# BiteOscope, an open platform to study mosquito biting behavior

1. Department of Bioengineering, Stanford University, United States
2. Insect-Virus Interactions Unit, Institut Pasteur, UMR2000, CNRS, France
3. Center for research and Interdisciplinarity, U1284 INSERM, Université de Paris, France
Tools and Resources

## Abstract

Female mosquitoes need a blood meal to reproduce, and in obtaining this essential nutrient they transmit deadly pathogens. Although crucial for the spread of mosquito-borne diseases, blood feeding remains poorly understood due to technological limitations. Indeed, studies often expose human subjects to assess biting behavior. Here, we present the biteOscope, a device that attracts mosquitoes to a host mimic which they bite to obtain an artificial blood meal. The host mimic is transparent, allowing high-resolution imaging of the feeding mosquito. Using machine learning, we extract detailed behavioral statistics describing the locomotion, pose, biting, and feeding dynamics of Aedes aegypti, Aedes albopictus, Anopheles stephensi, and Anopheles coluzzii. In addition to characterizing behavioral patterns, we discover that the common insect repellent DEET repels Anopheles coluzzii upon contact with their legs. The biteOscope provides a new perspective on mosquito blood feeding, enabling the high-throughput quantitative characterization of this lethal behavior.

## eLife digest

Scientists often sacrifice their own skin to study how mosquitos drink blood. They allow mosquitos to bite them in laboratory settings so they can observe the insects’ feeding behavior. By observing blood feeding, scientists hope to find ways to prevent deadly diseases like malaria, which is transmitted by bites from mosquitos carrying the malaria parasite. These studies are not only unpleasant for the volunteers, they also have important limitations. For example, it is too risky to use pathogen-infected mosquitos that could make the volunteers sick.

A device called the biteOscope developed by Hol et al. may give scientists and their skin a reprieve. The device has a transparent skin-like covering that attracts mosquitos and supplies them an artificial blood meal when they bite. The device captures high-resolution images of the insects’ behavior. It is small enough to fit in a backpack when disassembled, costs about \$900 to \$3,500 US dollars, and is suitable for use in the laboratory or in the field. Using machine-learning techniques, Hol et al. also developed an automated system for analyzing the images.

The researchers tested the device on four types of disease-transmitting mosquitos. In one set of experiments, Anopheles mosquitos were recorded interacting with a biteOscope partially coated with an insect repellent called DEET. The images captured by the biteOscope showed that the mosquitos are attracted to the warm surface and land on the part coated with DEET. But when their legs come in contact with the repellent, they leave.

The biteOscope provides scientists a new way to study blood feeding, even in mosquitos infected with dangerous pathogens. It might also be used to test new ways to prevent mosquitos from biting and spreading disease. Because the device is portable and relatively inexpensive, it may enable larger studies in a variety of settings.

## Introduction

Blood feeding is essential for the reproduction of many mosquito species, and in the process, mosquitoes transmit myriad pathogens to their (human) host. Yet, despite being the focal point of pathogen transmission, many aspects of blood feeding remain ill understood. The initial step in obtaining a blood meal, flying toward a host, is relatively well characterized (Dekker and Cardé, 2011; McMeniman et al., 2014; van Breugel et al., 2015). The steps that unfold after a mosquito has landed on a host, however, are much less understood. Once landed, mosquitoes exhibit exploratory bouts during which the legs and proboscis frequently contact the skin (Jones and Pilitt, 1973; De Jong and Knols, 1995; Clements, 2013). An increasing body of literature reports the presence of receptors involved in contact-dependent sensing on the legs and proboscis (Sparks et al., 2013; Matthews et al., 2019; Dennis et al., 2019), suggesting that these appendages evaluate the skin surface and thus serve an important role in bite-site selection. Yet, the role and mechanism of contact-dependent sensing in blood feeding is largely unclear (Benton, 2017). In addition to the body parts that come in contact with the skin surface, the skin piercing labrum also serves as a chemosensory organ, guiding blood feeding in currently unknown ways (Lee, 1974; Werner-Reiss et al., 1999; Jove et al., 2020).

In addition to external cues, an animal’s (internal) physiology may also affect its behavior. Nutrition, hydration, and pathogen infections, for instance, have been hypothesized to affect blood feeding behavior, for?example by altering feeding avidity (i.e. number of feeding attempts) or the size of the meal taken (Rossignol et al., 1984; Choumet et al., 2012; Cator et al., 2013; Vantaux et al., 2015; Hagan et al., 2018). These topics, however, remain a matter of debate, due to a lack of (standardized) assays to measure mosquito behavior (Stanczyk et al., 2017). Quantitative mapping of Drosophila behavior provides an important perspective, suggesting that innovative experimental approaches and computational tools can fuel the acquisition of new insights (e.g. Branson et al., 2009; Kain et al., 2013; Berman et al., 2014; Corrales-Carvajal et al., 2016; Robie et al., 2017; Moreira et al., 2019). Yet, apart from olfactometers and other flight chambers, very few assays to characterize the blood-feeding behavior of mosquitoes exist (Geier and Boeckh, 1999; Verhulst et al., 2011; McMeniman et al., 2014; van Breugel et al., 2015; Murray et al., 2020). Due to this paucity of assays, studies often expose human subjects to quantify the number of landings and/or bites, or the time it takes to complete a blood meal, and score experimental outcomes by hand (Jones and Pilitt, 1973; Ribeiro, 2000; Moreira et al., 2009; DeGennaro et al., 2013; Dennis et al., 2019; Hughes et al., 2020). The use of humans as bait constrains the number and type of experiments that can be done (e.g. prohibiting the use of infected mosquitoes) and limits the type, detail, and throughput of measurements that can be made. Furthermore, the opaque nature of skin prevents the visualization of the stylets after piercing the skin leaving this aspect of blood feeding almost entirely unstudied, except for one notable study using intravital imaging of dissected mouse skin (Choumet et al., 2012) and two much earlier descriptions (Gordon and Lumsden, 1939; Griffiths and Gordon, 1952).

To overcome these limitations, we developed the biteOscope, an open platform that allows the high-resolution and high-throughput characterization of surface exploration, probing, and engorgement by blood-feeding mosquitoes. The biteOscope consists of a rudimentary skin mimic: a substrate that attracts mosquitoes to its surface, induces them to land, pierce the surface, and engage in blood feeding. The bite substrate can be mounted in the wall of a mosquito cage allowing freely behaving mosquitoes access. By virtue of its transparent nature, the substrate facilitates imaging of mosquitoes interacting with it, including the visualization of the skin piercing mouthparts of the mosquito. We developed a suite of computational tools that automates the extraction of behavioral statistics from image sequences, and use machine learning to track the individual body parts of behaving mosquitoes. These capabilities enable a detailed characterization of blood-feeding mosquitoes. We demonstrate that the biteOscope is an effective instrument to study the behavior of several medically relevant species of mosquito and describe behavioral patterns of the two main vectors of dengue, Zika, and chikungunya virus (Aedes aegypti and Aedes albopictus), and two important malaria vectors (Anopheles coluzzii and Anopheles stephensi). The biteOscope allows detailed tracking of the complex interactions of mosquitoes with a substrate and can be used to characterize behavioral alterations in the presence of chemical surface patterns. Using this capability, we provide evidence that DEET repels Anopheles coluzzii upon contact with their legs, demonstrating the utility of body part tracking to understand behaviors mediated by contact-dependent sensing. We anticipate that the biteOscope will enable studies that increase our understanding of the sensory biology and genetics of blood feeding, and the effects external (environmental) and internal (physiology) variables have on this behavior. Given its relevance for pathogen transmission, dissecting the interplay between the mosquito sensory system and host-associated cues during blood feeding is of clear interest, and may suggest new avenues to interfere with blood feeding, and eventually curb pathogen transmission.

## Results

### The biteOscope

To allow mosquitoes to engage in blood feeding and feed to full repletion, a device needs to attract mosquitoes, allow them to explore and pierce the surface, and subsequently imbibe a blood meal. To design a tool that can easily be used in a variety of ‘mosquito labs’ (including (semi-)field settings), we sought to recapitulate this behavioral sequence using readily available and low-cost laboratory materials. Heat is a dominant factor in short-range mosquito attraction and can be used to attract mosquitoes to a surface and elicit probing behavior (Healy et al., 2002; Corfas and Vosshall, 2015; Zermoglio et al., 2017; Greppi et al., 2020). We constructed a bite substrate using an optically clear flask filled with water as a controllable heat source (see Figure 1A). An artificial blood meal is applied on the outside of the flask and covered using Parafilm (a commonly used membrane in laboratory blood feeders) creating a thin fluid cell on which mosquitoes can feed (see Figure 1—figure supplement 1). To elicit blood feeding in a transparent medium, we use adenosine triphosphate (ATP) as a strong phagostimulant, which, together with an osmotic pressure similar to that of blood and the presence of sodium ions, is sufficient to induce Aedes mosquitoes to feed to full engorgement (Galun et al., 1963; Duvall et al., 2019). Anopheles also require sodium ions and a tonicity similar to blood to feed to full engorgement, but interestingly their feeding rate on artificial meals is independent of ATP (Galun et al., 1985).

Figure 1 with 2 supplements see all

To allow freely behaving mosquitoes access to the bite substrate, we constructed acrylic cages having an opening in the wall or floor where the bite substrate can be mounted. The bite substrate is transparent, facilitating imaging with a camera mounted outside the cage (Figure 1A shows a schematic of the set up). For the majority of data presented here, we used a 4.3 × 4.3 cm field of view (see Figure 1C) which allows up to 15 mosquitoes to explore and feed simultaneously while providing images at a resolution where small body parts like the stylets can easily be resolved. Depending on experimental requirements, the field of view (and correspondingly assay throughput) can be much larger at the expense of resolution. Figure 1B, for example, shows a 13 × 13 cm field of view. Individual mosquitoes can be easily tracked at that resolution, yet the visualization of small body parts is challenging. Experiments on Ae. aegypti and Ae. albopictus, both active during the day, were performed using white light illumination; we used an infrared (IR) LED array as light source during experiments on An. coluzzii and An. stephensi which were performed in the dark, corresponding to their peak activity during the night. Figure 1B demonstrates that Ae. aegypti mosquitoes show strong attraction to the bite substrate (surface indicated using a dashed line) and spend more time on its surface compared to the surrounding wall. Figure 1C–F shows Ae. aegypti undertaking the full blood feeding trajectory on the substrate: starting with surface exploration (Figure 1C and G), piercing of the membrane and insertion of the stylet into the artificial meal (Figure 1D–F), and feeding to full engorgement, as evidenced by the expanded abdomen (Figure 1E). Videos 1, 2, 3 and 4 show blood feeding Ae. albopictus, Ae. aegypti, An. stephensi, and An. coluzzii, respectively. Imaging the stylet (Videos 1 and 5) as it evaluates the artificial meal reveals the striking dexterity of the organ as it rapidly bends, extends, and retracts—aspects of feeding that normally remain hidden inside the skin.

Video 1
Video 2
Video 3
Video 4
Video 5

### Automatic characterization of the blood-feeding behavior of multiple species

We created a computational pipeline to extract behavioral statistics from image sequences (see Figure 1—figure supplement 2 for an overview and Materials and methods for details). The position of individual mosquitoes is tracked over time to yield locomotion statistics (see Figure 1G and Video 6), and select all time slices that make up a single behavioral trajectory (e.g. landing, exploration, feeding, and take off). The error rate of tracking was 0.045 (5 errors in a validation data set of $n=111$ tracks, see Materials and methods for details) with the majority of errors arising from erroneously assigned identities when two mosquitoes cross. Validation videos (see Video 7 for an example) make it straightforward to manually correct such errors yielding near-perfect tracking. To determine a mosquito’s engorgement status, we take advantage of the dilation of the mosquito abdomen when it takes a blood meal (Figure 1E). We determine a mosquito’s body shape (excluding appendages) using an active contour model to quantify feeding dynamics and engorgement status at each timepoint of a trajectory, and detect full engorgement with a sensitivity of 81% and a specificity of 100% (see Figure 1 G1-3, Video 8, and Materials and methods for details). Together with locomotion statistics, engorgement data provides a high-level description of the behavioral trajectory.

Video 6
Video 7
Video 8

To assess the capability of the biteOscope to characterize the behavior of different species of mosquito, we performed experiments with the two most important vectors of arboviral diseases (Ae. aegypti and Ae. albopictus) and two dominant malaria vectors (An. stephensi and An. coluzzii, formerly known as Anopheles gambiae M molecular form). Figure 2 and Figure 2—figure supplement 1 show locomotion and feeding statistics for the four species. All species land readily on the bite substrate and undertake exploratory bouts leading to full engorgement in 18%, 7%, 4%, and 14% of all trajectories and 46%, 22%, 10%, and 31% of all >10 second trajectories, for Ae. aegypti, Ae. albopictus, An. stephensi, and An. coluzzii, respectively, when offered a meal consisting of 1 mM ATP in phosphate buffered saline (PBS). Figure 2A–D shows summary statistics of 349 behavioral trajectories of An. coluzzii obtained from a total of 1 hr and 15 min of imaging data (five 15-min experiments with 15 females per experiment), demonstrating the throughput of the biteOscope.

Figure 2 with 1 supplement see all

Figure 2E shows the time spent on the surface versus the distance covered for trajectories that did (large opaque circles) and did not (small transparent dots) lead to full engorgement for the four species. As expected, rather short trajectories do not lead to engorgement, yet less intuitive is the observation that exploratory trajectories that do not lead to engorgement rarely exceed the duration of successful feeding trajectories (8% of non-feeding trajectories takes longer than the mean time to engorge). This suggests that a mosquito’s search for blood has a characteristic timescale that is independent of success, and when blood is not found within the time a typical meal takes, the search is aborted.

We further explored this observation using individual Ae. albopictus which were offered a bite substrate with a meal of PBS with or without ATP. As PBS alone does not lead to engorgement, mosquitoes offered the PBS only feeder never engorged whereas mosquitoes interacting with the PBS + ATP feeder engorged to full repletion in the majority of cases (55%). High-resolution trajectory analysis enables us to dissect behavioral patterns that lead to (non-)feeding; a trajectory here is defined as landing, the ensuing behavioral sequence, followed by leaving the bite substrate by walking or flying (see Videos 9 and 10 for two example trajectories). The velocity of a mosquito’s centroid can be used to classify locomotion behaviors (stationary, walking, flight) with high accuracy (89% see Figure 3—figure supplement 1 and Materials and methods for details). Figure 3 presents ethograms of Ae. albopictus on these two bite substrates, and in agreement with the data in Figure 2E, shows that trajectories on feeders without ATP (non-feeding) have an approximately equal maximum duration as trajectories leading to full engorgement on the feeder with ATP. While mosquitoes do not increase the duration of exploratory trajectories when not feeding to repletion, the number of exploratory bouts mosquitoes undertook on the PBS only substrate was significantly higher compared to the PBS + ATP case (Wilcoxon rank-sum test p?<?0.05), resulting in a slightly longer total exploration time (Figure 3C). This suggests that mosquitoes not finding their desired resource increase the frequency with which they initiate searches rather than the duration of individual searches. This observation may be interpreted in the context of the dangers associated with blood-feeding: while on a host, a mosquito runs the risk of being noticed and subsequently killed. When not finding blood, it may therefore be beneficial to abort the search and evacuate from a risky, yet unproductive situation to try elsewhere. The trade-off between exploiting a potential resource versus exploring other options has been shown to depend on the internal state of individuals in other insects (Katz and Naug, 2015; Corrales-Carvajal et al., 2016), it is possible that such mechanisms play a role here too. Figure 3 furthermore shows a strong behavioral heterogeneity between individual mosquitoes. While all individuals are from the same mosquito population (and raised and maintained under identical conditions) and interact with the same bite substrate, there is a clear heterogeneity in the number of times a mosquito visits the surface (Figure 3C, middle panel), the amount of time she spends exploring the surface (Figure 3C, left panel), and the behaviors they engage in. Automatic classification of locomotion behaviors, shows that some individuals often land on the surface to engage in short interactions, while other individuals undertake much longer trajectories. These long trajectories, in turn, vary in the amount of stationary versus locomotion behaviors. The richness of these data highlight the potential of the biteOscope to quantitatively characterize the intricacy of individual behaviors hidden in population averages.

Figure 3 with 1 supplement see all
Video 9
Video 10

### Pose estimation, behavioral classification, and contact-dependent sensing

We next turned to body part tracking to acquire a more detailed description of behavioral trajectories. Body part tracking is powerful to address a variety of questions, for?example by determining points of surface contact of specific appendages, or to estimate the pose of an animal, which when tracked over time can be translated into a behavioral sequence. We used a recently developed deep learning framework, DeepLabCut (Mathis et al., 2018), to train a convolutional neural network (CNN) to detect the head, proboscis, abdomen, abdominal tip, and six legs of Ae. aegypti and Ae. albopictus. Due to their morphological similarity, the same CNN can be used to track the body parts of both Aedes species with a mean accuracy of 11 pixels (275 micrometer, see Materials and methods for details) in a 4.3 × 4.3 cm field of view. Tracking stylet insertions into the artificial meal during probing and feeding using DeepLabCut was challenging, and therefore not included.

Figure 4A–C shows body part tracking results of Ae. albopictus and reveals the choreography of three distinct behaviors. Anterior grooming is characterized by circular motion of the forelegs followed by the proboscis, while the middle legs remain stationary (see Figure 4—video 1). During walking, the tips of all six legs oscillate along the body axis while the proboscis explores laterally (see Figure 4—video 2), while during probing, the fore and middle legs pull toward the body and the proboscis remains stationary (see Figure 4—video 3). Inference is done on raw images and the obtained coordinates thus subject to movement of the mosquito. To correct for this, the coordinates are translated and rotated to align along the body axis taking the abdominal tip as the origin. Figure 4D–I shows time series of the obtained egocentric coordinates and their corresponding wavelet transforms. The three behaviors each are associated with distinct periodic movements: smooth periodic motion of the forelegs during anterior grooming (x, and y coordinates), punctuated oscillations along the body axis during walking (x coordinate), and faster jerky movement during probing (x, and y coordinate of forelegs, y coordinate of middle legs). These trajectories can be used in concert with locomotion and body-shape features as inputs for behavioral classification algorithms. The data outputted by our computational pipeline is ideally suited for classification in either a supervised (e.g. Kain et al., 2013; Kabra et al., 2013) or unsupervised (e.g. Berman et al., 2014; Marques et al., 2018; Calhoun et al., 2019; Tao et al., 2019) approaches (see Figure 4—figure supplement 1).

Figure 4 with 4 supplements see all

#### DEET repels An. coluzzii upon contact with legs

Next, we explored the use of body part tracking within the context of contact-dependent sensing by An. coluzzii. Anopheles and Aedes mosquitoes have an overall similar body plan, yet the length of their maxillary palps (an olfactory appendage projecting from the head) is very different with anophelines having maxillary palps with a length comparable to the proboscis, while Aedes palps are much shorter. We therefore trained a CNN for Anopheles body parts, which additionally tracks the position of the maxillary palps (mean accuracy for Anopheles body parts: six pixels, 150 micrometer). Through this approach, we addressed the open question if An. coluzzii is repelled by N,N-diethyl-meta-toluamide (DEET) upon contact. DEET has been in use as an effective insect repellent for decades and is thought to act on mosquitoes through several mechanisms that are either olfactory- or contact-based (DeGennaro, 2015). Afify et al., 2019 recently observed that volatile DEET does not activate olfactory neurons in An. coluzzii and reported that An. coluzzii does not avoid DEET by smelling it (Afify et al., 2019; Afify and Potter, 2020). Afify et al., 2019 proposed that DEET may prevent An. coluzzii from locating humans by masking odorants emanating from potential hosts. However, it remained an open question if An. coluzzii is repelled by DEET upon direct contact.

We addressed this question by imaging An. coluzzii offered a bite substrate partly coated with DEET. Figure 5 shows that An. coluzzii do land on both the DEET-coated and uncoated surface, and there is a moderate decrease in landing rate on the DEET-coated portion (the landing rate is 1.9 times lower, normalized for surface area). The time An. coluzzii spend on the DEET-coated surface, however, is much shorter: trajectories on the DEET-coated surface ($n=34$) are on average seven times shorter when compared to the uncoated surface ($n=412$). Furthermore, the longest residence time observed on the DEET-coated surface was less than 6 s, whereas individual An. coluzzii spent up to 52 s on the uncoated surface. From these data, we conclude that An. coluzzii do approach and land on the DEET-coated surface, but avoid (prolonged) contact with it, indicating that An. coluzzii indeed is not strongly repelled by volatile DEET at very close range, yet avoids it on contact.

Figure 5

We next asked what appendages mediate this contact dependent avoidance. The 34 trajectories in which An. coluzzii visited the DEET area consisted of 25 ‘touch and go’ events in which an individual approached the DEET surface in flight, landed, and immediately took off after first contact (residence time on DEET surface <0.5 second, see Video 11 for a typical ‘touch and go’ event played at 1/4 speed). In the remaining nine trajectories, An. coluzzii landed outside the DEET area and moved onto it (see Figure 5 and Video 12), the reverse scenario in which a mosquito would land on the DEET surface and move onto the non-coated surface was never observed. We performed body part tracking on the trajectories where An. coluzzii moved from the non-coated surface to the DEET-coated surface and developed analysis software that scores how often a specific body part visits an arbitrarily shaped region of interest. We observed that the legs of individuals came in contact with the DEET surface in all cases, whereas the proboscis only came in contact with the DEET surface in 5/9 cases (in cases where no proboscis contact was observed, the entire proboscis remained outside the boarders of the DEET-treated area). Together, these observations demonstrate that An. coluzzii are indeed repelled upon contact with DEET, and indicate that this behavior is mediated by sensilla on the legs, and likely not the proboscis. While contact-dependent sensing (e.g. by tarsal neurons) seems the most plausible mechanism to explain this contact-dependent avoidance, we cannot rule out that physical properties of the DEET coating play a role as well.

Video 11
Video 12

## Discussion

The biteOscope provides an alternative for current methods using human subjects or mice to study mosquito blood feeding. The elimination of the need for a human subject opens new avenues of research, for?example allowing blood-feeding studies with pathogen-infected mosquitoes, enabling precise surface manipulations and characterization of the associated behavior, and facilitates the use of high-resolution imaging and machine-learning-based image analysis. Through these innovations, the biteOscope increases experimental throughput and expands the type of experiments that can be performed and measurements that can be made. We developed computational tools that allow the behavioral monitoring of mosquitoes at an unprecedented level of detail. Behavioral research on other animals, including fruit flies (Werkhoven et al., 2019; Pereira et al., 2019) and zebrafish (Marques et al., 2018; Johnson et al., 2020) shows that high spatiotemporal resolution data describing the posture of animals can be very informative to dissect behavioral trajectories and compare behavioral statistics across individuals and experimental treatments. While the details of computational approaches differ, a common theme is the two dimensional embedding of a high-dimensional representation of an animal at a given time point (e.g. body part coordinates and derived features), data points in two dimensions can subsequently be clustered to reveal behavioral classes (see Figure 4—figure supplement 1 for an illustration of this concept using tSNE to embed the data represented in Figure 4). Translating such advances in computational ethology to mosquito research is a very promising avenue for future research.

We used the biteOscope to describe behavioral patterns of four medically relevant mosquito species and anticipate that such datasets will provide a useful ‘behavioral baseline’ for future studies quantifying the effect of a mosquito’s physiology on blood feeding behavior. The role of pathogen infections is particularly interesting in this respect, as infections may alter feeding behavior, for?example by affecting the structural integrity of the salivary glands or other tissues, or inducing systemic change through the immune system or infection of neural tissues (Rossignol et al., 1984; Girard et al., 2007; Cator et al., 2013; Turley et al., 2009). A quantitative understanding of such behavioral alterations, however, is lacking. Gaining such insights is of high epidemiological relevance, as mathematical models suggest that (pathogen induced) changes in bite behavior can have important implications for pathogen transmission (Cator et al., 2014; Abboubakar et al., 2016). In addition to pathogen-induced behavioral changes, there are many other promising lines of inquiry, including the behavioral influence of the microbiome (Dickson et al., 2017), which, in other insects such as Drosophila, influences locomotor behavior (Schretter et al., 2018) and food choice (Leit?o-Gon?alves et al., 2017; Wong et al., 2017). Drosophila research furthermore shows interesting examples of collective behaviors mediated by for?example olfaction or direct contact between animals (Schneider et al., 2012; Ramdya et al., 2015; Lihoreau et al., 2016; Ramdya et al., 2017), it would be interesting to explore if mosquitoes also take advantage of collective intelligence when searching for food or avoiding noxious stimuli. Tools enabling high-throughput behavioral monitoring may also be useful to characterize population intervention strategies aimed at curbing pathogen transmission, such as Wolbachia infected Ae. aegypti, or Anopheles genetically engineered to be refractory to P. falciparum infection. Quantifying the behavioral effects of such interventions is an important step toward assessing the competitiveness of engineered mosquitoes in the field. As the biteOscope enables novel high-throughput experiments with a variety of mosquito species, we anticipate that it will prove useful for the characterization of various behaviors relevant to pathogen transmission.

By tracking the individual body parts of An. coluzzii,?we discovered that they are repelled by DEET upon leg contact—a mechanism that may work in concert with other ways in which DEET prevents anopheline mosquitoes to locate humans. Our findings regarding An. coluzzii are in agreement with observations in Ae. aegypti which are also repelled by DEET upon leg contact (DeGennaro et al., 2013; Dennis et al., 2019). However, in contrast to An. coluzzii, olfactory neurons of Ae. aegypti are activated by volatile DEET (Davis and Rebert, 1972; Boeckh et al., 1996; Stanczyk et al., 2010) and Ae. aegypti has been reported to avoid volatile DEET in recent studies (Stanczyk et al., 2013; Afify and Potter, 2020) (in contrast, an earlier study reported attraction of Ae. aegypti by DEET [Dogan et al., 1999]). Together, these observations suggest that contact-based repellency may be conserved across Anopheles and Aedes mosquitoes and thus may be a potentially interesting target for the design of new repellents. It is less clear, however, what degree of conservation exists for the olfactory modes of action, as the only study comparing the olfactory effects of volatile DEET on Anopheles and Aedes mosquitoes in the same assay, suggests that the former is not repelled at all by volatile DEET, while the latter showed moderate repulsion (these behavioral responses may be concentration dependent) (Afify and Potter, 2020). This observation, together with the observation that volatile DEET activates olfactory neurons in Ae. aegypti while it does not seem to do this in An. coluzzii, suggest that volatile DEET may modulate the response of olfactory neurons to attractive stimuli (‘scrambling of the odor code’ [Pellegrino et al., 2011]) and/or trigger repulsion in Ae. aegypti, while these mechanisms seem less appropriate for An. coluzzii. In addition to effects on olfactory signaling, DEET has also been suggested to decrease the amount of volatile odorants emanating from hosts through chemical interactions between DEET and the odorants resulting in the masking of a host (Afify et al., 2019). As in this scenario the amount of attractive odorants reaching a mosquito is reduced, it may affect the behavior of a variety of species. The observation that both Ae. aegypti and An. coluzzii avoid DEET upon leg contact, while the effects of volatile DEET may partly overlap and partly differ, may guide efforts aimed at uncovering the underlying molecular mechanisms.

Our results highlight the use of body part tracking in assigning roles to the various sensory appendages the mosquito body has. The recent surge in genetic tools available to manipulate mosquitoes is shedding light on the genetic elements that mediate pathogen transmission relevant behaviors (Matthews et al., 2019; Ingham et al., 2020; Raji et al., 2019; Greppi et al., 2020). Combining such molecular level insights with detailed behavioral tracking and chemical surface patterning, may enable a deep understanding of how contact-dependent sensing drives blood feeding, and other important phenotypes such as insecticide resistance and egg laying preferences.

When studying animal behavior in the lab a trade-off exists between the level of experimental control and detail of observation on the one hand, and an accurate representation of natural conditions and behaviors on the other. In case of the biteOscope, an engineered bite substrate opens up a variety of possibilities including surface modifications and high-resolution imaging impossible on human skin, yet the bite substrate does not offer the full set of cues (and thus behavioral responses) a human host would. It would therefore be interesting to add more human-associated cues, for instance using materials that resemble the texture of skin, or by coating the bite substrate with attractive human odorants (Okumu et al., 2010). In addition to more closely mimicking human hosts by presenting olfactory stimuli, surface coatings could be used to dissect the role of contact-dependent gustatory behaviors on the skin surface in bite site selection. It is important to note that many of the factors that may change behavior mentioned above (e.g. infections/nutritional status or components of the microbiome) are best assessed in a relative manner, for?example comparing non-infected to infected individuals. Comparing cohorts of mosquitoes undergoing different experimental treatments puts less emphasis on the absolute attractiveness of the bite substrate and thus mitigates potential issues related to the fact that a synthetic bite substrate is likely less attractive than a real live host.

We took advantage of the possibility to elicit engorgement on a transparent meal to facilitate imaging. It seems feasible to add a dye to the meal to provide visual cues to the mosquito without interfering with image quality. Using whole blood, however, is challenging in the current system. It would therefore be worthwhile to explore the use of microfluidics to incorporate blood flow into the bite substrate while maintaining optical access. A recent study took advantage of the biteOscope to quantify stylet contact with artificial meals (Jove et al., 2020), combining such efforts with artificial vasculature presents exciting opportunities to characterize the role of the stylets in the search for blood.

The biteOscope is designed with a variety of possible users in mind. It has a relatively modest price tag (900–3500 USD depending on the configuration), uses readily available materials and components, and when disassembled fits in a backpack—characteristics we hope will facilitate adoption. Beyond the lab, we foresee interesting applications of the behavioral tracking of mosquitoes in (semi-)field settings, and expect that innovative tools that provide high-quality quantitative data will enable discoveries in this space. We anticipate that the techniques and computational tools presented here will provide a fresh perspective on mosquito behaviors that are relevant to pathogen transmission, and enable researchers to gain a detailed understanding of blood feeding without having to sacrifice their own skin.

## Materials and methods

Key resources table

### Mosquito rearing and maintenance

Request a detailed protocol

The mosquito species/strains used in this study are described in Key resources table. Larvae were hatched and reared in water at a density of approximately 200 larvae per liter on a diet of fish food. Adult mosquitoes were maintained at $28?°C$, 75% relative humidity, and a photoperiod of 12 hr light : 12 hr dark in 30 × 30?×?30 cm screened cages having continuous access to 10% sucrose. Prior to experiments, mosquitoes were deprived of sucrose for 6–12 hr while having access to water. Mosquitoes aged 6–25 days old were used for behavioral experiments. Experiments using Ae. aegypti and Ae. albopictus were performed during light hours, while experiments with An. stephensi and An. coluzzii were performed during dark hours. Mosquitoes had no access to water during experiments.

### biteOscope hardware

A full list of components necessary to build the biteOscope is available in Appendix 1—table 1. Depending on the experimental requirements, several components can be easily adapted (e.g. cage geometry or bite substrate) or replaced by more economical alternatives (e.g. imaging components).

#### Cage, bite substrate, and environmental control

Request a detailed protocol

#### Imaging and illumination

Request a detailed protocol

Images were acquired at 25 or 40 frames per second using a Basler acA2040-90um camera controlled using Pylon 5 software running on an Ubuntu 18.04 computer (NUC8i7BEH). The camera was equipped with a 100 mm macro lens (Canon macro EF 100 mm f/2.8L). Illumination for Aedes experiments was provided by two white light LED arrays (Vidpro LED-312), while IR LEDs (Taobao) were used for Anopheles experiments. The same camera was used for white light and IR illuminated experiments. Thorlabs components were used to arrange all optical components and the experimental cage at suitable distance.

### Computational tools

All image processing and downstream analysis code was written in Python 3 and is available from Github (https://github.com/felixhol). Raw images were background subtracted, thresholded, and subjected to a series of morphological operations to yield binary images representing mosquito bodies of which the center of mass was determined using SciPy (Virtanen et al., 2019). The Crocker–Grier algorithm (Crocker and Grier, 1996) was used to link the obtained coordinates belonging to an individual mosquito in time using trackPy (Allan et al., 2016). The obtained tracking data is used to select all images that make up a single behavioral trajectory (e.g. landing, exploration, feeding, and take off) and store cropped image sequences centered on the focal mosquito. In addition to the computationally extracted data described below, such image sequences can also be used for the manual annotation of other events (e.g. stylet insertion as done in Jove et al., 2020).

We verified the tracking results of 111 individual trajectories across 12293 images resulting in an error rate of 0.045 (5/111). The validation dataset includes data from both Aedes and Anopheles experiments and consists of images having a variety of densities ranging from 0.05 to 0.4 mosquitoes per cm2. The most common error (4/5) is caused by wrongly assigning the identity of two mosquitoes that cross (e.g. an individual moving over another one and thus overlapping in the image). Interestingly, the validation videos (e.g. Video 7) make it straightforward to correct such errors by manually re-assigning the correct identity to the track. A rather minor amount of manual interventions therefore results in nearly perfect tracking.

#### Classifying locomotion behaviors

Request a detailed protocol

Locomotion behaviors (as presented in Figure 3A and B) can be automatically assigned based on the velocity of the centroid of a mosquito. To estimate the accuracy of this procedure, we manually labeled the behavior Ae. albopictus mosquitoes exhibited in 1124 frames of the dataset presented in Figure 3 and compared the labeled behaviors to the computationally detected behaviors. The overall accuracy of behavioral classification was 89%, with a per class accuracy of 90% (stationary), 89% (walking), and 97% (flight), with accuracy defined as: $TP+TNO$, with TP denoting true positives, TN true negatives, and O the number of observations. The classification of locomotion behaviors depends on the velocity thresholds set to distinguish flight, walking, and stationary behaviors. Figure 3—figure supplement 1 shows that classification accuracy peaks at 89% accurate classifications using a stationary – walking threshold of 2 mm/s and a walking – flight threshold of 12 mm/s, and exceeds 80% for a range of parameters.

#### Detecting engorgement

Request a detailed protocol

Images cropped on the focal mosquito (above) are used to determine a mosquito’s body shape at each timepoint to infer engorgement status by computationally removing all appendages and fitting an active contour model (using OpenCV [Bradski, 2000]) to the remaining body shape. For a mosquito to be computationally defined as engorged, two empirically determined conditions need to be met:

1. The abdominal area needs to expand 1.3 fold. Fold expansion is calculated as the ratio of the 90th percentile of abdominal area along the full trajectory and the 10th percentile of abdominal area in the first 10 s of the trajectory.

2. The 90th percentile of abdominal area measurements needs to exceed 2.4 mm2 for An. stephensi and An. coluzzii, or 3.0 mm2 for Ae. aegypti and Ae. albopictus.

We estimated the performance of the engorgement detection algorithm by validating all data presented in Figure 2 and Figure 2—figure supplement 1 and observed an overall sensitivity of engorgement detection $T?PP=0.81$ ($n=130$), with a sensitivity of 0.97 ($n=29$) and 0.76 ($n=101$) for Aedes and Anopheles mosquitoes, respectively. The overall specificity was $T?NN=1.0$ ($n=1184$). The difference in sensitivity for detecting engorgement in Aedes versus Anopheles may have two reasons: (1) Anopheles excrete excess liquid during feeding to a much larger extent than Aedes mosquitoes, resulting in a less pronounced dilation of the abdomen and?(2) some Anopheles experiments had a higher density of mosquitoes on the bite substrate leading to mosquitoes touching more often resulting in less accurate fitting of the body shape.

#### Body part tracking

Request a detailed protocol

The DeepLabCut framework (Mathis et al., 2018) was used to train a convolutional neural network (ResNet architecture) to detect the most distal part of the six legs, the abdominal tip, the center of the abdomen, the head, the tip of the proboscis, and for anophelines the tip of the maxillary palps. Due to their similar appearance, Ae. aegypti and Ae. albopictus can be analyzed using the same network, while a second network was trained for An. stephensi and An. coluzzii. Approximately 350 images were used to train the Aedes dataset, while approximately 300 images were used for the Anopheles dataset. To ensure robustness of training, the Aedes and Anopheles models were trained on 4 and 2 shuffles of the training set,?respectively. Averaged across shuffles training yielded an accuracy, defined as the mean average Euclidean error between the manual labels and predicted labels, of 11 pixels (275 μm) and six pixels (150 μm) in a 4.3 × 4.3 cm field of view, for Aedes and Anopheles, respectively. In addition to the mean performance across all body parts, prediction accuracies per groups of body parts (core: head, proboscis, abdomen, abdominal tip, (and palps for Anopheles); and legs: tips of all six legs) was 1.7 pixels (43 μm) for core body parts, and 1.6 (40 μm) for the tips of legs for the best performing Aedes model; and 5.2 pixels (130 μm) and 3.7 pixels (93 μm) for core and legs for the best performing Anopheles model. Trained models are available on GitHub.

We used cropped image sequences (described above) for inference. To facilitate downstream analysis of body part tracking data, body part coordinates can be aligned along the body axis (defined along the abdominal tip and center of the abdomen) yielding coordinates invariant of body orientation or movement. The wavelet transforms shown in Figure 4 are obtained by applying the Morlet continues wavelet transform to this data. Two-dimensional embedding of the aligned body part coordinates and their wavelet transform (Figure 4—figure supplement 1) was done by scaling the data (subtracting the mean and scaling to unit variance) and using t-distributed stochastic neighborhood embedding (tSNE) in two dimensions (Maaten and Hinton, 2008).

### Experiment-specific procedures

#### Feeding experiments

Request a detailed protocol

Population experiments (Figure 1 and Figure 2) were performed with 15–30 individuals in a 10 × 10?×?10 cm cage. Groups of mosquitoes were recorded for up to 1 hr and replaced by a new group for a subsequent recording (mosquitoes were not re-used and discarded after experiments). We noticed that activity is typically highest in the first 15–30 min of an experiment, depending on the question being addressed multiple short experiments may therefore yield more data compared to a single long experiment. Individual Ae. albopictus females (Figure 3) were recorded for 10 min per mosquito and discarded after the experiment. Movement status (Figure 3A and B) was classified using the velocity derived from tracking.

#### DEET experiments

Request a detailed protocol

As DEET dissolves Parafilm and plastics, a glass surface was placed on top of the heated culture flask (no artificial meal was present during DEET experiments). The glass surface was partly coated with 50% N,N-diethyl-meta-toluamide (DEET) using a cotton swab. Groups of 20 An. coluzzii females (14 days old) were released into a 10 × 10?×?10 cm cage with the DEET-coated substrate mounted in the floor. Images were acquired at 40 frames per second for 1 hr. Mosquito and body part tracking was performed as described above. The landing rate was calculated by summing the number of trajectories that started on the surface in question (DEET coated versus non-coated) and normalizing this value by the area of the surface. The dwell time was calculated as the average duration of all trajectories on the surface in question. The duration of trajectories moving from the non-coated surface to the DEET-coated surface was split proportionally to the time spend on the respective surface, trajectories moving from the DEET-coated surface to the non-coated surface were not observed indicating that the dwell time on the DEET surface was not limited by the size of the surface.

## Appendix 1

Appendix 1—table 1

## References

1. 1
2. 2
3. 3
4. 4
5. 5
6. 6
7. 7
8. 8
The opencv library, version 5.0 (2000)
Dr Dobb’s J. Software Tools.
9. 9
10. 10
11. 11
12. 12
13. 13
14. 14
The Physiology of Mosquitoes: International Series of Monographs on Pure and Applied Biology: Zoology
(2013)
Elsevier.
15. 15
16. 16
17. 17
Methods of digital video microscopy for colloidal studies (1996)
Journal of Colloid and Interface Science 179:298–310.
https://doi.org/10.1006/jcis.1996.0217
18. 18
19. 19
20. 20
21. 21
22. 22
23. 23
24. 24
25. 25
26. 26
27. 27
28. 28
29. 29
30. 30
31. 31
32. 32
33. 33
34. 34
35. 35
36. 36
37. 37
38. 38
39. 39
40. 40
41. 41
Sensory discrimination of blood and floral nectar by aedes aegypti mosquitoes
(2020)
Neuron In press.
42. 42
43. 43
44. 44
45. 45
Structure and function of the fascicular stylets, and the la bral and cibarial sense organs of male and female Aedes aeg ypti (L.) (Diptera, Culicidae)
(1974)
Quaestiones Entomologicae 10:187–216.
46. 46
47. 47
Collective selection of food patches in Drosophila (2016)
The Journal of Experimental Biology 219:668–675.
https://doi.org/10.1242/jeb.127431
48. 48
Visualizing data using t-sne
(2008)
Journal of Machine Learning Research 9:2579–2605.
49. 49
50. 50
51. 51
52. 52
53. 53
54. 54
55. 55
56. 56
57. 57
58. 58
59. 59
60. 60
61. 61
62. 62
63. 63
64. 64
Increased intradermal probing time in sporozoite-infected mosquitoes (1984)
The American Journal of Tropical Medicine and Hygiene 33:17–20.
https://doi.org/10.4269/ajtmh.1984.33.17
65. 65
66. 66
67. 67
68. 68
69. 69
70. 70
71. 71
72. 72
73. 73
74. 74
75. 75
76. 76
77. 77
78. 78
79. 79
80. 80

## Decision letter

1. Elena A Levashina
Reviewing Editor; Max Planck Institute for Infection Biology, Germany
2. Dominique Soldati-Favre
Senior Editor; University of Geneva, Switzerland
3. Matthew DeGennaro
Reviewer; Florida International University
4. Carlos Ribeiro
Reviewer; Champalimaud Centre for the Unknown, Portugal
5. Philip McCall
Reviewer

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

This manuscript is of great importance to the field as it represents an important technical advance that will support further studies of mosquito biting and disease-transmitting behavior. By combining image-based tracking, computer vision algorithms, and deep learning, the authors quantify the parameters which are of high relevance to future studies of the neurobiology controlling mosquito blood-feeding and, hence, transmission of human pathogens.

Decision letter after peer review:

Thank you for submitting your article "BiteOscope, an open platform to study mosquito blood-feeding behavior" for consideration by eLife. Your article has been reviewed by Dominique Soldati-Favre as the Senior Editor, a Reviewing Editor, and three reviewers. The following individuals involved in review of your submission have agreed to reveal their identity: Carlos Ribeiro (Reviewer #2); Philip McCall (Reviewer #3).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

We would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (http://www.asadoresteatre.com/articles/57162). Specifically, when editors judge that a submitted work as a whole belongs in eLife but that some conclusions require a modest amount of additional new data, as they do with your paper, we are asking that the manuscript be revised to either limit claims to those supported by data in hand, or to explicitly state that the relevant conclusions require additional supporting data.

Our expectation is that the authors will eventually carry out the additional experiments and report on how they affect the relevant conclusions either in a preprint on bioRxiv or medRxiv, or if appropriate, as a Research Advance in eLife, either of which would be linked to the original paper.

Summary:

The manuscript describes an experimental framework for studying female mosquito foraging and feeding behavior. By combining previously established stimuli promoting female feeding behavior, with image-based tracking, computer vision algorithms, and deep learning, the setup is able to record trajectories from multiple mosquitos within an acrylic box containing an area that contains an artificial meal which the mosquitos can reach by piercing a parafilm membrane. The authors provide data allowing them to quantify different parameters which are of high relevance to future studies of the neurobiology controlling mosquito blood-feeding and hence the potential transfer of pathogens. As proof of principle, they explore the impact of DEET on An. coluzzii behavior, which leads them to suggest that chemosensory neurons in the legs but not the proboscis are required for the deterring effect of DEET.

This manuscript is of great importance to the field as it represents an important technical advance that will support further studies of mosquito blood-feeding, a disease-transmitting behavior. The method is super useful, but the technical description and the discussion of the data is a bit superficial. Also, the method is limited to clear liquids, this technical limitation that is difficult to overcome and as such should be acknowledged and discussed in the manuscript. While the different aspects of the method are not novel on their own, the experimental framework is intriguing both in terms of the technical aspects as well as the potential to other researchers in the field. Importantly, the authors make the hardware design as well as the software openly available. The manuscript is well written and concise. However, the reviewers raised some major concerns that will need to be addressed.

The Title:

The mention of blood-feeding in the title is misleading. It should reflect the experimental setup. Therefore, blood-feeding should be replaced by feeding. Temperature is not a good proxy for blood.

Essential revisions:

1) The authors have completely avoided a large literature on the difference between the effects of DEET alone versus DEET and human odor. The authors need to review the literature on this topic more thoroughly and address their interpretation of their data. There is substantial data suggesting that DEET cannot repel mosquitoes in the vapor phase without human odor or other attractive odors. For a review on the topic, the authors should read "The mysterious multi-modal repellency of DEET" (https://doi.org/10.1080/19336934.2015.1079360). It is also untrue that there is evidence presented that Aedes aegypti mosquitoes can sense DEET without human odor in the vapor phase in DeGennaro et al., 2013. The section on DEET needs to be revised to address these issues and those below to fairly describe the authors results in context. DeGennaro et al., 2013 also should be referenced when discussing the separation between contact and olfactory actions of DEET in the mosquito as that was one of the key findings of the publication.

2) It is not clear whether Anopheles mosquitoes are any different that Aedes mosquitoes in regard to the effect of DEET in the vapor phase. There is not enough evidence presented in the paper to come to this conclusion for the reasons listed above.

3) There is some literature that states that ATP is not a phagostimulant in Anopheles species (https://doi.org/10.1111/j.1365-3032.1985.tb00029.x). ATP works well in Aedes species to stimulate blood-feeding behavior. In this manuscript, the authors conclude that ATP has no effect on Anopheline feeding when compared to Aedes aegypti. Key components of the feeding solution are important to induce engorgement, but not the ATP. The authors should provide their arguments about the choice of the feeding solution used in the study place their findings in the context of earlier literature.

4) Another major concern is the lack of description and validation of the behavioral classification methods used in the manuscript. In its current form the authors do not explain how they segment the behavior of the animals into approach/take off, stationary, walking, exploration, engorged etc. The quality of the analysis will largely depend on how well these classifications capture the actual behavior. Likewise, the authors never benchmark their algorithms. It is critical that the authors quantify how often their algorithm misses or wrongly assigns a specific behavior. Given that the quantification of the engorgement volume is a key parameter it would be especially important to focus on that aspect of behavior (e.g. how is, for example, full engorgement defined?). Ideally, the authors would validate the video-based quantification of the ingested volume by measuring the actual ingested volume experimentally. But given the difficulty in performing experiments at the moment, a validation of the video data using manual annotations and acknowledging the limitation in terms of quantifying actual volume should suffice.

5) The authors should also validate and benchmark the performance of the deep learning-based detection of the appendages.

6) The authors mostly analyze movies from experiments with multiple animals. It is widely acknowledged that reliably tracking the identity of multiple animals is challenging. The authors should benchmark their algorithm and provide an error rate for assigning the correct identity to animals. This is key for the correct interpretation of the results.

7) While the use of a membrane to visualize the actual feeding behavior of mosquitoes is a key aspect of the setup, the authors did not fully exploit it. It would be important to go beyond the anecdotal data in the first figure and show analyses of the piercing and stylet behavior highlighting this key aspect of the setup.

8) Some of the statements in the manuscript are rather anecdotal and would be better supported by including their quantification in figures. Furthermore, statistical analysis needs to be described in more details for Figure 3, i.e. include exact p-values in the figure. It also seems that the number of samples (n=9-10) is relatively low for making solid interpretations. Finally, some of the numbers described in the main text do not match the caption label for Figure 2.

9) The quantitative analysis shown in Figure 5 is insufficient, especially because it does not fully support the statements made in the main text. How is the landing rate (and dwell time etc.) calculated? Are these values normalized to the area coated by DEET and inhomogeneities for mosquito landing observed on the arena? Furthermore, the authors should control or at least discuss the possibility that aversiveness is being caused by physical attributes of the coated surface (i.e., slippery surface).

10) The authors' efforts to make the setup openly available including parts descriptions and code repository are highly appreciated. However, reproducibility and openness could be further improved by making the software easier accessible and understandable by structuring the code in the repository and documenting it, because currently, it does not explain which files to use to reproduce the findings. I also could not find the source data of Figure 2 and Figure 3 as described in the data availability statement. Data from all figures should be made available, clearly labeled, code should be provided for reproducing all figures, and well documented for others to use.

11) The Discussion section is rather superficial. A more thorough comparison of how the observed behavior compares to feeding and foraging behavior of other animals, especially insects would be a valuable addition. Also, discussing the limitations of the method would be advisable. The authors should openly recognize and discuss how prudent is an extrapolation of questions around vectorial capacity and host-vector interactions from a minimalist system with synthetic skin, blood, and without human-specific attractants to 'real world'. If the authors believe that it would not be difficult to augment the experimental setup with a human odor (synthetic or real) or any other attractant, then the text should state this clearly.

Revisions expected in follow-up work:

While the current experimental design of the BiteOscope provides advantages to tracking mosquito feeding behavior on humans or animals, a key question which remains unanswered is to which extent the behavior observed on the membrane is comparable to the behavior on a living host. Except for the actual blood feeing behavior, tracking animals foraging on a host should be feasible. It would be an extremely important addition to compare the behavior of mosquitoes in such a naturalistic setting with the behavior on the membrane. Understandably, in the current COVID situation performing experiments is challenging. Therefore, the authors should at least discuss this caveat and consider performing such experiments in follow-up work.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "BiteOscope, an open platform to study mosquito blood-feeding behavior" for further consideration by eLife. Your revised article has been evaluated by Dominique Soldati-Favre (Senior Editor) and a Reviewing Editor.

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

1) To avoid confusion and false expectations, the title should not include "blood-feeding" but "biting" behavior.

2) The authors should tone down the enthusiasm about the quality of the stylet imaging data in subsection “Automatic characterization of the blood feeding behavior of multiple species” and also mention that using DeepLabcut to track the stylet is not trivial.

3) Please modify the text to clarify the questions of the reviewer 3 regarding responses of the two mosquito species to DEET.

Reviewer #1:

This manuscript presents an exciting new approach to visualizing and characterizing mosquito blood-feeding behavior. This version of the manuscript is substantially revised. It addresses my prior concerns. In particular, I would point to the improved discussion of DEET and how the results presented in this paper fit into our understanding of DEET-mediated repellency. This paper will be of interest to eLife's broad readership and is ready for publication in its current form.

Reviewer #2:

The authors have done a superb job at revising the manuscript and addressing the concerns of the reviewers. Especially given the difficult times.we are all facing. I especially appreciate the thorough validation of the algorithms and the improved description of the methods and the curation of the code on GitHub.

Reviewer #3:

The manuscript is much improved but I'd like some feedback on the DEET story before going any further.

This is a system with many different elements each of which has resolution limits, and the bulk of the reviewers' comments were directed towards getting them recognised and acknowledged. The authors have addressed everything and, in most cases, they seem to have edited and have altered the manuscript sufficiently.

Nonetheless it is ultimately an imaging system and even the best pictures never tell the complete story. For me, a few issues remain.

Blood feeding – given the artificial membrane, the absence of blood/ necessity for clear liquid and presumably subsequent digestion (e.g. peritrophic mem from. Line brane formation?), this is 'biting' behaviour rather than bloodfeeding? This is likely to be relevant to many of the applications listed in the Discussion section.

Similarly, is engorgement an accurate term for what's being measured? Engorgement = fed to repletion, but here that is not always the case and mosquitoes are simply 'fed'.

Also, I wondered whether viewing from directly beneath the ventral abdomen is the most reliable position to measure an abdomen expanding with ingested volume of fluid – i.e. does the abdomen of all individuals expand similarly in every time (e.g. parous vs. nullipars?); what about 3D?

DEET – I found the authors' reply confusing (which read as if Afify and Potter provided more convincing evidence than the authors had.) but the text in the revised manuscript text was much clearer. Nonetheless, I still have reservations: the contact vs. non-contact observations are fine but is this conclusion justified? Can imaging [alone] provide the evidence to solve this question?

1) If the two genera differ in responses to DEET vapour, then in the real world' Anopheles coluzzii would land frequently on DEET-treated skin, whereas Aedes aegypti would rarely/never land. I have no data but having used DEET as a repellent for over 30 years in Africa and elsewhere, I remember Anophelines being repelled completely.

2) In the insecticide world, we use the terms 'contact-irritancy' and 'repellent-induced response', the latter being a change occurring prior to, or without contact. Both are usually bundled together for convenience, often viewed as being a question of exposure dosage from low/vapour to high/contact. I've always had doubts, increasingly so with the recent papers by Ingham et al.

Is it possible that the different responses reported for the 2 genera are the result of different response thresholds, with Aedes being more sensitive at lower levels (vapour) than Anopheles?.… also, have the olfactory neurons in Anopheles coluzzii been explored (which is not mentioned)?

3) Can results from experiments with DEET in the absence of host stimuli be reliable or indicative of anything other than the mosquito can/cannot detect it?

https://doi.org/10.7554/eLife.56829.sa1

## Author response

Summary:

The manuscript describes an experimental framework for studying female mosquito foraging and feeding behavior. […] The manuscript is well written and concise. However, the reviewers raised some major concerns that will need to be addressed.

We thank the reviewers and editors for the positive response and are please to read that our work is deemed of great importance to the field. Following the suggestions in the review report we have thoroughly revised our manuscript and now include a more detailed description of our method, and extensive discussion, and we have performed several additional analyses to validate and benchmark our algorithms. We thank the reviewers and editors for the constructive comments.

The Title:

The mention of blood-feeding in the title is misleading. It should reflect the experimental setup. Therefore, blood-feeding should be replaced by feeding. Temperature is not a good proxy for blood.

We agree with the reviewers that temperature is not a good proxy for blood. However, we note that full engorgement (the significant swelling of the abdomen when imbibing blood) only occurs when female mosquitoes feed on blood, while on nectar mosquitoes imbibe a much smaller volume (the volume of a typical nectar meal versus a blood meal in Ae. aegypti differs by a factor of 3 to 4 (Jove et al., 2020)). Furthermore, the piercing of a membrane and subsequent insertion of the stylets is a hallmark of blood feeding (Jove et al., 2020). As our method induces these hallmarks of blood feeding (and not nectar feeding) using a liquid that contains the essential components to elicit engorgement (osmotic pressure similar to blood, sodium ions, ATP) and a membrane that the stylets need to pierce to obtain the meal, we feel it is more accurate to refer to it as a method to study blood feeding as opposed to using the more general term ‘feeding’ (which would include e.g. nectar feeding).

Essential revisions:

1) The authors have completely avoided a large literature on the difference between the effects of DEET alone versus DEET and human odor. The authors need to review the literature on this topic more thoroughly and address their interpretation of their data. There is substantial data suggesting that DEET cannot repel mosquitoes in the vapor phase without human odor or other attractive odors. For a review on the topic, the authors should read "The mysterious multi-modal repellency of DEET" (https://doi.org/10.1080/19336934.2015.1079360). It is also untrue that there is evidence presented that Aedes aegypti mosquitoes can sense DEET without human odor in the vapor phase in DeGennaro et al., 2013. The section on DEET needs to be revised to address these issues and those below to fairly describe the authors results in context. DeGennaro et al., 2013 also should be referenced when discussing the separation between contact and olfactory actions of DEET in the mosquito as that was one of the key findings of the publication.

We apologize for not discussing the literature on the repellency of DEET clearly. We have revised this part of the Results section thoroughly and have added a section on DEET in the Discussion section to accurately place our observations in the context of the current literature. We now cite DeGennaro et al., (2013) when discussing contact-dependent repulsion of Ae. aegypti and have removed the reference to DeGennaro et al., (2013) when discussing the sensing of DEET by Ae. aegypti in the absence of human odor, as indeed DeGennaro et al., (2013) do not report on this.

We now furthermore discuss (Discussion section) a recent study comparing olfactory behaviors in response to DEET across several species by Afify and Potter (Afify and Potter, 2020) reporting that Ae. aegypti are repelled by DEET in the vapor phase presented in the absence of human odors. In contrast to previous studies that were difficult to interpret because they could not discriminate between olfactory-mediated repulsion and contact-dependent repulsion (e.g. Syed and Leal, (2008)), Afify and Potter, (2020) prevented mosquitoes from coming into contact with DEET to measure a purely olfactory response and observed that volatile DEET (100% concentration, presented alone) repelled Ae. aegypti (An. coluzzii did not respond to DEET in this assay, while Cx. Quinquefasciatus showed stronger repulsion compared to Ae. aegypti). We realize that this finding (reported by Afify and Potter, (2020)) contrasts with the statement brought up in the review report: “There is substantial data suggesting that DEET cannot repel mosquitoes in the vapor phase without human odor or other attractive odors”. However, we believe that this discrepancy may reflect the fact that many studies reporting on the effects of DEET do not assay repulsion by volatile DEET per se, yet test the absence of attraction in the presence of volatile DEET when mosquitoes face a choice between attractive odor alone versus attractive odor presented with DEET. In addition to Afify and Potter, (2020), we now also cite Stanczyk et al., (2013) which measured attraction of Ae. aegypti to a heat source covered with either DEET or solvent impregnated cloth. No human odor was presented during this experiment and the heat source with impregnated cloth was placed outside the mosquito cage (0.5 cm from the mesh) to prevent contact. Stanczyk et al., (2013) observed a strongly reduced attraction of Ae. aegypti to the DEET treated heat source, compared to the control heat source treated with solvent (í 30~ attracted to the control, 0% to the DEET treated heat source).

We also note that, in contrast to these two studies reporting repulsion of Ae. aegypti by volatile DEET without the presence of attractive odors (Stanczyk et al., 2013; Afify and Potter, 2020), an earlier study (Dogan et al., 1999) reported attraction of Ae. aegypti by volatile DEET alone. We now mention this diversity of results in the Discussion section and additionally point to (DeGennaro, 2015) for a review of the multi-modal repellency of DEET.

In summary, to place our observations regarding DEET repellency in context we now discuss:

1) Contact-based repellency by DEET observed in:

- An. coluzzii (our results)

- Ae. aegypti (DeGennaro et al., 2013; Dennis et al., 2019).

2) Olfactory responses to volatile DEET; we note that not all of these may be exhibited by a mosquito species, and (a subset of) these could potentially act in concert:

- Avoidance of DEET, also in the absence of human odors (the ‘smell and avoid’ hypothesis) e.g. observed in (Stanczyk et al., 2013; Afify and Potter, 2020)

- Inhibition of attraction to human odorants either through modulation of the olfactory system by ‘scrambling of the odor code’ (e.g. the ‘confusant’ hypothesis) and/or ‘smell and avoid’, for example observed in Ae. aegypti (DeGennaro et al., 2013)

- ‘Masking’ of odorants by DEET through direct chemical interactions between DEET and odorants, e.g. described in An. coluzzii (Afify et al., 2019)

2) It is not clear whether Anopheles mosquitoes are any different that Aedes mosquitoes in regard to the effect of DEET in the vapor phase. There is not enough evidence presented in the paper to come to this conclusion for the reasons listed above.

The DEET experiments described in our manuscript sought to assess if An. coluzzii was repelled upon direct contact with DEET (and we conclude this is indeed the case). For effects in the vapor phase we refer to the literature and in response to point #1 have expanded this discussion. We cite a study by Afify and Potter, (2020) that observed a difference between Ae. aegypti and An. Coluzzii responding to DEET in the vapor phase: An. coluzzii was not repelled by 100% DEET whereas Ae. aegypti showed moderate repulsion. See Afify and Potter, (2020).

Several studies report the activation of olfactory neurons in the Ae. aegypti antennae by volatile DEET presented alone without human body odors (Davis and Rebert, 1972; Boeckh et al., 1996; Stanczyk et al., 2010). Activity-dependent Ca2+ imaging in olfactory neurons of An. coluzzii on the other hand, suggests that volatile DEET presented alone does not activate olfactory neurons in the An. coluzzii antennae (Afify et al., 2019). This indicates that DEET in the vapor phase has a different effect on activation of the olfactory neurons of Aedes and Anopheles mosquitoes, and the repulsion assay by Afify and Potter, (2020) suggests this may have distinct behavioral effects.

Our experiments focussed on the contact-dependent effects of DEET on An. coluzzii, and it is interesting to note that our observations in An. coluzzii are very similar to those of DeGennaro et al., (2013) and Dennis et al., (2019) reporting on contact-dependent repulsion in Ae. aegypti. In contrast to the similarities in contact-dependent repulsion by DEET, the literature discussed above (and in response to point #1) suggests that the olfactory responses to volatile DEET may to differ to some degree between the two species. To place our results in context, we now discuss these observations extensively in the Results section and the Discussion section.

3) There is some literature that states that ATP is not a phagostimulant in Anopheles species (https://doi.org/10.1111/j.1365-3032.1985.tb00029.x). ATP works well in Aedes species to stimulate blood-feeding behavior. In this manuscript, the authors conclude that ATP has no effect on Anopheline feeding when compared to Aedes aegypti. Key components of the feeding solution are important to induce engorgement, but not the ATP. The authors should provide their arguments about the choice of the feeding solution used in the study place their findings in the context of earlier literature.

We are aware of the study by Galun et al., (1985) demonstrating that ATP is not required to induce engorgement in several Anopheles species. However, as we did not perform feeding experiments with Anopheles mosquitoes without the presence of ATP, we did not discuss this study in the manuscript. We agree with the reviewers that this could be of interest to the readers and now cite Galun et al., (1985) on line 101 mentioning that ATP may not be required to induce engorgement in Anopheles.

All feeding experiments presented in the manuscript were performed using phosphate buffered saline (PBS) with 1 mM ATP as the artificial meal (with the exception of ‘PBS only’ experiments presented in Figure 3). We based our choice for the artificial meal on Galun et al., (1963) and Duvall et al., (2019) which demonstrate that eliciting the feeding response in Ae. aegypti requires sodium ions and an osmotic pressure close to blood. Galun et al., (1963) observed the highest percentage of feeding mosquitoes using 10 mM ATP in 150 mM NaCl (the same concentration as typical in PBS) and Duvall et al., (2019) observed robust feeding using 1 mM ATP in 120 mM NaHCO3. We chose 1 mM ATP in PBS as the artificial meal as it is a commonly used buffer that fits the criteria described by Galun et al., (1963) and Duvall et al., (2019). As Galun et al., (1985) report engorgement of Anopheles mosquitoes on 150 mM NaCl (neutralized with NaHCO3) that is independent of ATP, we chose for consistency across our experiments and used the same formulation for experiments with Aedes and Anopheles mosquitoes.

A separate study led by the Vosshall lab in which we collaborated characterizing blood detection in the Ae. aegypti stylet (currently under review and available on bioRxiv (Jove et al., 2020)) used the biteOscope and showed that Ae. aegypti also engorge on a simplified saline meal consisting of 1 mM ATP in 110 mM NaCl and 20 mM NaHCO3 (compared to PBS: 150 mM NaCl, 1 mM KH2PO4, 3 mM Na2HPO4). We have verified that Ae. albopictus, An. stephensi, and An. coluzzii also feed robustly on this simpli1ed formulation and now mention the possibility to use 1 mM ATP in 110 mM NaCl and 20 mM NaHCO3 in the Materials and methods section.

4) Another major concern is the lack of description and validation of the behavioral classification methods used in the manuscript. In its current form the authors do not explain how they segment the behavior of the animals into approach/take off, stationary, walking, exploration, engorged etc. The quality of the analysis will largely depend on how well these classifications capture the actual behavior. Likewise, the authors never benchmark their algorithms. It is critical that the authors quantify how often their algorithm misses or wrongly assigns a specific behavior. Given that the quantification of the engorgement volume is a key parameter it would be especially important to focus on that aspect of behavior (e.g. how is, for example, full engorgement defined?). Ideally, the authors would validate the video-based quantification of the ingested volume by measuring the actual ingested volume experimentally. But given the difficulty in performing experiments at the moment, a validation of the video data using manual annotations and acknowledging the limitation in terms of quantifying actual volume should suffice.

We agree with the reviewers that adding additional details on the accuracy of the behavioral classification methods is a very useful addition to the manuscript. We performed a thorough benchmarking of our locomotion behavior classification and engorgement detection algorithms and have now included the results of this analysis in the text as indicated below and as Figure 3—figure supplement 1.

Locomotion behavior classification

The classification of locomotion behaviors is based on the velocity derived from centroid tracking. In the original manuscript we mentioned velocity based locomotion classification very briefly in the Discussion section. Following the suggestions of the reviewers, we now include information regarding the performance of our classifier of locomotion behaviors as presented in Figure 3A and B of the manuscript. To quantify performance, we manually labeled the behavior of Ae. Albopictus mosquitoes exhibited in 1124 frames of the dataset presented in Figure 3 of the manuscript, and compared the labeled behaviors to the computationally detected behaviors. The overall accuracy of behavioral classification was 89%, with a per class accuracy of 90% (stationary), 89% (walking), and 97% (flight), with accuracy defined as: $TP+TNO$, with TP denoting true positives, TN true negatives, and O the number of observations. The classification of locomotion behaviors depends on the velocity thresholds set to distinguish flight, walking, and stationary behaviors, we therefore performed a sensitivity analysis to estimate the dependence of the classification accuracy on the thresholds used. Figure 2 shows the results of this analysis, demonstrating that classification accuracy peaks at 89% accurate classifications using a stationary – walking threshold of 2 mm/s and a walking – flight threshold of 12 mm/s, and is superior to 80% accurate for a range of parameters. We now include this information in the Results section and the Materials and methods section, and have included the figure detailing the threshold-dependence of classification accuracy as Figure 3—figure supplement 1.

Engorgement detection

We now included a thorough characterization of the performance of the engorgement detection algorithm, and have included additional details in the Materials and methods section regarding the computational detection of full engorgement. To determine engorgement, we fit an active contour model to the mosquito’s body (with appendages computationally removed) and we use the area of the fitted shape as a proxy for engorgement. As can be seen in Video 7 and panels 1 and 2 of Figure 1H of the manuscript, the abdominal width increases during imbibing and saturates when a mosquito is fully engorged. After full engorgement, mosquitoes retract their mouthparts from the artificial blood meal and often remain stationary for a period to excrete excess liquid. In anophelines this is clearly visible as a growing droplet at the abdominal tip, whereas in Aedes this is less pronounced yet occasionally droplet excretion can be seen. For a mosquito to be computationally defined as engorged, two conditions need to be met:

1) The abdominal area needs to expand by 1.3 fold, where fold expansion is calculated as the ratio of the 90th percentile of abdominal area measurements and the 10th percentile of abdominal area measurements in the first 10 seconds of the trajectory.

2) The ninetieth percentile of abdominal area measurements needs to exceed 2.4 mm2 for An. stephensi and An. coluzzii, or 3.0 mm2 for Ae. aegypti and Ae. albopictus.

Following the reviewers’ suggestion (as indeed our ability to do experiments is severely limited during the Covid-19 crisis), we estimated the performance of the engorgement detection algorithm by validating all data presented in Figure 2 and Figure 2—figure supplement 1. Visual inspection of all video data indicated that the overall sensitivity of engorgement detection $TPp=0.81(n=130)$, with a sensitivity of 0.97 (n = 29) and 0.76 (n = 101) for Aedes and Anopheles mosquitoes, respectively. The overall specificity was $TNN=1.0(n=101)$. We now include these numbers in the manuscript on lines 136-137 and details regarding the benchmarking in subsection “Detecting engorgement”.

We note that the sensitivity for detecting engorgement is lower for Anopheles compared to Aedes mosquitoes (0.76 versus 0.97). Two possible reasons may be the source of this discrepancy: (1) Anopheles excrete excess liquid during feeding to a much larger extent than Aedes mosquitoes, resulting in a less pronounced dilation of the abdomen making it harder to detect the dilation, (2) we noticed a higher density of mosquitoes on the bite substrate in experiments conducted with Anopheles compared to Aedes mosquitoes, the higher density more often leads to mosquitoes that physically touch resulting in challenging situations for the detection algorithm as the active contour model may fit the body shape less accurate when mosquitoes touch. Figure 3 shows the relation between the average number of mosquitoes present on the bite substrate versus the accuracy of engorgement detection. A lower number of mosquitoes on the substrate results in higher accuracy, yet at similar mosquito densities the algorithm performs better on Aedes than on Anopheles, likely due to a larger increase in abdominal area during feeding in Aedes.

5) The authors should also validate and benchmark the performance of the deep learning-based detection of the appendages.

In the original manuscript we described the accuracy of the deep learning-based tracking of appendages in subsection “Pose estimation, behavioral classification, and contact-dependent sensing”, and the Discussion section stating an average accuracy of the detection of body parts of 10 pixels (250 μm) and 8 pixels (200 μm) for Aedes and Anopheles, respectively (average distance between manually labeled and computationally predicted body part location). We now include additional benchmarking and details regarding the performance of deep learning-based body part detection in subsection “Body part tracking”.

In order to assess the robustness of the trained model, we created multiple shuffles of the training data set. ‘Shuffle’ here refers to the generation of a training set from a pool of manually labeled images: labeled images are randomly split into a ‘train’ and a ‘test’ set, the random split differs per shuffle allowing one to assess the ‘robustness’ of training.

Below we report the accuracy of body part detection defined as the mean average Euclidean error between the manual labels and the ones predicted by the algorithm. Averaged over 4 shuffles of the training set Aedes body parts were detected with a mean accuracy of 11 pixels, (275 μm). The mean accuracy of predicting Anopheles body parts averaged over 2 shuffles of the training set was 6 pixels (150 μm). In addition to the mean performance across all body parts, we now also include accuracies per groups of body parts for the best performing model (core: head, proboscis, abdomen, abdominal tip, (and palps for Anopheles); and legs: tips of all 6 legs). In the best performing model for detecting Aedes body parts the mean distance between manual labels and predicted positions was 1.7 pixels (43 μm) for core body parts, and 1.6 (40 μm) for the tips of legs. The best performing model for Anopheles body parts had an accuracy of 5.2 pixels (130 μm) and 3.7 pixels (93 μm) for core and legs, respectively. We now include these metrics in the manuscript in subsection “Body part tracking” (and briefly in the Results section). We furthermore provide the final trained models on Github https://github.com/felixhol/biteOscope (mentioned subsection “Body part tracking”). Depending on the exact imaging conditions of users, these trained models may work without modification on newly acquired data, in other cases the provided models can be used as a convenient starting point for training on new datasets. In the latter case, starting from the pre-trained models will reduce the required training time substantially making training on a standard computer without GPU feasible.

6) The authors mostly analyze movies from experiments with multiple animals. It is widely acknowledged that reliably tracking the identity of multiple animals is challenging. The authors should benchmark their algorithm and provide an error rate for assigning the correct identity to animals. This is key for the correct interpretation of the results.

We agree with the reviewers that correctly tracking the identity of multiple mosquitoes that are simultaneously present on the bite substrate is an important aspect of our image analysis pipeline. It indeed is a great suggestion to present the validation in the manuscript.

In order to assess the performance of our mosquito tracking algorithm, we created videos in which the results of the tracking algorithm are overlaid on the raw imaging data. The validation videos indicate the location where the centroid of a mosquito is detected and plot the numeric ID assigned to a mosquito. We verified the tracking results of 111 individual trajectories across 12293 images resulting in an error rate of 0.045 (5/111). The validation dataset includes data from both Aedes and Anopheles experiments and consists of images having a variety of densities ranging from 0.05 – 0.4 mosquitoes per cm2. The most common error (4/5) is caused by erroneously assigning identities when mosquitoes cross (e.g. an individual moving over another one and thus overlapping in the image). Interestingly, the validation videos make it straightforward to correct ID swap errors by manually re-assigning the correct identity to the track. A rather minor amount of manual interventions therefore results in nearly perfect tracking. We now include information regarding the performance of the tracking algorithm in subsection “Automatic characterization of the blood feeding behavior of multiple species” and Subsection “Computaional tools”. We include an example validation video, Video 1—figure 1, and provide code to create the validation videos in the Github repository.

7) While the use of a membrane to visualize the actual feeding behavior of mosquitoes is a key aspect of the setup, the authors did not fully exploit it. It would be important to go beyond the anecdotal data in the first figure and show analyses of the piercing and stylet behavior highlighting this key aspect of the setup.

We agree with the reviewers that visualizing the stylet of a probing/feeding mosquito presents very interesting opportunities. This imaging capability for instance was of value to one of the ‘early adopters’ of the biteOscope (the lab of Leslie Vosshall) in a study that characterizes the sensory perception of blood in Ae. aegypti mouthparts. In this study mosquitoes were presented membrane feeders containing a variety of artificial meals and the question was raised whether the stylet came into contact with all meals presented, or only with those that elicited engorgement (i.e. are non-appetitive meals rejected based on evaluation by the stylet, or by other means). We used the biteOscope to answer this question and demonstrated that the stylet indeed came in contact with all meals and evaluates the meal before engorging. This study in which we collaborated is currently under review and available on bioRxiv (Jove et al., 2020). We now discuss this use-case of stylet imaging in the Discussion section.

We believe that there are many more questions that can be addressed by directly visualizing the stylet of a feeding mosquito and anticipate that computationally detecting stylet insertion would be a very valuable tool. As the reviewers note’ it would be convenient to use DeepLabCut to detect the stylet. We have indeed tried this however, our efforts have not produced an algorithm that detects the stylet with high enough accuracy to be useful. The reason for this is (at least) two fold: (1) the stylet appears as a rather subtle feature in images, having low contrast (compared to other body parts) and adopting a variety of conformations complicating reliable detection; and (2) the stylet is only visible during probing and feeding and therefore only appears in a minor subset of images resulting in a low number of training images (and a high rate of false positive detections at the tip of the proboscis). We are currently exploring other deep learning methods (not based on DeepLabCut) for stylet detection and hope that this will result in a high-confidence stylet detection algorithm. Given the challenging nature of this problem, and the range of (computational) capabilities we currently present, we feel that high-accuracy computational detection of the stylet is beyond the scope of the current work.

In addition to a number of behavioral statistics, our current computational pipeline saves cropped videos of all individually tracked mosquitoes and we mention in subsection “Computational tools” that these videos may serve as a convenient starting point for manual annotation of stylet piercing as has been done in Jove et al., (2020).

8) Some of the statements in the manuscript are rather anecdotal and would be better supported by including their quantification in figures. Furthermore, statistical analysis needs to be described in more details for Figure 3, i.e. include exact p-values in the figure. It also seems that the number of samples (n=9-10) is relatively low for making solid interpretations. Finally, some of the numbers described in the main text do not match the caption label for Figure 2.

In response to the points raised above, we have performed several analyses (see points 4–6 above) and now include additional quantification in the manuscript. Following the reviewers’ suggestion, we now include the exact p-values in the caption of Figure 3. We furthermore noticed that some data points were accidentally excluded from panel 3 of Figure 3C (bout length), we have updated this graph to include all data points. The conclusions regarding this Figure are not affected by this. We agree that using a larger number of individual mosquitoes (9–10 for the data presented in Figure 3) would give our statistical analysis more power. However, the differences between the two experimental conditions are large enough to observe statistically significant differences between the treatments with the current number of samples. Furthermore, we discuss the data from Figure 3 (single mosquito experiments) in relation to the results from Figure 2E which have a much larger n (population experiments, n = 111 trajectories for Ae. albopictus, n = 1184 trajectories for all species together) and note that these different data sets are in agreement. It will be interesting to follow up on these findings at a larger scale once we can resume experiments post Covid-19. We thank the reviewers for pointing out a typo in the n quoted in the text discussing Figure 2. The caption of Figure 2 indeed specified n = 349, whereas the main text read n = 350 trajectories for the An. coluzzi experiment. The correct number is n = 349 and we have corrected this.

9) The quantitative analysis shown in Figure 5 is insufficient, especially because it does not fully support the statements made in the main text. How is the landing rate (and dwell time etc.) calculated? Are these values normalized to the area coated by DEET and inhomogeneities for mosquito landing observed on the arena? Furthermore, the authors should control or at least discuss the possibility that aversiveness is being caused by physical attributes of the coated surface (i.e., slippery surface).

The landing rate was calculated by summing the number of trajectories that started on the surface in question (DEET coated versus non-coated) and normalizing this value by the area of the surface (we now explicitly mention normalization by the surface area in subsection “DEET repels An. coluzzii upon contact with legs”). The dwell time was calculated as the average duration of all trajectories on the surface in question. The duration of trajectories moving from the non-coated surface to the DEET coated surface was split proportionally to the time spent on the respective surface. Trajectories moving from the DEET coated surface to the non-coated surface were not observed indicating that the dwell time on the DEET surface was not limited by the size of the surface. We now include this information in subsection “DEET experiments”. Following the reviewers’ suggestion we now also discuss the possibility that physical attributes of the DEET coating may influence the mosquitoes’ interaction with the surface in the Discussion section.

10) The authors' efforts to make the setup openly available including parts descriptions and code repository are highly appreciated. However, reproducibility and openness could be further improved by making the software easier accessible and understandable by structuring the code in the repository and documenting it, because currently, it does not explain which files to use to reproduce the findings. I also could not find the source data of Figure 2 and Figure 3 as described in the data availability statement. Data from all figures should be made available, clearly labeled, code should be provided for reproducing all figures, and well documented for others to use.

We agree with the reviewers that making our setup and code openly available is important, and the reviewers are right in pointing out that the ‘user friendliness’ of the Github repository at the time of initial submission could be improved. We have now restructured the repository to make it easier to navigate. In the repository, we provide a README file that specifies the function of each algorithm and we have commented the code extensively indicating the function of code blocks and variables that should be modified to provide experiment specific details (e.g. directories where data is stored). We furthermore provide test data in the repository to enable testing of the code by new users. In addition to these improvements with respect to accessibility of the code, we now also provide design files for laser cutting the mosquito cages. Data will be available on Github upon acceptance of the manuscript.

11) The Discussion section is rather superficial. A more thorough comparison of how the observed behavior compares to feeding and foraging behavior of other animals, especially insects would be a valuable addition. Also, discussing the limitations of the method would be advisable. The authors should openly recognize and discuss how prudent is an extrapolation of questions around vectorial capacity and host-vector interactions from a minimalist system with synthetic skin, blood, and without human-specific attractants to 'real world'. If the authors believe that it would not be difficult to augment the experimental setup with a human odor (synthetic or real) or any other attractant, then the text should state this clearly.

Following these suggestions, we have thoroughly revised the Discussion section. We now include a more extensive discussion of feeding and foraging behavior in other animals and methods to study those behaviors in the Discussion section. In addition, we point out several interesting opportunities for future research where the biteOscope could be a useful tool (Discussion section). We furthermore discuss several limitations of our setup and the interesting possibility to add odorants to the setup to more closely mimic a real host (Discussion section). We note that a synthetic bite substrate will likely never by exactly as attractive as a real host, yet we also that many factors that may change behavior (e.g. infections/nutritional status or components of the microbiome) are best assessed in a relative manner, e.g. comparing non-infected to infected individuals. We indicate that comparing cohorts of mosquitoes undergoing different experimental treatments puts less emphasis on the absolute attractiveness of the bite substrate and thus mitigates potential issues related to the fact that the a synthetic bite substrate is likely less attractive than a real live host. We discuss the potential use of non-clear liquids in the Discussion section.

Revisions expected in follow-up work:

While the current experimental design of the BiteOscope provides advantages to tracking mosquito feeding behavior on humans or animals, a key question which remains unanswered is to which extent the behavior observed on the membrane is comparable to the behavior on a living host. Except for the actual blood feeing behavior, tracking animals foraging on a host should be feasible. It would be an extremely important addition to compare the behavior of mosquitoes in such a naturalistic setting with the behavior on the membrane. Understandably, in the current COVID situation performing experiments is challenging. Therefore, the authors should at least discuss this caveat and consider performing such experiments in follow-up work.

As mentioned above, we now briefly discuss the comparison of behavior in our synthetic system to behavior on a real living host in the Discussion section. We furthermore agree with the reviewers that this would be an interesting follow up and will explore this in future work (when Covid-19 related limitations are lifted).

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

1) To avoid confusion and false expectations, the title should not include "blood-feeding" but "biting" behavior.

We have changed the title accordingly, it now reads “BiteOscope, an open platform to study mosquito biting behavior”.

2) The authors should tone down the enthusiasm about the quality of the stylet imaging data in subsection “Automatic characterization of the blood feeding behavior of multiple species” and also mention that using DeepLabcut to track the stylet is not trivial.

We have toned down subsection “Automatic characterization of the blood feeding behavior of multiple species” and now mention that tracking the stylet using DeepLabCut is not trivial in subsection “Pose estimation, behavioral classification, and contact-dependent sensing”.

3) Please modify the text to clarify the questions of the reviewer 3 regarding responses of the two mosquito species to DEET.

We respond to the questions of reviewer #3 below and have revised the text accordingly.

Reviewer #1:

This manuscript presents an exciting new approach to visualizing and characterizing mosquito blood-feeding behavior. This version of the manuscript is substantially revised. It addresses my prior concerns. In particular, I would point to the improved discussion of DEET and how the results presented in this paper fit into our understanding of DEET-mediated repellency. This paper will be of interest to eLife's broad readership and is ready for publication in its current form.

We thank the reviewer for their support and we are glad that the reviewer appreciates our improved discussion on DEET-mediated repellency.

Reviewer #2:

The authors have done a superb job at revising the manuscript and addressing the concerns of the reviewers. Especially given the difficult times we are all facing. I especially appreciate the thorough validation of the algorithms and the improved description of the methods and the curation of the code on GitHub.

We thank the reviewer for their support and we are glad that the reviewer appreciates the thorough validation of our algorithms and our efforts at making our code (available at GitHub) more user friendly.

Reviewer #3:

The manuscript is much improved but I'd like some feedback on the DEET story before going any further.

We are pleased to read that the reviewer evaluates our manuscript as much improved and we provide feedback regarding the DEET story below.

This is a system with many different elements each of which has resolution limits, and the bulk of the reviewers' comments were directed towards getting them recognised and acknowledged. The authors have addressed everything and, in most cases,, they seem to have edited and have altered the manuscript sufficiently.

Nonetheless it is ultimately an imaging system and even the best pictures never tell the complete story. For me, a few issues remain.

Blood feeding – given the artificial membrane, the absence of blood/ necessity for clear liquid and presumably subsequent digestion (e.g. peritrophic mem from. Line brane formation?), this is 'biting' behaviour rather than bloodfeeding? This is likely to be relevant to many of the applications listed in the Discussion section.

Following this comment (and the Editor’s suggestion) we have changed the title of the manuscript which now states ‘biting behavior’ instead of ‘blood feeding behavior’.

Similarly, is engorgement an accurate term for what's being measured? Engorgement = fed to repletion, but here that is not always the case and mosquitoes are simply 'fed'.

Also, I wondered whether viewing from directly beneath the ventral abdomen is the most reliable position to measure an abdomen expanding with ingested volume of fluid – i.e. does the abdomen of all individuals expand similarly in every time (e.g. parous vs. nullipars?); what about 3D?

We presented a validation of our engorgement detection algorithm in the first revision of our manuscript, and as suggested by the reviewers, we used manual annotations to validate all video data and calculate the sensitivity (0.97 for Aedes and 0.76 for Anopheles) and specificity (1.0 for both) of engorgement detection. This validation was mentioned in subsection “Automatic characterization of the blood feeding behavior of multiple species”, and subsection “Detecting engorgement”. While 3D imaging and/or adding another camera (e.g. a side view) could add another way of assessing engorgement status, we feel that this would significantly complicate the optical set up while the sensitivity and specificity of our detection algorithm are already quite high. We furthermore note that the artificial meal we use (phosphate buffered saline + 1 mM ATP) has been used to study blood feeding in several other studies. Notably, Duvall et al., (2019) and Jove et al., (2020) compare the weight of Ae. aegypti mosquitoes fed on sheep blood to the weight of mosquitoes fed on saline + ATP and both studies report no significant difference in the post-feeding weight of mosquitoes feeding on either meal (please see Figure 1C of (Duvall et al., 2019) and Supplemental Figure S1F and H of (Jove et al., 2020)). As weight is a good proxy for ingested volume, these observations indicate that there is no significant difference between the ingested volume when using sheep blood versus our artificial meal.

DEET – I found the authors' reply confusing (which read as if Afify and Potter provided more convincing evidence than the authors had.) but the text in the revised manuscript text was much clearer. Nonetheless, I still have reservations: the contact vs. non-contact observations are fine but is this conclusion justified? Can imaging [alone] provide the evidence to solve this question?

We apologize if our response was confusing, yet are glad to read that the manuscript text is clear. We would like to stress that we report on a different aspect of DEET repellency than Afify and Potter did: We tested if An. coluzzii is repelled upon contact with DEET (a hypothesis not tested before), whereas Afify and Potter studied olfactory responses to volatile DEET (i.e. responses in the absence of contact). As we study a different aspect of DEET-mediated repellency, our results are complimentary to, instead of more/less convincing than, those of Afify and Potter. In the Discussion section we mention studies by Afify and Potter, and others to place our observations regarding contact-dependent repellency in the broader context of the several modes in which DEET may repel mosquitoes. We extensively discuss our results concerning DEET in relation to the prior literature as this was an explicit request in the initial review report.

1) If the two genera differ in responses to DEET vapour, then in the real world' Anopheles coluzzii would land frequently on DEET-treated skin, whereas Aedes aegypti would rarely/never land. I have no data but having used DEET as a repellent for over 30 years in Africa and elsewhere, I remember Anophelines being repelled completely.

This is an interesting observation which may be explained by the various possible modes of action of DEET (these different modes of action are discussed in the Discussion section). Our results show that An. coluzzii is repelled by DEET upon contact, and DeGennaro et al., (2013) and Dennis et al., (2019) report that this is also the case in Ae. aegypti. Regarding olfactory responses to volatile DEET in the absence of contact, several studies suggest that the olfactory response to volatile DEET of these two species differs (Davis and Rebert, 1972; Boeckh et al., 1996; Stanczyk et al., 2010, 2013; Afify et al., 2019) which may result in differences in olfactory behavioral responses to volatile DEET: Ae. aegypti has been observed to avoid volatile DEET in the absence of attractive cues, whereas An. coluzzii does not seem to avoid volatile DEET in the absence of attractive cues (Afify and Potter, 2020). A third mode of action of DEET is “masking” in which DEET acts directly on the odorants emanating from a host: through chemical interactions DEET decreases the odorants volatility and thereby reduces the amount of attractive odorants capable of activating mosquito olfactory receptors and thus inhibiting behavioral responses (Syed and Leal, 2008). A recent study (Afify et al., 2019) observed that the olfactory neurons of An. coluzzii are not activated by DEET, and also showed that the neuronal response to otherwise attractive compounds (e.g. 1-octen-3-ol) was strongly inhibited when the volatile attractive compound was co-presented with volatile DEET (i.e. compound that normally elicit a strong response in olfactory neurons did not elicit a strong neuronal response when presented together with volatile DEET). This observation suggests that DEET interacts directly with odorants and through masking indeed may inhibit behavioral responses (such as flying towards a host). Together, our results and prior literature thus suggest that DEET has (at least) two effects on An. coluzzii: contact-based repellency, and the masking of odorants. It is thus likely that DEET has kept reviewer #3 safe from anopheline bites through the combined effect of odorant masking and contact-dependent repulsion.

2) In the insecticide world, we use the terms 'contact-irritancy' and 'repellent-induced response', the latter being a change occurring prior to, or without contact. Both are usually bundled together for convenience, often viewed as being a question of exposure dosage from low/vapour to high/contact. I've always had doubts, increasingly so with the recent papers by Ingham et al.

Is it possible that the different responses reported for the 2 genera are the result of different response thresholds, with Aedes being more sensitive at lower levels (vapour) than Anopheles?.… also, have the olfactory neurons in Anopheles coluzzii been explored (which is not mentioned)?

This is an interesting suggestion, yet as our DEET assay was designed to primarily test for contact dependent behaviors, our data do not allow us to draw conclusions regarding the concentration dependency of olfactory (and other non-contact) behaviors. As mentioned above, prior literature suggests that at the same concentration volatile DEET has distinct effects on the olfactory neurons of Ae. aegypti and An. coluzzii.

Following the reviewers suggestion we now mention that these effects may be concentration dependent in the Discussion section.

The olfactory neurons of An. coluzzii have indeed been studied. It may be a misunderstanding that this was not mentioned as we discussed studies reporting on the response of the olfactory neurons of An. coluzzii to DEET in subsection “DEET repels An. coluzzii upon contact with legs” and Discussion section. In summary, prior literature shows that the olfactory neurons of Ae. aegypti are activated by DEET (Davis and Rebert, 1972; Boeckh et al., 1996; Stanczyk et al., 2010) whereas the olfactory neurons of An. coluzzii are not activated by DEET (Afify et al., 2019; Afify and Potter, 2020).

3) Can results from experiments with DEET in the absence of host stimuli be reliable or indicative of anything other than the mosquito can/cannot detect it?

This is an interesting question. However, we note that in our assay we do provide heat which is an important factor in short-range host attraction. Our results are therefore best interpreted in the presence of a host stimulus (heat).

https://doi.org/10.7554/eLife.56829.sa2

## Article and author information

### Author details

1. #### Felix JH Hol

1. Department of Bioengineering, Stanford University, Stanford, United States
2. Insect-Virus Interactions Unit, Institut Pasteur, UMR2000, CNRS, Paris, France
3. Center for research and Interdisciplinarity, U1284 INSERM, Université de Paris, Paris, France
##### Contribution
Conceptualization, Resources, Data curation, Software, Formal analysis, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing - original draft, Project administration
##### For correspondence
felix.hol@pasteur.fr
##### Competing interests
No competing interests declared
2. #### Louis Lambrechts

Insect-Virus Interactions Unit, Institut Pasteur, UMR2000, CNRS, Paris, France
##### Contribution
Resources, Project administration, Writing - review and editing
##### Competing interests
No competing interests declared
3. #### Manu Prakash

Department of Bioengineering, Stanford University, Stanford, United States
##### Contribution
Conceptualization, Resources, Funding acquisition, Project administration, Writing - review and editing
##### Competing interests
No competing interests declared

### Funding

• Felix JH Hol

• Felix JH Hol

• Felix JH Hol

#### Agence Nationale de la Recherche (ANR-16-CE35-0004-01)

• Louis Lambrechts

#### Agence Nationale de la Recherche (ANR-18-CE35-0003-01)

• Louis Lambrechts

#### Agence Nationale de la Recherche (ANR-10-LABX-62-IBEID)

• Louis Lambrechts

• Manu Prakash

#### United States Agency for International Development (Grand Challenges: Zika and Future Threats)

• Felix JH Hol
• Manu Prakash

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

### Acknowledgements

We thank Emilie Giraud, Gregory Murray, Haripriya Vaidehi Narayanan, Shailabh Kumar, William Gilpin, Hongquan Li, Leslie Vosshall, Veronica Jove, and all members of the Prakash and Lambrechts labs and the Center for Research and Interdisciplinarity for valuable discussions; and Catherine Lallemand, Sylvain Golba, and Patricia Baldacci for help with the rearing of mosquitoes. We thank the reviewers for their constructive comments. An. stephensi strain Sda500 and An. coluzzii strain N’Gousso were provided by the Centre de Production et Infection d’Anopheles of Institut Pasteur; Ae. aegypti strain Liverpool was provided by Leslie Vosshall (Rockefeller University); Ae. aegypti strain D2S3 (NR-45838) was provided by Centers for Disease Control and Prevention for distribution by BEI Resources, NIAID, NIH. FJHH was supported by a Rubicon fellowship for the Netherlands Foundation for Scientific Research, a Career Award at the Scientific Interface from the Burroughs Wellcome Fund, and a Marie Curie Fellowship from the European Union. L.L. is supported by the French Agence Nationale de la Recherche (grants ANR-16-CE35-0004-01 and ANR-18-CE35-0003-01), and the French Government’s Investissement d’Avenir program Laboratoire d’Excellence Integrative Biology of Emerging Infectious Diseases (grant ANR-10-LABX-62-IBEID). MP was supported by NIH DP2-AI124336 New Innovator Award and USAID Grand Challenges: Zika and Future Threats Award.

### Senior Editor

1. Dominique Soldati-Favre, University of Geneva, Switzerland

### Reviewing Editor

1. Elena A Levashina, Max Planck Institute for Infection Biology, Germany

### Reviewers

1. Matthew DeGennaro, Florida International University
2. Carlos Ribeiro, Champalimaud Centre for the Unknown, Portugal
3. Philip McCall

### Publication history

2. Accepted: September 5, 2020
3. Accepted Manuscript published: September 22, 2020 (version 1)
4. Version of Record published: October 5, 2020 (version 2)

? 2020, Hol et al.

## Metrics

• 2,176
Page views
• 267
• 2
Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

A two-part list of links to download the article, or parts of the article, in various formats.

### Open citations (links to open the citations from this article in various online reference manager services)

1. Ecology
2. Evolutionary Biology

# Sexual dimorphism in trait variability and its eco-evolutionary and statistical implications

Susanne RK Zajitschek et al.
Research Article Updated

Biomedical and clinical sciences are experiencing a renewed interest in the fact that males and females differ in many anatomic, physiological, and behavioural traits. Sex differences in trait variability, however, are yet to receive similar recognition. In medical science, mammalian females are assumed to have higher trait variability due to estrous cycles (the ‘estrus-mediated variability hypothesis’); historically in biomedical research, females have been excluded for this reason. Contrastingly, evolutionary theory and associated data support the ‘greater male variability hypothesis’. Here, we test these competing hypotheses in 218 traits measured in >26,900 mice, using meta-analysis methods. Neither hypothesis could universally explain patterns in trait variability. Sex bias in variability was trait-dependent. While greater male variability was found in morphological traits, females were much more variable in immunological traits. Sex-specific variability has eco-evolutionary ramifications, including sex-dependent responses to climate change, as well as statistical implications including power analysis considering sex difference in variance.

1. Ecology
2. Microbiology and Infectious Disease

# Metapopulation ecology links antibiotic resistance, consumption, and patient transfers in a network of hospital wards

Julie Teresa Shapiro et al.
Research Article Updated

Antimicrobial resistance (AMR) is a global threat. A better understanding of how antibiotic use and between-ward patient transfers (or connectivity) impact population-level AMR in hospital networks can help optimize antibiotic stewardship and infection control strategies. Here, we used a metapopulation framework to explain variations in the incidence of infections caused by seven major bacterial species and their drug-resistant variants in a network of 357 hospital wards. We found that ward-level antibiotic consumption volume had a stronger influence on the incidence of the more resistant pathogens, while connectivity had the most influence on hospital-endemic species and carbapenem-resistant pathogens. Piperacillin-tazobactam consumption was the strongest predictor of the cumulative incidence of infections resistant to empirical sepsis therapy. Our data provide evidence that both antibiotic use and connectivity measurably influence hospital AMR. Finally, we provide a ranking of key antibiotics by their estimated population-level impact on AMR that might help inform antimicrobial stewardship strategies.