香京julia种子在线播放

Hydrogels containing a vascular network were designed using Blender, an open-source 3D CAD software (Blender Foundation, Amsterdam, Netherlands). Designs were exported in STL format, and photomasks were generated using Creation Workshop software (https://makershop.co/envision-labs/). These photomasks were modified using a custom MATLAB script that adjusted the grayscale value of photomasks in select regions to 60% intensity. The resulting custom photomasks were used to fabricate hydrogels via 3D printing.

Hydrogels were fabricated as previously described (Grigoryan et al., 2019) with some modifications. Pre-hydrogel solution was prepared using 3.25% wt/vol 3.4 kDa poly(ethylene glycol) diacrylate (PEGDA), 10% wt/vol gelatin methacrylate (GelMA), 10% wt/vol glycerol, 17 mM Lithium phenyl-2,4,6-trimethylbenzoylphosphinate (LAP), and 3.25 mM tartrazine (Sigma-Aldrich, St. Louis, MO, United States). This solution was 3D printed via digital light processing (DLP) on an Lumen X bioprinter (Cellink, Boston, MA, United States). The layer height for printing was fixed at 50  μ m and the bioink was exposed to light for 12.0 s per layer with approximately 21 mW/cm² in intensity. Three hydrogels were printed simultaneously for a total print time of 23 min. Following printing, hydrogels were allowed to remain in PBS for a minimum of 2 days before experiments were performed. PBS was replaced several times as excess tartrazine leached out of the hydrogels until the PBS solution was completely clear. To make a model system that would enable mosquito feeding, hydrogels with a simple vascular architecture were used to mimic a skin sample. All of the hydrogels described here were used in one experiment and were discarded after use.

Research-grade defibrinated animal blood was purchased from HemoStat Laboratories (Dixon, CA, United States). All feeding experiments used either defibrinated chicken, sheep, or cow blood. Blood was loaded into a 30 mL syringe and was placed in a syringe pump (New Era Pump Systems, Inc., Farmingdale, NY, United States). Blood was flowed at a rate of 100  μ L/min through Tygon^® tubing (Cole-Parmer, Vernon Hills, IL, United States) for the duration of all mosquito feeding experiments. Tight fluidic connections were achieved via blunt syringe tips (Nordson Corporation, Westlake, OH, United States). Blunt 90-degree needle tips were modified by removing the metal tip from the luer lock connector using pliers, and tubing was connected directly to the detached metal tip. These tips and the hydrogels were held in place with 3D printed perfusion chambers printed using poly(lactic acid) (PLA) filament (Ultimachine, South Pittsburg, TN, United States) on a Prusa i3 MK3S+ printer (Prusa Research, Prague, Czech Republic). Perfusion chambers were designed using a pre-defined protocol (Kinstlinger et al., 2021) with subsequent modifications using Fusion 360 software (Autodesk, San Rafael, CA, United States). Tubing was used to connect multiple hydrogels in series such that a single syringe pump could supply blood flow to up to six hydrogels.

Printed circuit boards (PCBs) (Conclusive Engineering, Katowice, Poland) were designed with six evenly spaced sites for attachable hydrogels arranged in a 3 × 2 grid. Each site contained a resistive heater, LEDs, and a thermistor. The settings for the heating elements and LEDs were controlled using custom software. Individual heating elements and LEDs were independently controlled according to the requirements of particular experiments. The platform was positioned vertically and held in place with a combination of custom 3D printed parts, T-slotted framing rails, and nuts and bolts. The 3D printed perfusion chambers securing needle tips and hydrogels were bolted to the PCBs. These chambers were printed using white filament to provide contrast against the black PCBs in an effort to attract mosquitoes. Machined parts were purchased from McMaster-Carr (Atlanta, GA, United States). Raspberry Pi 4 Model B computers were outfitted with PoE+ HAT attachments to support power delivery and data communication among Raspberry Pis via ethernet cables. Both the Raspberry Pi computers and PoE+ HATs were purchased from Adafruit (New York, NY, United States). A PoE+ network switch (Netgear, San Jose, CA, United States) supplied power to all Raspberry Pi computers through the ethernet cables. Raspberry Pi camera boards and lenses (Arducam, Nanjing, China) were fixed in place with adjustable 3D printed parts. Each Raspberry Pi camera was directed at a single hydrogel housed in a perfusion chamber. Raspberry Pis recorded video directly to their respective hard drives using custom Python scripts.

Feeding experiments were performed using lab raised Aedes aegypti (Ae. aegypti) Rockefeller strain mosquitoes. The mosquitoes were raised in a controlled insectary at 25°C–27°C, 75%–80% relative humidity and a photoperiod of 16:8 (light:dark) hours. Dried egg papers less than 3 months old were hatched in deoxygenated water and transferred to polypropylene Nalgene pans at a density of 200 larvae/2 L water and fed on algae tablets (Hikari, Kyorin Food Ltd. Japan). Pupae were transferred to beakers and placed in cages to emerge; approximately 500 adult males and females were kept in 30 × 30 × 30 mesh cages (BugDorm, MegaView Science Co., Ltd. Taiwan) and allowed to mate freely. Adults were provided 10% wt/vol sucrose solution delivered by cotton wick. Sucrose solution in select cages was removed overnight prior to feeding experiments.

Hydrogels were loaded into 3D printed perfusion chambers and tight fluidic connections were created to support perfusion (see above). These perfusion chambers were bolted to a PCB, which was held in place vertically. After flowing blood through all hydrogels to ensure their structural integrity, a solid glass cage was fitted over the PCB and 20–30 female mosquitoes were introduced into the cage through a hole. Mosquitoes were selected by mechanical aspiration of the subset of the colony attracted to a human hand held close to a mesh wall of the colony cage. Several male mosquitoes were included with aspirated females to encourage feeding. The cage’s edges did not allow mosquitoes to escape, and the hole was plugged with a cotton ball after mosquitoes were fully released into the cage. A Raspberry Pi camera was aimed at each hydrogel so that mosquito behavior and feeding could be recorded. Feeding experiments were stopped after 30–45 min of activity and mosquitoes were subsequently isolated for manual quantification of feeding and egg counts.

A total of 3 mosquito feeding cages were prepared the same way as for feeding experiments, with minor changes. Hydrogels in each of the three cages were perfused with either defibrinated blood, red India ink, or PBS. The hydrogels in a given cage all received the same liquid, and each of the three liquids was heated to 37°C before entering the cage. Approximately 20–30 Ae. aegypti mosquitoes were introduced as normal and were observed for feeding behavior.

Similar to meal choice experiments, 3 feeding cages were prepared with distinct experimental conditions. One of the cages contained 6 hydrogels coated with 10 mg/mm² of 25% DEET (N,N-diethyl-meta-toluamide), another contained 6 hydrogels coated with 10 mg/mm² of a 30% concentration of a plant-based repellent derived from the oil of lemon-eucalyptus plants (OLE), and the last cage contained 6 uncoated hydrogels and served as a control. All hydrogels were perfused with blood that was heated to 37 °C before entering the cage. 20–30 female Ae. aegypti mosquitoes were introduced into each cage and activity was observed for 30–45 min. The mosquitoes introduced to the control, DEET, and OLE groups were selected from a population of approximately 500 mosquitoes hatched on the same day based on their attraction to a human hand held near a mesh wall of the colony cage. Repellent experiments were repeated for five total replicates to reduce noise due to sample differences in feeding activity. Each experimental replicate used a different population of mosquitoes to minimize batch effects.

Following the platform assembly and blood perfusion methods described above, videos of mosquito feedings were collected. During the early stages of model development, approximately 50 still images were extracted from these videos. These images were pre-processed by downsizing to 416 × 416 pixels to reduce computation time and by disregarding images that do not contain any mosquitoes. Training data were supplemented with augmented images that contained several transformations (see Supplementary Methods) to improve model performance. Images were manually labeled with bounding boxes for mosquitoes, abdomens of feeding or engorged mosquitoes, and abdomens of non-feeding mosquitoes (see Supplementary Methods). A machine learning model was trained and deployed via a custom algorithm developed by Roboflow, Inc. (San Francisco, CA, United States, see Supplementary Methods), and performance was evaluated based on precision, mean average precision (mAP), and recall. New images were labeled and added to the model as experiments progressed. Initial training data was heavily skewed toward representing non-feeding mosquitoes, as the number of frames containing feeding or engorged mosquitoes was drastically smaller. To compensate for this issue, the model was deployed on unlabeled videos to identify images containing feeding or engorged mosquitoes. These images were automatically uploaded to the dataset for future incorporation if they were less than 97% similar to previously analyzed frames. Our final dataset consisted of 3,292 images, which were segmented into training, validation, and testing datasets in a ratio of 7:2:1.

Machine learning model outputs were reduced to a single number per label type (i.e. mosquito, non-feeding, and feeding) for each video to quantify activity. Raw counts of each label’s abundance were tabulated for each frame, creating a function of label abundance over time. These counts were then summed to obtain the total number of labels throughout a video. This metric can be represented by the following expression: A c t i v i t y = ∑ i = 1 n D t Where n represents the number of frames in a video and D(t) represents the number of model detections at time t for a given class. Because mosquito presence and incidence of biting are the most relevant indicators of repellent effectiveness, we chose to only analyze mosquito and feeding labels with our metric. In some cases (noted in Results) we limited our video data to a region of interest that included hydrogels and their immediate surroundings to ignore mosquitoes within the field of view but are far away from hydrogels. To decrease computation time and resources, videos were often uniformly downsampled by a factor of 16 (i.e., every 16th frame was kept). Our current computational setup allows us to use our model to interpret videos at a rate of 1.67 frames per second, and that analysis rate could be amplified by running several analyses in parallel on different machines.

Hydrogels were fabricated as described above. To visualize the vascular channel and verify its patency and distance from the hydrogel surface, a solution containing a 1:200 dilution of red fluorescent beads (Magsphere, Pasadena, CA, United States), Irgacure 2959 (Ciba), and 20% wt/vol 6 kDa PEGDA with PBS as a diluent was injected into the channel and crosslinked under UV light. The hydrogel was then inserted into a 3D printed chamber and carefully sliced with a razor blade. The 3D printed chamber featured slots for the razor blade for the sake of consistency. The hydrogel cross sections were imaged on a Ti-E inverted microscope (Nikon, Melville, NY, United States) and a Zyla 4.2 sCMOS camera (Andor, Belfast, United Kingdom). The resulting images were linearly adjusted for brightness and contrast to improve visualization.

Based on the results of pilot experiments, hydrogels were softened in precisely patterned regions directly above the vascular channels to ease mosquito feeding. To visualize the differences between these patterned regions (termed “compliant regions” for their increased compliance) and unaltered regions on hydrogel surfaces, pre-hydrogel solution was supplemented with FITC conjugated to 150 kDa dextran at a concentration of 1 mg/mL. Hydrogels were printed in sets of three, each of which contained softened compliant regions at regular intervals on hydrogel surfaces. For each set, one of these hydrogels was sliced and imaged (see above) immediately after printing, and the remaining two were stored in PBS and were imaged at 2 and 5 days post-printing, respectively. Hydrogels were imaged using the same exposure settings for all three timepoints, and brightness was uniformly adjusted to assist visualization. Differences of hydrogel fluorescence intensity between compliant and adjacent control regions of hydrogels were compared at 0, 2, and 5 days post-printing. Quantification of fluorescence intensity was performed by summing the pixel intensity within a consistently sized rectangular region over compliant regions and control regions of the hydrogel. Every compliant and control region was measured this way, and regions within the same hydrogel were treated as technical replicates.

Statistical analyses of all data were performed using R software and GraphPad Prism 9.4. Data for the diffusion assay were collected for N = 6 hydrogels on days 0 and 2 post-printing, and for N = 5 hydrogels on day 5 post-printing. It was determined that fluorescence readings were not normally distributed after day 0 (p < 0.01, Shapiro-Wilk test), so measurements were compared using a non-parametric Mann-Whitney test for each day. The data are shown in Figure 3 as mean ± S.D. For experiments consisting of multiple feeding cages with isolated conditions (i.e. meal choice experiments and repellent screening experiments), a chi-square test was used to compare feeding preferences of multiple mosquito populations. Comparisons of egg counts between mosquitoes feeding on hydrogels (N = 85) vs the laboratory standard as a control (N = 45) were performed using a two-tailed unpaired t-test (p > 0.05) after confirming that the data for both groups was sufficiently normally distributed (p > 0.05, Shapiro-Wilk test).

The hydrogels that served as a mosquito food source consisted of a single channel that sweeps back and forth several times in a serpentine pattern (Figure 1A). The switchbacks of the serpentine were spaced far enough apart so that rupturing during perfusion was unlikely, but close enough to maximize the available feeding area for mosquitoes. After several iterations of testing, it was determined that vessels 400  μ m from the feeding surface would consistently print and support perfusion while still enabling mosquitoes to easily reach the artificial blood vessel during feeding.

To gather a rich dataset of mosquito feeding behavior, we sought to record videos of mosquitoes feeding on our hydrogels. After several rounds of optimization (see Supplementary Methods), we adopted a system that allowed us to record videos to a series of Raspberry Pi computers that allow recording at 1920 × 1080 resolution at 30 frames per second (fps) while occupying approximately 2.5 L worth of space (Figure 1B). Recording video data of mosquito feeding experiments enabled us to develop a machine learning model and perform experiments with multiple mosquitoes, perfused liquids, and surface repellents (Figure 1C, D).

It has been well documented that mosquitoes rely on a variety of cues to find hosts, one of which is heat (Potter, 2014; Liu and Vosshall, 2019). We therefore hypothesized that heating hydrogels during experiments could improve mosquito attraction to our skin mimics. Additionally, illuminating hydrogels might enhance photography and data collection, although colored light may also affect mosquito attraction (Gjullin, 1947; Van Breugel et al., 2015; Liu and Vosshall, 2019; Alonso San Alberto et al., 2022). To incorporate heating and lighting capabilities into our platform, we designed a custom printed circuit board (PCB) with heating elements and LED lights at six fixed positions. Our preliminary experiments revealed that mosquitoes feeding on a horizontal surface will orient themselves at a variety of body angles relative to a stationary camera, but mosquitoes feeding on a vertical surface tend to orient themselves so that their abdomens are pointing down. To support a vertical feeding surface, our PCB features through holes to enable bolting hydrogel perfusion chambers to it. This capability improved consistency of hydrogel placement and firmly anchored them to the PCB, allowing us to position the entire PCB vertically. Video data suggested that mosquitoes were more likely to orient themselves the same way on this vertical surface, so this angle was maintained for future experiments.

Next, we leveraged our PCB’s heating capabilities to quantify the impacts of heat on mosquito attraction. Prior to performing experiments, we empirically determined that PCBs were able to heat hydrogels to a maximum temperature of 34 °C, which is approximately equal to skin temperature (Metzmacher et al., 2018). We then designed an experiment that offered mosquitoes an equal choice of three unheated hydrogels and three hydrogels heated to 34 °C to see whether heating increased attraction. After introducing groups of laboratory-reared Ae. aegypti mosquitoes, we observed that the mosquitoes appeared distracted by the perfusion chambers and nuts and bolts securing the chambers to the PCB and exhibited host-seeking behavior on these surfaces as well. We hypothesize that heat transfer inefficiencies heated up hardware surrounding the hydrogels, resulting in unexpected mosquito attraction to these surfaces. Surprisingly, repeating this experiment with Anopheles quadrimaculatus (An. quadrimaculatus) mosquitoes yielded less distracted behavior and revealed that heated hydrogels are required to elicit feeding for this species (Supplementary Figure S1). We therefore concluded that our PCB’s heating successfully increased mosquito attraction in some species but was experimentally detrimental for others.

After developing a feeding platform and establishing a consistent experimental procedure, a total of over 180 recordings of mosquito feeding experiments were collected over a period of 6 months. These videos were used to train a machine learning object detection model to identify mosquitoes (Figure 2A). Some videos exhibited slightly different camera angles and lighting, which is beneficial because it does not overfit to a particular experimental setup, expanding the applicability of the model to a variety of laboratory settings. This machine learning model was developed using Ae. aegypti mosquitoes during experiments that did not use PCB heating.

Our object detection model was assessed by applying it to video files and analyzing the returned predictions. We first trained our model to identify mosquitoes within the camera’s field of view. Next, we further characterized the behavior of visible mosquitoes by training the model to identify two additional objects: feeding/engorged mosquito abdomens and non-feeding mosquito abdomens. When assessed in aggregate, the model’s mean average precision (mAP) of 92.5% implies that the model performs well across all three classes (see Supplementary Methods). The model’s relatively high recall of 89.2% indicates that it has a low false negative rate, and its 92.1% overall precision indicates that it has a low false positive rate as well. In the case of precision, more meaning can be extracted when examining each class individually.

The developed model can identify mosquitoes with 98% precision, meaning that 98% of all mosquito detections made by the model are accurate. The effectiveness in distinguishing feeding or engorged from non-feeding mosquitoes was assessed using the detection classes for feeding and non-feeding mosquito abdomens (see methods). Using this approach, the model was able to correctly identify feeding or engorged mosquitoes with 94% precision and non-feeding mosquitoes with 90% precision. It should be noted that these classes are treated independently, so it is possible for a single mosquito to be labeled as both feeding and non-feeding simultaneously. However, if these classes are combined into a single “abdomen” label regardless of feeding status, the model identifies abdomens correctly with 96% precision. Although the model makes some mistakes, it often performs well in challenging situations such as when mosquitoes are partially obscured or out of focus.

After developing a model that accurately identifies three objects of interest, we sought to use this model to automatically parse mosquito feeding videos and extract meaning from them. The performance statistics we obtained from our model were impressive, but it was unclear whether the model’s inaccuracies would render it useless for analyzing videos. Starting with small test cases, we first evaluated the model’s performance on trimmed videos that were cropped to only show a single feeding mosquito. Overall, mosquitoes analyzed using this method were identified by the model as non-feeding while their abdomens remained slim and feeding when their abdomens became swollen and red (Figure 2B). Visualizing model predictions for an entire video revealed the existence of a transition period, during which time the abdomen was classified as both feeding and non-feeding simultaneously (Figure 2C). This transition period roughly corresponds to the time during which a human observer would have difficulty telling whether a mosquito has begun to feed or not. Finally, once the abdomen undergoes substantial change from its original state, the model consistently identifies the abdomen as that of a feeding mosquito.

While we were pleased with our model’s predictions for individual mosquitoes over the course of feedings, limiting analysis to a particular ROI requires manual intervention and prior knowledge of where feeding mosquitoes are in a video. Although this intervention is minimal, it still limits the throughput and applicability of our technology. To further improve automation, we applied our model to an entire unmodified video to return raw counts of each label per frame (Figure 2D). We displayed the data in a similar way as with individual mosquitoes, but added color gradients to show label density throughout time. This new visual representation reveals the density of mosquitoes and feeding events for any given frame and elucidated trends in mosquito behavior for feeding experiments. It is also possible to visualize the time and duration of specific feeding events if they are temporally separated. This visualization tool provides informative qualitative and quantitative data about collected videos and obviates the need for researchers to manually watch videos to make similar observations.

Early attempts to induce mosquito feeding with these PEGDA-GelMA hydrogels revealed that mosquitoes struggled to puncture the hydrogel deeply enough to reach the channel. A mosquito proboscis is long enough to reach artificial vasculature, implying that vessel depth is not the issue. To address this problem, the surface of these hydrogels was selectively softened directly above the blood channel to increase their compliance (Figure 3A, see methods). Analysis of these compliant regions showed that they are more porous than adjacent regions of the hydrogel surface, as confirmed by a diffusion assay (Figure 3B, C, see methods). Although this change did not drastically change the overall number of feedings, mosquitoes appeared to have less difficulty consuming a blood meal. We therefore incorporated compliant regions into all subsequent hydrogels. The fact that mosquitoes were able to feed on our hydrogels confirmed the feasibility of our material as a mosquito food source.

Although we previously demonstrated that mosquitoes could feed on our hydrogels, it was unclear whether mosquitoes were attracted to the blood itself, the color contrast between the blood and the translucent hydrogels, or a property of the hydrogels themselves. To disentangle these variables, we tested the extent to which blood successfully attracted mosquitoes to our hydrogels. In three independent feeding cages, we offered Ae. aegypti mosquitoes hydrogels perfused with blood, red India ink, or PBS and observed their feeding tendencies. In this experiment, the PBS acted as a negative control—since it does not contain any sugars and is isotonic with the hydrogels, any mosquito attraction to these hydrogels is likely not due to the perfused liquid. This experiment showed that mosquitoes only fed on the hydrogels containing blood, as expected (Figure 4; Table 1). Although mosquitoes spent a substantial amount of time around the hydrogels in both of the other experimental groups, none of them engaged in host-seeking behavior. We therefore concluded that a chemical component of blood, rather than its visual appearance, attracts mosquitoes. This experiment also confirmed that hydrogels do not inherently attract mosquitoes, indicating that no further optimization of hydrogel composition is necessary.

To validate our feeding platform as a screening assay, we assessed the effectiveness of two available mosquito repellents as feeding deterrents. To ensure independence of experimental conditions, we increased the scale of our assay to include three isolated feeding cages which contained hydrogels coated with DEET, OLE, or PBS (Figure 5A, see methods). These screening experiments demonstrated that the DEET and OLE repellents significantly negatively impacted mosquito attraction (Table 2). There are no visible differences between hydrogels in each experimental group, indicating that DEET and OLE repel mosquitoes by non-visual mechanisms. In both experimental groups, none of the mosquitoes fed on any of the hydrogels. This experiment performed as expected, with mosquitoes avoiding hydrogels coated with repellent while feeding on control hydrogels.

We next investigated whether our machine learning model could expedite data analysis for our repellent screening experiments. Although previously explored visual representations of our model’s predictions can show striking differences between mosquito behavior in different videos (Supplementary Figure S2), further quantification would more convincingly demonstrate behavioral differences in unique experimental groups. To that end, we developed a metric to reduce our model predictions to a single numerical output for each label (see methods) using the frequency of machine learning model predictions as a proxy for mosquito activity and feeding activity. We then used this metric to analyze our repellent screening data. To filter out irrelevant data that manual tabulation would ignore anyway, we restricted our model’s detections to an area encompassing only hydrogels and their immediate surroundings (see methods). This approach yielded the same results and overall conclusions as with our manually tabulated data—both DEET and OLE effectively repelled mosquitoes and discouraged feeding on hydrogels (Figure 5C). Therefore, reducing our data to a single metric increases autonomy, decreases interpretation subjectivity, and improves throughput without altering experimental conclusions. The fact that this metric agrees with our manually tabulated observations (Table 2) and implies the same conclusions regarding the effectiveness of repellents suggests that our metric could be useful for other experiments in the future.