Rapid age-grading and species identification of natural mosquitoes for malaria surveillance

Abstract

The malaria parasite, which is transmitted by several Anopheles mosquito species, requires more time to reach its human-transmissible stage than the average lifespan of mosquito vectors. Monitoring the species-specific age structure of mosquito populations is critical to evaluating the impact of vector control interventions on malaria risk. We present a rapid, cost-effective surveillance method based on deep learning of mid-infrared spectra of mosquito cuticle that simultaneously identifies the species and age class of three main malaria vectors in natural populations. Using spectra from over 40, 000 ecologically and genetically diverse An. gambiae, An. arabiensis, and An. coluzzii females, we develop a deep transfer learning model that learns and predicts the age of new wild populations in Tanzania and Burkina Faso with minimal sampling effort. Additionally, the model is able to detect the impact of simulated control interventions on mosquito populations, measured as a shift in their age structures. In the future, we anticipate our method can be applied to other arthropod vector-borne diseases.

Fig. 1. Experimental setup for capturing variation in MIRS caused by the laboratory of origin, individual genetic differences and natural environment.

To disentangle genetic and environmental effects, mosquitoes were obtained from either laboratory-bred colonies or from genetically heterogeneous wild larvae; half of the laboratory larvae were then reared and allowed to develop through the adult stage in semi-field conditions, which offer ecologically realistic conditions while still allowing control of mosquito age.

Fig. 2. Variation in MIRS, machine-learning model architecture, the sensitivity of the trained model.

We collected the MIRS of 41,151 female mosquitoes belonging to three species from diverse laboratories, genetic backgrounds, and environments and three age classes spanning 1-17 days post pupal emergence. a, b Unsupervised clustering of MIRS measurements using Uniform Manifold Approximation and Projection of MIRS in two dimensional space (plot axes) from An. arabiensis, An. coluzzii and An. gambiae coloured according to site of origin (a) and source of variation (b). c Representative variation of mid-infrared absorption spectra of An. arabiensis, An. coluzzii and An. gambiae and of three age classes. d Schematic representation of the deep convolutional neural network that takes MIRS inputs and outputs mosquito age and species. The input layer (wavenumber values) is fed through five 1-dimensional convolutional layers, comprising of 16 filters each (convolutional layers region), followed by a dense layer of 500 features and age and species output layers (dense layers) that were used to make predictions.

Fig. 3. Confusion matrices of model prediction accuracies and transfer-learning power.

DL-MIRS was trained using mosquitoes from laboratory larvae reared in the lab (LV, laboratory variation), larvae from the field reared in the lab (GV, genetic variation), and laboratory larvae reared in semi-field (EV, environmental variation). To improve model generalisation from lab to field-reared mosquitoes, we used transfer learning by freezing the convolutional layers of a model trained on LV+GV datasets only and calibrated using a smaller number of EV mosquitoes (here, 1294 examples) to train only the dense layers, resulting in highly accurate identification of (a) mosquito age and (b) mosquito species. c Classification accuracy improved from ~50% to 94% for both age group and species with a training set comprising 0 (i.e. effects of increasing sampling of lab-reared mosquitoes only) through 1452 semi-field (EV) mosquitoes used to re-train the transfer learned model. The solid and shaded lines indicate the mean and standard deviation of the mean of 20 trained models, respectively.

Fig. 4. Average model sensitivities to different wavenumber values and comparison with the features of the average absorption spectrum (grey line) of each output class.

The coloured stripes show the regions associated with the particular vibration of a functional chemical group. The upper part (maxima) displays the intervals of wavenumber values in which the maximum of the absorption peaks of each vibration appear for each of the three most abundant components in the cuticle of a mosquito. Here, the vibration of the same bonds appears in different wavenumber values depending on which cuticular component they belong to (chitin, protein or wax), which modifies the shape of the peaks.

Fig. 5. DL-MIRS generalisation and detection of vector control intervention.

a Computer simulations were used to assess the power of DL-MIRS a 'rapid kill' (long-lasting insecticide-treated nets; LLIN) or ‘slower kill’ intervention (attractive toxic sugar baits; ATSB) relative to a population with no intervention (control). b, c Power to detect an effect of the vector control intervention was estimated over 10 levels of training set size represented by coloured points, with EV mosquitoes ranging from 0 to 1452 and seven sample sizes per population from 20 to 300 (Supplementary Table 5). The dashed red line shows the power that would be achieved with 100% accurate age group classification. The difference between the solid and dashed lines represents the cost in power due to prediction error.

Fig. 6. DL-MIRS validation on wild mosquito populations.

The proportion of wild female mosquitoes with 0, 1 or ≥2 gonotrophic cycles (G0, G1, G2+G3+G4) was determined by ovarian dissection and morphological characterisation (yellow) or predicted by DL-MIRS on non-dissected mosquitoes (blue). The same number of mosquitoes for each group was analysed on each day of collection, either from Burkina Faso (a, An. coluzzii) or from Tanzania (b, An. arabiensis). The mean proportions and 95% credible intervals of the age proportion from dissected mosquitoes (yellow) were estimated with a Dirichlet distribution. The age proportion predicted by the DL-MIRS (blue) is presented as box-whisker plots showing the median, interquartile range (IQR, box), lowest/highest data within 1.5 IQR (whiskers), and outliers (red points) of the probability distribution of predictions from ten different models.