important features spread across many regions, including the 630 to 720 per-centimeter range associated with C–S vibrational modes of sulfur-containing amino acid residues, along with features near 244 to 313 and 900 to 1,573 per-centimeter. Under the stricter cross-cultivar regime, importance concentrated dramatically in the low Raman shift region from roughly 200 to 275 per-centimeter, bands tied to metal–adsorbate interactions and substrate phonon modes rather than internal molecular vibrations, together with contributions near 715 to 731 per-centimeter within the C–S band range. A truncation experiment removing the low-shift region actually worsened performance, showing the model was extracting genuinely useful, if partly substrate-derived, information rather than noise.

The team also stress-tested the network against simulated additive noise, scaling noise according to signal-averaging theory to emulate effective scan counts from 64 down to 1. Performance degraded gracefully: R-squared fell only from 0.868 at 64 scans to 0.837 at 8 scans, then dropped more steeply to 0.608 at a single scan. This relative stability at low to moderate noise levels suggests practical potential for faster, cheaper acquisitions, although the authors caution that the simulated noise was not calibrated to specific instrument settings and should not be read as a guarantee of performance at particular acquisition parameters.

Overall, the researchers frame the work as a proof of concept rather than a validated analytical method, noting that to their knowledge it is the first application of deep learning to quantify a specific amino acid in legume extracts using SERS. Future improvements will require repeated HPLC measurements, independent extraction replicates, additional growing environments, and, above all, a larger and more diverse set of independent reference samples spanning a wider concentration range. If those hurdles can be cleared, the approach could evolve from a laboratory curiosity into a routine breeding tool, letting plant scientists rapidly identify high-cysteine pea lines and helping the plant-protein industry close the nutritional gap with animal protein, one spectrum at a time.

Subject of Research: AI-based quantification of cysteine in pea cultivars from surface-enhanced Raman spectra

Article Title: Cysteine quantification in pea cultivars from SERS spectra using AI

Article References: Gorgannejad, E., Liu, Q., Findlay, C. R. J., Nadimi, M., Ko, A. C.-T., Bhowmik, P., & Paliwal, J. (2026). Cysteine quantification in pea cultivars from SERS spectra using AI. Smart Agricultural Technology, 15, Article 102557. https://doi.org/10.1016/j.atech.2026.102557

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102557

Keywords: SERS, cysteine, pea cultivars, deep learning, convolutional neural network, Raman spectroscopy, plant protein, food analysis, machine learning, legume nutrition, SHAP interpretability, HPLC reference values