Delimiting cryptic morphological variation among human malaria vector species using convolutional neural networks

Abstract Background Among mosquitoes that transmit human pathogens there are several species complexes, morphologically indistinguishable, distinct in their behavior and ecology, particularly as it relates to vectorial capacity. Morphological demarcation has remained largely intractable. Identification using molecular markers is the favored method in applied settings, prohibitively so, as the expense of such methods prevents the majority of mosquito surveillance programs to distinguish species within species complexes. Here, we apply computational methods in machine learning to delimit vector species within the Anopheles gambiae complex, which transmit human malaria and are notoriously difficult to distinguish based on morphology.Results We introduce a library of $1,709$ images collected from $15$ species colonies of mosquito vector species originating from five regions worldwide, with four sibling species from Africa that cannot be readily distinguished even amongst trained medical entomologists. Among them we include images of two strains of a single species, Anopheles gambiae s.s. . We varied storage method within a species to include two methods of storage, freshly frozen or dried, to test whether the same cage population could be distinguished by deep learning on this basis. We present a machine vision approach using Convolutional Neural Networks (CNN) with data augmentation that can achieve up to $97\%$ accuracy in distinguishing the genera, species, cryptic species, and between two strains of one species as well as storage method. We included three outgroup vector species from the Aedes and Culex genera known to transmit other human pathogens.Conclusion We demonstrate that deep learning models can delimit species with cryptic morphological variation, strains that are genetically differentiated but morphologically identical, and automate delineation of sex and storage method. These findings have broad reaching implications for the study of morphological variation and applications to the surveillance and study of human malaria.