Accurate age-grading of field-aged mosquitoes reared under ambient conditions using surface-enhanced Raman spectroscopy and artificial neural networks

Abstract

Age-grading mosquitoes are significant because only older mosquitoes are competent to transmit pathogens to humans. However, we lack effective tools to do so, especially at the critical point where mosquitoes become a risk to humans. In this study, we demonstrated the capability of using surface-enhanced Raman spectroscopy and artificial neural networks to accurately age-grade field-aged low-generation (F₂) female Aedes aegypti mosquitoes held under ambient conditions (error was 1.9 chronological days, in the range 0–22 days). When degree days were used for model calibration, the accuracy was further improved to 20.8 degree days (approximately equal to 1.4 chronological days), which indicates the impact of temperature fluctuation on prediction accuracy. This performance is a significant advancement over binary classification. The great accuracy of this method outperforms traditional age-grading methods and will facilitate effective epidemiological studies, risk assessment, vector intervention monitoring, and evaluation.

Graphical Abstract

Graphical Abstract.

Introduction

Mosquito-borne pathogens, including malaria, Zika, dengue, and chikungunya, continue to be a major public health concern globally (World Health Organization 2019, 2022). Over 50% of the world’s population is considered at risk of being infected by these mosquito-borne diseases (World Health Organization 2014). With few effective registered vaccines available or antiviral therapies to combat the most important vector-borne infections, disease control relies on vector control. Controlled laboratory studies (Whitmire et al. 1987, Beier 1998) demonstrate that the period of pathogen/parasite infection in the mosquito, also known as the extrinsic incubation period, is long relative to mosquito lifespan. Consequently, only older mosquitoes are infectious and represent a risk to human health. For example, considering the earliest age when the dengue vector, Aedes aegypti, ingest their first blood meal (2 days after emergence) and mean ambient temperatures in the tropics, females must be 14 days or older to be potentially infective with dengue virus (Whitmire et al. 1987) and malaria transmission risk is associated with vectors older than 10 days (Beier 1998). For decades, scientists have sought to age-grade mosquitoes based on this understanding; however, no reliable, cost effective, and practical tools exist to accurately age-grade mosquitoes despite the tremendous epidemiological value of this approach. In the absence of such tools, making rapid and critical point-of-control decisions for assessing the risk of vector-borne diseases in a particular area is difficult, which will prevent public health experts from responding quickly to implement vector control practices and public health education messaging. In addition, the lack of such tools makes it challenging to evaluate the efficacy of new vector control strategies rapidly and effectively, especially those targeted at resistant populations.

Spectroscopic techniques have gained great attention due to their rapid data acquisition capability with minimum sample pretreatment. The applications of infrared spectroscopy, including near IR (NIR) and mid IR (MIR), for differentiating a variety of mosquito traits (age, species, and infection) have been reported (Johnson et al. 2020, Goh et al. 2021). For age-grading, the NIR technique has demonstrated great accuracy for classifying lab-reared mosquitoes into binary (parous/nulliparous) age categories (e.g., less than and greater than 7 days old, which presents) (Liebman et al. 2015, Sikulu-Lord et al. 2016, 2018, Lambert et al. 2018, Milali et al. 2019, Ong et al. 2020). However, when testing this technique for individual field-collected mosquitoes, accurate age prediction is still challenging. Krajacich et al. (2017) reported that the highest classification rate of an independent cohort of wild-caught Anopheles gambiae larvae reared to set ages was only 69.6%. Recently, Ong et al. (2020) concluded that they were unable to predict age of field-derived Aedes albopictus mosquitoes, and Joy et al. (2022) did not find NIR to be suitable for predicting the precise age of individual field-collected Ae. aegypti mosquitoes. In addition to NIR, a few studies using MIR have been reported for age prediction of mosquitoes. Khoshmanesh et al. (2017) and González-Jiménez et al. (2019) used MIR to classify very young and very old lab-reared mosquitoes ). A recent improvement reported by Siria et al. (2022) using machine learning algorithms achieved 95% accuracy in classification of ages 1–4, 5–10, and 11–17 days.

In our prior work (Wang et al. 2022), we established an approach based on surface-enhanced Raman spectroscopy (SERS) and chemometrics (i.e., principal component analysis [PCA] and partial least square [PLS] analysis) to quantitatively age-grade laboratory-reared Ae. aegypti mosquitoes with high accuracy (error within 1 day). Raman spectroscopy measures characteristic molecular vibrations based on polarizability. Compared with IR spectroscopy, Raman spectral characteristics are more distinct and carry richer information in general (Lin-Vien et al. 1991). Transition metal nanosubstrates, such as silver nanoparticles (AgNPs), greatly enhance the sensitivity and selectivity of Raman spectroscopy. A key innovation of SERS as compared to other spectroscopic methods is the interactions between the analyte and nanosubstrates which not only enhances sensitivity of analysis but also increases the selectivity of Raman spectroscopy for complex biomaterial analysis (Guo et al. 2017, Zhang et al. 2017, Lin and He 2019). AgNPs interact with specific age-related biomolecules (adenine-containing compounds and proteins) enabling SERS to generate unique and predictable spectral information for chemometrics models to determine the age of mosquitoes. However, this prior work focused on lab-reared mosquitoes under uniform conditions only.

The objective of this study was to test the capability of the SERS method combined with artificial neural networks (ANNs) to accurately age-grade low generation (F₂) Ae. aegypti mosquitoes, aged under ambient field temperatures. Compared to standard chemometric methods (e.g., PCA and PLS), nonlinear machine learning models have shown superior capability of quantitative and qualitative analysis of complex data matrices. More specifically, they have been shown to improve the robustness and accuracy of NIR and MIR analysis of mosquitoes (Liebman et al. 2015, Siria et al. 2022). To the best of our knowledge, this is the first study that demonstrates a significant advancement over binary or ternary classification and improved the accuracy of grading field-aged mosquitoes using traditional IR-based techniques. This developed method will enable prompt and effective assessment of mosquito-borne diseases risk in endemic regions, as well as monitoring and evaluation of the efficacy of mosquito control interventions.

Materials and Methods

Rearing of Field-Aged Mosquitoes Under Ambient Conditions

Adult Ae. aegypti were collected from ovitraps in the city of Medellin, Colombia in May 2022. Gravid females that emerged from the eggs deposited in the ovitraps were transferred to cartons with a moistened cup to collect deposited eggs. Medellin is a current location of mosquito releases for the World Mosquito Program. After PCR-screening to detect and remove females with Wolbachia, eggs from negative females (n = 11) were hatched. These F₁ progeny were reared to adults, mated, and blood fed to collect F₂ eggs. Those eggs were hatched and placed in containers of approximately 200 larvae/1 liter water and reared in an incubator at 27 °C and 80% RH. The resulting pupae were placed into 5-ml tubes and checked twice daily (10:00 and 15:00 h) to collect newly eclosed adults. Those adults were transferred to a bucket and labeled with the date of transfer “day 0” and supplemented with a moist pad containing 10% sucrose. The utilization of 10% sucrose rather than blood was chosen to prioritize the feasibility testing of the technique under relevant field conditions. The buckets were placed outside in an enclosure under ambient conditions representing natural resting sites for Ae. aegypti (6°17ʹN 75°34ʹW). A Hobo data logger was placed next to the container to record hourly temperature and relative humidity. A day 0 subset (n = 18) of females was removed and frozen at −80 °C in 2 sterile microcentrifuge tubes with 8–9 mosquitoes each. The same number of females were removed on days 3, 6, 10, 14, 18, and 22 and frozen at −80 °C. The collection period lasted from 18 June to 11 July 2022. Replicate 2 was performed following the same methods, and mosquitoes of days 0, 3, 6, 12, 14, 18, and 22 were collected from 24 July to 12 August 2022. The collected mosquitoes were frozen and shipped on dry ice from Colombia to our laboratory for analysis.

Degree Day Calculation

Data loggers recorded the temperature each hour over the mosquito holding period. Degree day was calculated by accumulating hourly thermal units above the developmental threshold for Ae. aegypti (14 °C) and dividing by 24 (Grech et al. 2015).

Sample Preparation

A similar sample preparation protocol (Wang et al. 2022) with some modifications was employed to analyze the water extract of mosquitoes. The key steps were shown in Fig. 1. Mosquitoes were thawed immediately before use. For sample preparation, 1 mosquito was immersed in 200 μl Milli-Q water and homogenized using an ultrasonic disruptor (Branson, Fisher Scientific, Waltham, MA) for 2 min on ice. The homogenates were centrifuged at 9,000×g for 5 min at 4 °C to separate the water-soluble components (middle layer) from the fat (top layer) and tissue debris (bottom layer). The middle layer (100 µl) was carefully collected. Then 5 µl of sample from the middle layer was mixed with 5 µl of a commercial AgNP colloidal solution (50 nm, citrate coating, NanoComposix) for 15 min and deposited onto an aluminum foil covered glass slide and air-dried for SERS analysis.

Fig. 1.

Flow of the experimental steps for collecting SERS spectra of mosquito water extract.

Instrumentation

SERS spectra were obtained using a DXRi Raman microscope (Thermo Fisher Scientific, Madison, WI) with a 780-nm laser and a 20× objective. The laser power was set to be 5 mW and the acquisition time was 0.5 s. Spectra collection on each dried water extract droplet was illustrated in Supplementary Fig. S1, where 5 different areas were randomly selected, and each area was scanned to obtain 6 × 6 = 36 spectra. Then the average of these 36 spectra was obtained. Therefore, for each mosquito, there were 5 spectra used for data analysis.

Data Analysis

Standard chemometric methods (e.g., PCA and PLS).

PCA and PLS were performed using TQ analyst software (Thermo Fisher Scientific). Data preprocessing, such as standard normal variate (SNV) and second derivative transformation with Norris derivative filter were applied to reduce variation, eliminate baseline effect, separate overlapped peaks, and enhance spectral resolution. For PCA, 10 PCs were used for calculation and 3D PCA plots were drawn to visualize the data separation in each age group. For PLS models, optimum factors were chosen based on the lowest root mean square error of prediction (RMSEP). About 75% of randomly selected data were used for model calibration and the remaining 25% were used for validation. In PCA, percentage of accuracy was calculated based on the validation set. Correlation coefficient R and RMSEP were calculated for evaluating the quality of the PLS model.

ANNs.

A simple multilayer perceptron (MLP) consisting of 2 hidden layers and 100 neurons as the feedforward ANN was used for mosquito age-grading based on the SERS spectra. The same preprocessing steps as in the standard chemometric methods are used, and models are built and trained using the PyTorch library in Python. The qualitative ANN model was calibrated using a cross-entropy loss function to classify mosquitoes according to their age into 7 categories. On the other hand, the ANN regression model for quantitative mosquito age analysis was built by replacing the loss function with a mean-squared-error loss. About 75% of randomly selected data were used for model calibration and the remaining 25% were used for validation. Both models are trained using the full-batch Adam optimizer (Kingma and Ba 2015) for 1,000 iterations with an initial learning rate set to 0.001. The percentage accuracy, R and RMSEP were calculated to evaluate the quality of the models.

Results

Spectral Characterization

As shown in Fig. 2, the spectral features of the field-aged F₂ female mosquitoes (field spectra) were similar but had more peaks identified when compared to SERS spectra of lab-reared female mosquitoes (lab spectra). Table 1 summarized the major characteristic peaks and their tentative assignments to either lab and/or field mosquitoes(Feng et al. 2015, Kubryk et al. 2016, Premasiri et al. 2016, Li et al. 2018, Shen et al. 2018). Both of them contained peaks at 655, 729, 958, 1,092, and 1,328 cm⁻¹, with the most distinct peaks being the adenine-containing compounds at 729 cm⁻¹. Additional peaks at 800, 839, 1,021, 1,248, 1,400 cm⁻¹ were identified in the field spectra, indicating the variation of protein profiles between lab and field mosquitoes. Compared to the lab spectra, where the difference between young (0–3 days), old (6–18 days), and very old (day 22) females were clearly visible, the difference between different ages were smaller in the field spectra, except at day 10 which exhibited the greatest difference from others, indicating a nonlinear feature of the spectra data for field-aged mosquitoes.

Fig. 2.

Average SERS spectra of (A) lab- (Wang et al. 2022) and field-aged (B) mosquito females of various ages (0–22 days).

Table 1.

Tentative peak assignments of the SERS spectra of lab and field mosquitoes (adopted from Feng et al. 2015, Kubryk et al. 2016, Premasiri et al. 2016, Li et al. 2018, Shen et al. 2018)

Raman shift (cm−1)	Lab	Field	Peak assignment
1,582	x	x	Guanine, adenine, tryptophan
1,447	x (days 6–22)		CH2/CH3 deformation
1,440		x	CH2 scissoring
1,328	X	x	Adenine
1,248		x	Amide III
1,134	x (days 0–3)		C–C/C–N stretching
1,092	X	x	C–C skeletal and C–O–C stretching from glycosidic link
1,046	x (day 22)		C–N stretching
1,021		x	Phenylalanine, C–H in-plane bending
1,003	X		Phenylalanine, ring breathing
958	X	x	Random coil
839		x	Tyrosine
800		x	Tryptophan
729	X	x	adenine
655	X	x	C–S stretching

The discrepancy between lab and field spectra may be due to from the physiological difference between the lab and field mosquitoes, which may have been influenced by different environmental factors during rearing (e.g., temperature and food) (Chown and Nicolson 2004). Degree day measurements are a way of standardizing aging of insects and other organisms under different environmental conditions (Gerade et al. 2004, Orlova-Bienkowskaja and Bieńkowski 2022). Supplementary Table S1 is a chart of the chronological day age and corresponding degree day ages for our experiment. It shows that the lab-reared mosquitoes were older than the field-aged mosquitoes, because the temperatures in the field were lower than in the laboratory. For example, a mosquito that was 10 days old in field (replicate 1) would be the same degree days as one that was between 6 and 7 days old from the lab reared set. This difference in temperature could explain why the overall difference between each age group in the field spectra was smaller than that in the lab spectra. Ntamatungiro et al. (2013) found that the physiological status of Anopheles gambiae also influenced the chronological age prediction by NIR, which suggested that older insect predictions were less accurate (Ntamatungiro et al. 2013).

Another factor that would impact the spectral characteristics is the sample preparation protocol. The key steps of the sample preparation are provided in Fig. 1. The timeframe of this procedure is approximately 40 min per sample, in which the drying time is the longest (~20 min per sample). As we usually handle a set of 8 samples at a time, the total time for a set of 8 samples is ~2 h. In this preliminary study, we tested the impact of postdrying time (the time after drying and before analysis: 0, 2, 12, 24, and 48 h at room temperature) on the slide, as this can be varied depending on how many samples in total we would like to process at a time. As shown in Fig. 3, the postdrying time had a significant impact on the spectral characteristics. Peaks at 729 and 1,328 cm⁻¹ which are adenine compounds were consistent over 48 h, while other peaks (e.g., 800, 958, 1,023, 1,400 cm⁻¹) which are associated with proteins were significantly decreased over time, potentially due to protein degradation. This result demonstrates the importance of controlling the postdrying time to be less than 2 h to obtain rich and consistent characteristic spectra.

Fig. 3.

SERS spectra (A) and representative peak intensity (B) of mosquito (day 6) water extract collected by varying the postdrying time from 0 to 48 h.

PCA/PLS Analysis

Chronological day analysis.

To appropriately compare the PCA and PLS analysis between lab and field-aged mosquitoes in a similar number, only spectra from replicate 1 (N = 135) were used. Compared to the analysis result of lab mosquitoes, the classification and quantification performance of field-aged mosquitoes was decreased. In brief (Fig. 4 and Table 2), we were able to discriminate between each age group (0, 3, 6, 10, 14, 18, and 22) using PCA with accuracy of 84% for field-aged mosquitoes, as compared to 96% for lab-reared mosquitoes. The reduced accuracy was resulted from the fact that the data variation within the age group was larger and the discrimination distance between different age groups was smaller as we can see from the data distribution in the PCA plot (Fig. 4). The PLS model demonstrated the capability of this approach to predict the age of field-aged mosquitoes between 0 and 22 days with R = 0.950 and RMSEP = 2.3 day and R = 0.914 and RMSEP = 3.56 day for PLS model calibration and validation, respectively, which was not as accurate as the performance of lab model with an error less than 1 day. Furthermore, when both replicates (replicate 1 and 2, total N = 390) were considered, the accuracy dropped significantly for both PCA and PLS, likely due to the large variation between these 2 replicates caused by environmental (e.g., temperature) variation (Supplementary Table S1).

Fig. 4.

PCA and PLS plots of lab (A) (Wang et al. 2022) and wild (B) mosquito samples of various ages.

Table 2.

Classification and quantification model (chronological day vs. degree day) performance for age-grading lab- and field-aged mosquitoes

Sample	Model	Classification		Quantification (chronological day)		Quantification (degree day)
		Total N	Accuracy (%)	Prediction R	RMSEP (day)	Prediction R	RMSEP (day)
Lab	PCA/PLS	144	96	0.990	0.93	0.991	13.8
Field (Rep 1)	PCA/PLS	135	84	0.914	3.56	0.942	24.1
Field (Rep 1 + 2)	PCA/PLS	390	69	0.812	3.90	0.772	44.4
Field (Rep 1 + 2)	ANNs	390	86	0.955	1.90	0.955	20.8

R is correlation coefficient and RMSEP is root mean square error of prediction.

Degree day analysis.

PLS models were re-established using degree day (Supplementary Tables S1 and S2) which represents temperature fluctuations. Approximately, 1 chronological day equals to 15 degree days. The results (Table 2) demonstrated no significant difference for the lab mosquito data but significantly improved for the field (Rep 1) data (RMSEP is 24.1 degree days which approximately equals to 1.6 chronological days). When both replicates were used, the RMSEP slightly improved from 3.90 chronological day to 44.4 degree days (which is equivalent to 2.96 chronological days), although the R was decreased. This result indicates the incorporation of temperature fluctuation improved the prediction accuracy. This is significant as degree days are what drives all behaviors (blood feeding) and physiology (blood digestion, egg development, survival) and with that, we can compare across field locations and regions where temperatures naturally fluctuate (Bellone and Failloux 2020).

Improved Age-Grading Accuracy Using ANNs

The performance degradation of the standard chemometric methods (PCA and PLS) when applied to both mosquito replicates suggests that the intricate nonlinear spectrum-age relationship for field-aged mosquitoes cannot be accurately captured by linear statistical models. Nonlinear machine learning models, on the other hand, have shown superior capability of quantitative and qualitative analysis of complex data matrices than standard linear chemometric methods such as PCA and PLS (Liebman et al. 2015, Siria et al. 2022). Herein, the performance of ANNs and PCA/PLS was compared when both were applied on all 390 spectra collected from replicates 1 and 2 using the PyTorch library in Python. Using a simple multilayer perceptron (MLP) consisting of 2 hidden layers and 100 neurons, our qualitative ANN model calibrated with a cross-entropy loss achieves a validation accuracy of 86%, significantly outperforming the 69% accuracy obtained by PCA. Moreover, after replacing the cross-entropy loss function with a mean-squared-error loss, the ANN regression model for quantitative mosquito age analysis achieves a RMSE of 1.9 days with R = 0.955 compared to RMSEP = 3.9 days and R = 0.812 for the traditional PLS (Table 2). When degree days were used for quantification, the RMSEP was further reduced to 20.8 degree days (which is equivalent to 1.4 chronological days). These highly promising results demonstrate the immense potential of ANNs to enhance our SERS-based mosquito age-grading accuracy.

Conclusion

In conclusion, we demonstrated the capability of SERS with ANNs to successfully age-grade field-aged F₂ female Ae. aegypti with high accuracy (error less than 2 days). The SERS spectra collected from field mosquitoes have larger variations within the age group and less discrimination between different age groups which significantly reduced the prediction accuracy using PCA and PLS. The improvement of the prediction accuracy was achieved by using 2 strategies. The first one is the use of degree day instead of chronological day which corrected the impact of temperature fluctuation. The second strategy is the use of nonlinear ANNs which significantly improved the accuracy of prediction models as compared to PCA and PLS. The SERS-based method is a substantial improvement over the current IR-based method as summarized in Table 3. In addition, the cost of age grading can be simpler and lower than other approaches based on cuticular hydrocarbon analysis or gene expression. Only 5 µl of AgNPs was needed to mix with the water extract, making the consumable cost per mosquito approximately 0.1 USD per mosquito. While our method requires some standard laboratory equipment (e.g., centrifuge and sonicator), these items are common in many laboratories. The greatest expense would be the Raman spectroscope. Handheld versions may be a more affordable option (approximately 9,000 USD at the time of publication). Our study clearly demonstrates the reliability of Raman spectroscopy combined with machine learning to age grade mosquitoes exposed to ambient field conditions. Future studies should determine the impact of mosquito blood feeding status and infection stage and other abiotic factors that may influence the SERS spectral characteristics to further improve the accuracy and robustness of age grading models. In the long term, we expect our method will facilitate epidemiological studies, risk assessment, vector intervention monitoring, and evaluation.

Table 3.

Comparison of different spectroscopic methods on the capability of age discrimination

Spectroscopic method	Age discrimination capability
NIR (10+ studies)	• Great accuracy for binary discrimination (<7 and >7 days) for lab-reared mosquitoes • Challenge for field mosquitoes
MIR (3 studies)	• Two studies showed binary classification (very young vs. very old) • One study reported ternary classification of days 1–4, 5–10, and 11–17
SERS (our study)	• Great quantitative grading (days 0–22) for both lab (error <1 day) and field-aged (error <2 days) mosquitoes