BEEhaviourLab: A high-throughput platform for sublethal stressor screening in insects

BEEhaviourLab is a low-cost, automated, high-throughput platform that integrates multimodal video and audio tracking to detect sublethal neurotoxic effects in diverse insect species, offering a scalable solution to improve chemical risk assessment beyond traditional mortality endpoints.

The Gist

Imagine you are a doctor trying to figure out if a new medicine is safe for your patients. For a long time, the only way to test this was to see if the medicine killed the patient. If they survived, the doctor assumed the medicine was fine. But we now know that a medicine can leave a patient alive but make them too dizzy to drive, too tired to work, or too confused to recognize their family. These are "sublethal" effects—harmful but not fatal.

For decades, scientists testing pesticides on insects (like bees) have been stuck with this "kill or no kill" approach. They count how many bees die, but they miss the subtle ways pesticides might make bees too clumsy to find flowers or too groggy to navigate home.

Enter BEEhaviourLab. Think of this not as a lab, but as a high-tech, 24-hour "surveillance reality show" for insects.

Here is the story of how this new tool works and what it found, explained simply:

1. The Setup: A Smart, Cheap, Automated Zoo

The researchers built a system called BEEhaviourLab. Imagine a row of small, clear plastic boxes (like tiny terrariums). Inside each box, a bee (or a fly) lives with a little feeder.

• The Cameras: Above each box is a Raspberry Pi (a tiny, cheap computer) with a camera and a microphone.
• The Brains: Instead of humans watching 24/7 (which is boring and expensive), the system uses AI (Artificial Intelligence). It's like a security guard that never sleeps.
• The Magic: The AI doesn't just watch; it understands. It can tell the difference between a bumblebee, a honeybee, a solitary bee, and even a hoverfly (which looks like a bee but is a fly) without needing to put stickers or tags on them. It's like a bouncer at a club who can instantly recognize everyone in the crowd just by how they walk and sound, even if they are wearing different outfits.

2. The Experiment: Testing a "Veterinary" Drug on Bees

The team wanted to test Moxidectin, a very common medicine used to kill parasites in cows, dogs, and horses.

• The Problem: When farmers treat their cows, the medicine ends up in the cow's poop. Bees and other insects often land on or live in this poop. Scientists suspected this medicine might be hurting bees, but they didn't know how or how much.
• The Test: They gave bumblebees tiny, precise doses of Moxidectin (some got none, some got a little, some got a lot). Then, they let the AI watch them for two days.

3. The Discovery: The "Drunk Bee" Effect

The results were shocking because the bees didn't die immediately. Instead, they acted like they were drunk or on heavy sedatives.

• The Silence: The AI listened to the bees. Healthy bees buzz constantly (it's how they talk and fly). The bees with the medicine stopped buzzing. They went silent.
• The Stumble: The AI watched them move. The bees with the medicine didn't just move slower; they stopped moving entirely for long periods. They were "zoned out."
• The Timing: Even more interesting, the medicine messed up their internal clocks. Bees usually sleep at night and work during the day. The medicated bees lost this rhythm, acting confused about what time of day it was.

4. The Big Reveal: The "Canary in the Coal Mine"

Here is the most important part of the story.

• The Old Way: If you only looked at death, you would say, "This medicine is safe for bees until you give them a huge dose."
• The New Way (BEEhaviourLab): The AI detected that the bees were behaving strangely at doses 30 times smaller than the dose that actually killed them.

The Analogy:
Imagine a car.

• The Old Test: You drive the car until the engine explodes. If it doesn't explode, you say, "This car is safe!"
• The New Test: You use a sensor to listen to the engine. You hear a weird rattle and notice the steering wheel is vibrating long before the engine explodes. You realize, "Hey, this car is dangerous to drive even though it hasn't broken down yet."

Why This Matters

This paper shows that we can now catch "poison" much earlier.

1. It's Cheaper and Faster: You don't need a team of scientists staring at screens. One computer can watch hundreds of bees at once.
2. It's Universal: The same system can watch bees, flies, and other bugs, making it easy to compare how different species react.
3. It Saves Ecosystems: By detecting these subtle "drunk" behaviors early, regulators can ban or restrict dangerous chemicals before they wipe out entire populations of pollinators.

In a nutshell: BEEhaviourLab is a smart, automated watchtower that listens to the "whispers" of insect behavior. It tells us that a chemical might be making our pollinators stumble and fall long before it actually kills them, giving us a chance to fix the problem before it's too late.

Technical Summary

1. Problem Statement

Current chemical risk assessment for insects relies heavily on mortality endpoints (lethal dose, LD50) derived from a limited set of model species (e.g., honey bees). This approach fails to capture sublethal effects, which are increasingly recognized as ecologically critical. Sublethal impacts on behavior, development, and communication can propagate to population and ecosystem levels but are difficult to quantify in a standardized, scalable, and regulatorily defensible manner.

Specifically, there is a lack of quantitative frameworks to:

• Link sublethal behavioral impairments to lethal endpoints across exposure gradients.
• Compare toxic effects across diverse insect taxa beyond standard models.
• Detect neurotoxic disruption (e.g., from veterinary pesticides like moxidectin) at concentrations far below lethal thresholds.

2. Methodology: The BEEhaviourLab Platform

The authors developed BEEhaviourLab, a low-cost, automated, high-throughput platform designed for long-duration, standardized behavioral phenotyping.

Hardware Architecture:

• Modular Nodes: The system consists of distributed recording units built around Raspberry Pi 4B microcontrollers.
• Sensors: Each node integrates synchronized video (Raspberry Pi NoIR v2 camera), audio (RØDE smartLav+ microphone), and environmental monitoring (DHT22 temperature/humidity sensor).
• Lighting: Programmable LED strips allow for daylight (white) and night (red) modes to simulate natural circadian cycles.
• Control: A central computer coordinates multiple nodes via Wi-Fi/SSH, managing experimental protocols, metadata, and data transfer.

Software & Analysis Pipeline:

• Computer Vision: Insects are tracked without physical tags using YOLOv8n (You Only Look Once version 8 "nano"), a lightweight object-detection model.
  • Training Strategy: A "Model 1" trained only on bumble bees failed to generalize. A "Model 2" trained on a diverse dataset of four species (bumble bees, honey bees, ivy bees, and hoverflies) achieved high accuracy (>0.98 mAP) across all taxa, demonstrating robustness to morphological and kinematic differences.
  • Tracking Logic: Uses BoT-SORT for initial tracking, followed by a custom Hungarian algorithm step to enforce a fixed population size (IDs 1–N), preventing ID swaps or creation/deletion errors during occlusion.
• Audio Analysis: A custom pipeline processes raw audio (resampled to 16 kHz, bandpass filtered 200–8000 Hz). It uses energy-based thresholds (median + 5×MAD) to classify frames as "buzzing" or "silent," calculating the proportion of time spent buzzing.
• Statistical Framework:
  • Two-Part Mixed-Effects Models: To separate behavioral state (probability of moving/buzzing) from performance (speed/persistence given activity). This avoids the statistical pitfalls of averaging data where many observations are zero.
  • Circadian Analysis: Harmonic mixed-effects models (sine/cosine terms) to quantify 24-hour rhythmicity.
  • Sublethal Toxicity Index: A composite metric combining deviations in movement probability, speed, and buzzing probability, scaled by the natural variability of control groups.

3. Key Contributions

1. Scalable, Tag-Free Tracking: Demonstrated that a single, lightweight deep learning model can accurately track multiple untagged insects across different species (social bees, solitary bees, and hoverflies) simultaneously.
2. Multimodal Integration: Successfully integrated video and audio data to capture distinct behavioral dimensions (locomotion vs. acoustic signaling) within a unified analytical framework.
3. Methodological Innovation: Introduced a "two-part" statistical approach that distinguishes between the likelihood of an action occurring and the intensity of that action, revealing nuances in sublethal toxicity that single-metric analyses miss.
4. Open Source Framework: The hardware designs, control software, and training pipelines are fully open-source, lowering barriers to entry for behavioral toxicology.

4. Key Results

A. Cross-Species Behavioral Profiling:

• The platform successfully quantified species-specific circadian patterns.
• Bumble bees showed high movement probability but moderate speed; Honey bees showed lower probability but higher speed when active; Hoverflies showed low activity and low speed.
• Acoustic analysis revealed species-specific differences in buzzing frequency and duration, with bumble bees exhibiting the highest buzzing activity.

B. Moxidectin Toxicity in Bumble Bees (Bombus terrestris):

• Dose-Dependent Effects: Acute contact exposure to moxidectin caused significant, dose-dependent reductions in both locomotion and buzzing.
• State vs. Performance: The primary effect was an increase in the probability of inactivity (bees stopped moving/buzzing entirely). Among active bees, speed and buzzing persistence decreased more modestly.
• Circadian Disruption: Higher doses distorted the amplitude and phase of circadian rhythms, indicating neurobehavioral dysfunction beyond simple sedation.
• Sensitivity Comparison:
  • Behavioral LOEC (Lowest Observed Effect Concentration): 2.73 ng/bee (based on the multimodal toxicity index exceeding 1 SD of control variability).
  • Lethal LD50: 81.2 ng/bee (48-hour mortality).
  • Feeding Suppression: Detected only at 78.2 ng/bee.
  • Conclusion: Behavioral endpoints detected toxicity at concentrations >30 times lower than lethal endpoints and an order of magnitude lower than feeding suppression.

5. Significance

• Regulatory Impact: BEEhaviourLab provides a practical, standardized path to integrate sublethal behavioral endpoints into chemical risk assessment, addressing a critical gap in current ecotoxicology.
• Early Warning System: The platform demonstrates that behavioral metrics (especially multimodal ones) are far more sensitive indicators of neurotoxic stress than mortality or feeding assays. This is crucial for veterinary pesticides like moxidectin, which persist in the environment and affect non-target pollinators.
• Ecological Relevance: By detecting subtle disruptions in locomotion and communication at environmentally relevant concentrations, the system helps predict population-level risks that traditional mortality assays would miss.
• Future Directions: The framework is extensible for unsupervised motif discovery, state-based modeling, and integration with machine learning for automated annotation of complex behaviors (e.g., grooming, feeding), paving the way for next-generation, high-throughput toxicological screening.