Predicting the age of field Anopheles mosquitoes using mass spectrometry and deep learning

Abstract

Mosquito-borne diseases like malaria are rising globally, and improved mosquito vector surveillance is needed. Survival of Anopheles mosquitoes is key for epidemiological monitoring of malaria transmission and evaluation of vector control strategies targeting mosquito longevity, as the risk of pathogen transmission increases with mosquito age. However, the available tools to estimate field mosquito age are often approximate and time-consuming. Here, we show a rapid method that combines matrix-assisted laser desorption/ionization-time-of-flight mass spectrometry with deep learning for mosquito age prediction. Using 2763 mass spectra from the head, legs, and thorax of 251 field-collected Anopheles arabiensis mosquitoes, we developed deep learning models that achieved a best mean absolute error of 1.74 days. We also demonstrate consistent performance at two ecological sites in Senegal, supported by age-related protein changes. Our approach is promising for malaria control and the field of vector biology, benefiting other disease vectors like Aedes mosquitoes.

Fig. 1.. Deep learning frameworks for predicting the age of Anopheles mosquitoes from MALDI-TOF mass spectra.

We evaluated two neural network models: a convolutional neural network (CNN; green) and a temporal convolutional network (TCN; brown). The CNN and TCN architectures differ in their convolutional filters. CNN uses filters that slide over the input data linearly, focusing on nearby elements, while TCN uses filters with variable stride, allowing for the detection of dependencies between non-adjacent elements. The TCN is particularly suited for capturing complex temporal dependencies. To evaluate the performance of the models and improve mosquito age estimation, we compared three age prediction techniques: conventional classification, rank-consistent classification, and regression. Unlike conventional classification, rank-consistent classification considers the chronological order of mosquito ages when assigning them to age categories, and the regression method predicts the age as a continuous real value in days.

Fig. 2.. Validating MALDI-TOF MS coupled with deep learning for predicting the age of field Anopheles mosquitoes.

(A) To establish a genetically variable dataset, field larvae of Anopheles arabiensis were collected from two ecologically distinct sites in Senegal: urban (Senegal 1, blue, n = 183 specimens) and rural (Senegal 2, pink, n = 68 specimens). The mosquitoes were reared in the laboratory under uncontrolled temperature and humidity while controlling for the age of the mosquitoes. (B) Unsupervised clustering using principal components analysis of MALDI-TOF mass spectra of mosquito anatomical parts (head, thorax, and legs), with the complete datasets. The spectra are plotted in a two-dimensional space, with colors indicating the geographical origin. (C and D) Confusion matrices showing the accurate classification (diagonal) of the thorax mass spectra into three age categories (0 to 3 days, 4 to 10 days, and 11 to 28 days) using a CNN and a conventional classification for Senegal 1 (C) and Senegal 2 (D). (E and F) Box and whisker plots showing the intensity variation of the seven most discriminative peaks [mass/charge ratio (m/z) values in daltons] from protein profiling of thorax mass spectra for each of the three age categories (0 to 3 days, 4 to 10 days, and 11 to 28 days) in Senegal 1 (E) and Senegal 2 (F) datasets. The line represents the mean intensity, while the whiskers indicate the SD.

Fig. 3.. Applying other deep learning prediction techniques to improve the accuracy of Anopheles mosquito age prediction.

We evaluated the performance of prediction techniques using a CNN and thorax mass spectra. (A and B) Confusion matrices showing the accurate classification (diagonal) of thorax mass spectra into three age categories (0 to 3 days, 4 to 10 days, and 11 to 28 days) using rank-consistent classification for Senegal 1 (A) and Senegal 2 (B). (C and D) Regression analysis showing the best-fit line (dashed line) and the correlation coefficient (R-squared, R²) between the predicted and ground truth mosquito age using a regression prediction for Senegal 1 (C) and Senegal 2 (D). The ideal regression line, where the prediction is equal to the true age, is represented by the continuous line. (E and F) Modified Bland-Altman plots illustrating the concordance between actual and predicted age using regression. The difference between the predicted age and the true age values is plotted against the true age values for Senegal 1 (E) and Senegal 2 (F). The dispersion of the residual values of the predicted ages is shown by the dotted lines representing the means ± 1.96 SDs. MAE, mean absolute error.

Fig. 4.. Comparing the performance of prediction techniques and neural network models for Anopheles mosquito age prediction.

We compared the performance using the area under the receiver operator characteristic (AUROC) as a common metric. (A and B) Box and whisker plots of AUROC values (line: mean, whiskers: SD) to compare the performance of conventional classification, rank-consistent classification, and regression for Senegal 1 (A) and Senegal 2 (B) using a CNN. (C and D) Box and whisker plots of AUROC values (line: mean, whiskers: SD) to compare the performance of the CNN and the TCN for Senegal 1 (C) and Senegal 2 (D) using the regression prediction technique.