Non-invasive and early detection of tomato spotted wilt virus infection in tomato plants using a hand-held Raman spectrometer and machine learning modelling

The goal of the study was to develop a fast, non-invasive method for the early detection of Tomato spotted wilt virus (TSWV) in tomato plants, using a hand-held Raman spectrometer combined with machine learning (ML) techniques. The approach aimed to identify infected plants within 3 to 7 days after inoculation, well before visible symptoms emerge.

Using Raman spectral data and Partial Least Squares Discriminant Analysis (PLS-DA), the study demonstrated high accuracy (90–95%) in classifying infected versus healthy plants. The method was further validated on different tomato genotypes, confirming its robustness with over 85% accuracy. This portable and cost-effective tool enables in-field diagnostics, supporting sustainable agriculture by enabling earlier, more targeted responses to TSWV outbreaks.

The original article

Non-invasive and early detection of tomato spotted wilt virus infection in tomato plants using a hand-held Raman spectrometer and machine learning modelling

Ciro Orecchio, Camilla Sacco Botto, Eugenio Alladio, Chiara D'Errico, Marco Vincenti, Emanuela Noris

Plant Stress, Volume 15, March 2025, 100732

https://doi.org/10.1016/j.stress.2024.100732

licensed under CC-BY 4.0

Selected sections from the article follow. Formats and hyperlinks were adapted from the original.

Tomato spotted wilt virus (TSWV) (family Tospoviridae, genus Orthotospovirus, species Orthotospovirus tomatomaculae) infects >1000 different plant species from 90 botanical families, including ornamental, fruit, horticultural, and agronomic crops (Ruark-Seward et al., 2020). TSWV is one of the 10 most destructive viruses infecting horticultural crops in the world (Scholthof et al., 2011), causing more than a billion dollars damage each year, therefore being one of the major threats to both specialty and staple crops around the world. This virus is transmitted by several thrips species, mainly the western flower thrips (Frankliniella occidentalis), but also the onion thrips (Thrips tabaci), and the chili thrips (Scirtothrips dorsalis) (Ullman et al., 2002). The thrips vector too is infected by this virus. On tomato plants, TSWV induces different symptoms, including leaf bronzing, small brown flecks and chlorotic spots, stunting, inward cupping and dropping of leaves, unilateral plant growth, dieback of growing tips, and ultimately death.

The major management and containment strategies of this disease rely on the use of resistant cultivars and insecticides for thrips control. However, to reduce the spread of the disease, limit crop damage, calibrate the use of insecticides, and support the breeding procedures to identify new sources of genetic resistance, early detection of this plant pathogen is fundamental in a sustainable crop management context. ELISA assays and polymerase chain reaction (PCR) are the most frequently utilized diagnostic techniques for TSWV recognition (Chinnaiah et al., 2022; Gao and Wu, 2022; Iturralde Martinez and Rosa, 2023; Roberts et al., 2000). However, these tests are laborious, invasive, costly, and time-consuming, therefore inappropriate for a broad range screening. Recently, innovative techniques for non-invasive disease detection have been developed, based on Raman spectroscopy (RS) (Payne and Kurouski, 2021; Saletnik et al., 2024), volatile compound detection (Li et al., 2019), and hyperspectral imaging (Nguyen et al., 2021).

RS exploits the interaction between photons emitted by a laser hitting the sample and its molecular components. The inelastic scattering of photons results in an energy shift of these photons, which is related to the molecular structure of the sample components and their vibrational modes, ultimately providing information about the chemical composition of the sample. RS has been applied for the detection of abiotic (Altangerel et al., 2017; Sanchez et al., 2020a) and biotic (Baratto et al., 2022; Egging et al., 2018; Farber and Kurouski, 2018; Kong et al., 2024; Mandrile et al., 2019, 2022; Sanchez et al., 2019a, 2019b) stress responses, fruit quality (Nekvapil et al., 2018), and chemical contamination (Mandrile et al., 2018). Moreover, RS is useful for quick and precise plant phenotyping, as well as for evaluating the nutritional content of grains (Farber et al., 2020; Krimmer et al., 2019).

We have previously described the RS capability of detecting the TSWV infection in tomato plants at an early stage, i.e. when symptoms are not yet visible (Mandrile et al., 2019). Moreover, we showed that RS allows to discriminate TSWV infection from the attack of other endemic pathogens affecting this crop (Mandrile et al., 2019). Recently, a RS-based approach was applied to confirm the infection by different TSWV strains in resistant and susceptible tomato varieties, but in this case RS analysis was performed only on symptomatic leaves collected at late time points after the artificial inoculation (Juárez et al., 2024).

Our previous experiments were conducted with a highly sensitive benchtop Raman spectrometer, which is nevertheless expensive, cumbersome, and unsuitable for non-destructive field applications. In this study, we show that the presence of different isolates of TSWV can be detected in both susceptible and resistant tomato plants at a very early stage of infection with a hand-held Raman device, ideal for field usage. To optimize the diagnostic performance of RS technique, advanced Machine Learning (ML) elaboration of the spectroscopic data has been implemented, allowing to achieve high accuracy levels. The model obtained was externally validated on two independent sets of susceptible and resistant plants. This study further supports the suitability of hand-held RS for point-of-care applications in plant pathology and plant phenotyping.

2. Materials and methods

2.2. Raman spectroscopy

Raman spectra of tomato leaves were collected using a hand-held Bruker BRAVO spectrometer (Bruker Optik GMBH, Ettlingen, Germany) equipped with two laser sources (785 and 853 nm). To suppress fluorescence BRAVO uses a sequentially shifted excitation (SSE) technology, as detailed by Jehlička et al. (2017). The spectrometer recorded spectra over a range of 300–3200 cm^-1 with a sampling resolution of 10–12 cm^-1. The OPUS software (Bruker, version 8.2) was used for data transfer, while a Spectragryph (version 1.2.16.1) optical spectroscopy software was used for data conversion, in particular from 0.0 files to .csv files (Menges, 2001). For each plant, Raman spectra were acquired from the apical leaflet of the second and third leaves from the apex (Fig. 1a). Three spectra for each leaf were taken in different spots, at both 3 and 7 days post inoculation (dpi). Overall, 48 spectra, i.e. 24 spectra for each time point, were acquired for both mock-inoculated (‘Healthy’) and virus-inoculated (‘TSWV’) plants. In the second experiment, aimed at validating the ML model, additional 28 spectra for virus-inoculated (‘TSWV’) and 37 for mock-inoculated (‘Healthy’) plants were collected, using the Bruker BRAVO spectrometer. In the third experiment on York plants, Raman spectra were collected with the same procedure, using a hand-held Zolix Finder Edge spectrometer (Zolix Instruments Co., LTD, Beijing, China) equipped with a laser source at 1064 nm, recording spectra over a range of 200–2000 cm^-1 with a sampling resolution of 14 cm^-1. The Spectragryph (version 1.2.16.1) optical spectroscopy software was used for data acquisition and data conversion, in particular from .txt to .csv files. In this case, 46 spectra for virus-inoculated (‘TSWV’) and 24 for mock-inoculated (‘Healthy’) were collected at 7 dpi on non-symptomatic leaves.

3. Results and discussion

3.2. Preliminary analysis of tomato leaf spectra

The average leaf spectra of both Healthy and TSWV-infected tomato plants (acquired at 3 and 7 dpi) showed vibrational bands originating from the major chemical components of plants, i.e. cellulose, lignin, carotenoids, and chlorophyll (Fig. 2, Table 1). In detail, the most evident peaks were those relative to carotenoids (1004, 1156, 1186, 1526 cm⁻¹), chlorophyll (1224, 1326 - 1328 cm⁻¹), pectin (peaks between 740 and 746 cm⁻¹) and cellulose (peaks at 916 and 1094 cm⁻¹ associated to ν(C–O–C) and ν(CO) modes of cellulose), while the peaks at 920 and 1610 cm⁻¹ are related to lignin. Carotenoids generated peaks at 1004, 1156, and 1526 cm⁻¹, attributed to in-plane CH₃ rocking modes, and C = C and C—C stretching, respectively. Lastly, the peaks between 1650 and 1690 cm⁻¹ can be assigned to proteins.

Plant Stress, Volume 15, March 2025, 100732: Fig. 2. Average Raman spectra for Healthy (mock-inoculated, blue line) and TSWV-infected (isolate P105, orange line) leaves of tomato (cv. Marmande), collected in the range 600–1800 cm−1. The insets show the details of the peaks in the range 1175–1350 cm-1 and 1470–1500 cm-1 of the spectra.

By averaging the 48 spectra collected from ‘Healthy’ plants and the 48 from ‘TSWV’, and comparing the averaged spectra, some minor differences could be evidenced (Fig. 2). In particular, it was observed that the signals relative to chlorophylls at 1224 and 1326–1328 cm⁻¹, carotenoids at 1156 and 1185 cm⁻¹, polyphenols at 1440 cm⁻¹, and those due to proteins at 1650–1690 cm⁻¹ appeared altered in TSWV-infected samples, showing a slight reduction of peak intensity. Also, the differences observed in the shoulder peak at 1490 cm⁻¹ and attributed to CH bending are likely to be exploited to discriminate ‘Healthy’ from ‘TSWV’ plants. Except for these slight differences, the average spectra of ‘Healthy’ and ‘TSWV’ plants were almost completely superimposable. Therefore, the adoption of Machine Learning techniques appears to be mandatory to enhance the information provided by the Raman spectra toward the detection of the plant infection.

4. Conclusions

In the ever-increasing need of triggering more sustainable policies of crop production, the early detection of plant infections by affordable and easy-to-use instruments and methods represents a fundamental target. Unlike benchtop instruments, the less sophisticated portable devices, considerably cheaper and more maneuverable even though less performing, are likely to make practically feasible control strategies that are precluded to most agricultural enterprises.

The present study demonstrates that a simple portable Raman spectrometer is perfectly capable of recognizing asymptomatic TSWV-infected tomato plants two weeks before the infection effects become visible, provided that appropriate Machine Learning algorithms are targeted to the enhancement and interpretation of the minor alteration that the pathogen induces in the Raman spectra. Needless to say, this approach could be adopted to reduce or avoid the use of nucleic acid or protein-based diagnostic assays, allowing to save time and reduce costs and labor in horticultural practices.

The method's performing parameters were measured on independent test samples and were externally validated on two completely separated sets of plants, one of which performed on plants with a different genotype and with different virus isolates. Accuracies outscored 96 %, 86 %, and 92 % respectively, even on plants tested 3 days after inoculation. These repeated impartial test confirmations ensure the method's reliability and prevent overfitting that may arise from data-dependent ML modeling. Lastly, both spectral pre-processing and ML data elaborations were managed rapidly on a standard personal computer and could be applicable in routine controls.

From a broader perspective, the possibility to detect the presence of a pathogen when symptoms are visually undetectable represents an important advantage in the agricultural diagnostic sector, particularly considering that it can be achieved using a portable Raman device, allowing to perform field measurements in real time, with an instrument much less expensive than a corresponding benchtop spectrometer. Moreover, this allows the acquisition of a large number of spectra, as the procedure is not destructive and does not require storage and transfer of material, thus enabling to increase the accuracy and reliability of the technique.

However, for a more general application of such proximal sensor-based diagnostic techniques, deeper elucidations of the biological pathways altered during plant pathogen infection are sought, investigating and possibly verifying a commonality of responses induced by different isolates of a pathogen on different cultivars.

In a possible scenario of practical implementation of RS for diagnostic purposes of plant viruses, the most interesting example consists in breeding for resistance, which is currently the prevalent defense strategy against plant viruses. For tomatoes, due to the long domestication process of this crop (Ferrero et al., 2020), the majority of resistance genes against viruses are identified in wild germplasm and introduced into cultivated genotypes through hybrid breeding. Specifically, the most relevant disease management strategy for TSWV relies on the Sw-5 locus, providing durable resistance against different tospovirus species and even against strains from diverse geographic locations (Turina et al., 2016). However, the frequent onset of resistance-breaking isolates (Aramburu et al., 2010; Latham and Jones, 1998) requires a continuous search of new sources of resistance, pushing towards the use of rapid diagnostic tools during breeding.

Here, we approached this issue testing the performances of a portable RS instrument on an ancient commercial tomato cultivar (cv. Marmande) susceptible to the majority of tomato viruses, including TSWV (Peiró et al., 2014); we expanded our analysis using a commercial tomato hybrid line carrying the Sw-5 locus, testing two different TSWV isolates, a wild type one (I244) unable to systemically spread in this hybrid and the T992 isolate, obtaining rather high accuracy values. Nonetheless, a limitation of this diagnostic tool could result from its application on plants with different genetic background originating from crosses with wild relatives of domestic tomatoes to introgress new resistance traits (Qi et al., 2021), ultimately leading to uneven perturbations of the host basal metabolome and of altered metabolic responses following virus attack.

Indeed, substantial gaps in the knowledge of the biological pathways tackled during virus infection in cultivated vegetables with different genetic backgrounds and how such genetic variability modifies the response to pathogens have to be filled. A combinatorial approach integrating -omics studies with proximal sensor techniques could help to comprehensively elucidate the complex array of responses occurring during plant-pathogen interactions, for example combining transcriptomics and metabolomics studies of plants infected by TSWV (Lv et al., 2023; Liang et al., 2024) with RS techniques.