University of California researchers have applied machine learning to identify scents that repel honey bees, a discovery with potential to protect these vital pollinators from pesticide exposure. The team’s approach goes beyond traditional methods, predicting repellent compounds and then validating those predictions both in laboratory settings and within real agricultural fields.
“This important manuscript reports a very interesting view of how pesticides can be toxic to beneficial insects like the honeybee,” says Barbara F Baer-Imhoof of the University of California, Riverside. This research impacts ecology, pest control, and sensory biology by offering a new way to rationally discover aversive volatiles for insect control.
Machine Learning Predicts Honey Bee Repellents from Chemical Structures
An expanded computational model screened over 50 million compounds, ultimately identifying more than 130 candidate repellent substances for honey bees. Behavioral tests demonstrated a high degree of predictive accuracy for these compounds, a result subsequently validated through field studies assessing repellency against foraging bees near pesticide-treated crops. This dual validation, in controlled conditions and real-world environments, underscores the potential for machine learning to accelerate the discovery of effective insect repellents.
The challenge of identifying these repellents experimentally is considerable; despite two decades of research into insect olfactory receptors, the U.S. Environmental Protection Agency has not registered a new repellent odorant for insect control products. Researchers used machine learning, building on previous success in predicting mosquito repellents from chemical structure and modeling human olfactory behavior, achieving prediction accuracy for approximately 140 odor characters.
This prior work, detailed in publications from 2020 and 2024, provided a foundation for applying similar techniques to honey bee repellency. The team’s cheminformatics pipeline began with a review of existing literature on honey bee olfactory behavior, assembling a training set of 184 compounds, later expanded to 203, to form the basis for iterative machine learning model training and testing. “This ORCID iD identifies the author of this article,” a statement included with the publication, highlights the importance of researcher identification in modern scientific publishing.
The resulting models were then used to screen a vast chemical space, prioritizing compounds with the highest predicted repellency. Further investigation utilized video tracking protocols to analyze honey bee responses to these deterrent chemistries, providing detailed behavioral data.
Analysis of mean preference indexes in two-choice assays revealed that bees consistently favored moving away from the repellent-treated side, confirming the predictions generated by the machine learning models. This approach offers a powerful alternative to traditional methods, particularly for insects where limited behavioral data exists, and could significantly impact areas including ecology, pest control, and sensory biology. The researchers, based at the University of California, Riverside demonstrated that machine learning, combined with rigorous testing, provides a viable pathway for the rational discovery of aversive volatiles for insect control.
Iterative Modeling Refines Aversive Valence Prediction Accuracy
The predictive power of machine learning models identifying bee-repellent compounds was significantly refined through iterative testing with live insects, achieving an average area under the receiver operating characteristic curve (ROC AUC) of 0.88 in computational validation. This level of accuracy allowed researchers to distinguish between odorants triggering aversion and those that do not, a step beyond simply identifying potential candidates based on chemical structure alone.
The team’s approach involved generating behavioral data for both honey bees and Drosophila, using this information to improve the initial computational model’s ability to predict aversive valence. Following model refinement, a screen of over 50 million compounds yielded more than 130 repellent candidates, demonstrating the efficiency of the computational pipeline. Laboratory testing with honey bees confirmed a high rate of predictive success, but the validation process extended beyond controlled environments.
Additional assays utilizing freely foraging honey bees in field conditions substantiated strong repellency from the top seven candidates, suggesting a high probability of deterring bees from pesticide-treated crops. This real-world confirmation is critical, as it moves beyond theoretical predictions to demonstrate practical application.
The computational modeling began with the calculation of 5290 physicochemical features for both training and test chemicals, initially for 184 compounds and later expanding to 203. From the initial training set, 45 features were identified as important predictors of aversive valence, informing the selection of machine learning algorithms including regularized random forest, gradient-boosted decision trees, and a nonlinear support vector machine.
The team employed cross-validation techniques, dividing data into training and test sets to assess performance and graphically represent predictive success using ROC AUC. “Most odorants reaching a bee’s antennae activate or inhibit several of the >160 olfactory receptors at the same time,” the researchers noted, highlighting the complexity of the honey bee olfactory system and the challenge it presents to modeling.
The researchers acknowledge that a complete understanding of the honey bee olfactory system would require years and substantial funding, but their 3D structure-based computational approach effectively addresses a critical gap in identifying novel volatiles that trigger specific behavioral responses. The results from both laboratory and field assays demonstrate that a computational approach can successfully model honey bee aversive valence solely from chemical structure, and can efficiently screen millions of chemicals to find strong repellents. Data from honey bee robbing assays and evaluations of chemicals as attractants and repellents further support these findings, providing a comprehensive dataset for future research.
Computational Screening Identifies 130 Repellent Candidates
This broadened search, enabled by an improved predictive model, represents a significant increase in potential aversive volatiles compared to earlier, more limited investigations. Researchers then subjected these candidates to rigorous behavioral validation, first in laboratory settings and subsequently with free-flying honey bees exposed to field conditions. Laboratory assays demonstrated a high degree of predictive success for the machine learning approach, confirming its ability to accurately forecast repellency based on chemical structure and predicted olfactory interactions.
Further substantiation came from field testing, where the top seven candidates consistently repelled foraging bees, suggesting a strong probability of protecting pesticide-treated crops from unwanted visitation. The team estimated vapor pressure for each candidate, recognizing that evaporation rates influence the duration of repellent effectiveness in open fields.
Evaluations excluded plates where bees did not consume honey from either side, ensuring that only engaged bees contributed to the data, and workers from multiple colonies were used to account for potential variations in behavioral responses. To further validate findings, researchers elicited robbing behavior in honey bees, a natural foraging strategy involving the theft of unguarded honey, to assess repellency in a competitive context.
The updated repellent predictive model, based on area under the curve values showed improvement over previous iterations, allowing for a more refined screening process. The study’s findings have implications for a range of disciplines, extending beyond basic insect behavior to encompass practical applications in agriculture and pest management.
The ability to computationally identify potential repellents offers a powerful alternative to traditional, often broad-spectrum, insecticides, potentially reducing harm to beneficial insects and minimizing environmental impact. This work, however, demonstrates that even with current knowledge, machine learning can significantly accelerate the discovery of effective and targeted insect control strategies.
Laboratory Validation Confirms High Predictive Success of Model
88, demonstrating its capacity to differentiate between volatiles eliciting aversion and those that do not. This level of accuracy facilitated the in silico screening of approximately 45 million small molecules sourced from the MolPort database, ultimately identifying over 130 potential repellent candidates for further investigation. Researchers then prioritized the top seven candidates for behavioral testing, moving beyond purely computational predictions to assess efficacy in a controlled laboratory environment.
To ensure the robustness of these findings, the team excluded data from trials where bees failed to engage with the honey sources, focusing solely on instances of clear behavioral response. This meticulous approach extended to field assays, where freely foraging honey bees demonstrated strong repellency to the same compounds, suggesting a potential application in protecting pesticide-treated crops from unwanted bee visitation.
The iterative refinement of the predictive model proved important to its performance; incorporating data from newly validated compounds into the training set consistently improved its accuracy. Specifically, adding data from newly tested compounds led to enhanced computational validation results, as measured by AUC values.
Field Assays Demonstrate Repellency to Foraging Bees
Field trials confirmed the predictive power of a machine learning model identifying compounds that deter foraging honey bees, extending laboratory successes into real-world conditions. Analyses of bee behavior revealed strong repellency from seven candidate compounds tested in robbing assays, demonstrating a potential strategy for protecting pesticide-treated crops from unwanted visitation. These assays elicited robbing behavior, a natural phenomenon where bees steal unguarded honey or target weaker colonies, providing a realistic assessment of repellent effectiveness.
For 15 of the 28 test compounds, bees demonstrably consumed significantly less honey on the treatment side compared to control sides, a statistically significant difference supporting the model’s predictions. Additional testing involved compounds, with significant repellency still observed. The consistency of these effects was also confirmed across multiple honey bee colonies. Beyond honey bees, researchers assessed the compounds’ impact on fruit flies, Drosophila melanogaster, using a trap-based assay to determine broad-spectrum repellent activity.
Several compounds exhibited low levels of repellency to fruit flies, with less than a 50% reduction in trap entry compared to solvent controls, suggesting a potential for selective targeting of honey bees without harming other beneficial insects. Detailed observation of bee behavior involved filming assays and counting individuals on different wax foundations in five-minute increments, allowing for precise quantification of repellent effects. Kruskal-Wallis tests confirmed significantly fewer bees on treated wax foundations compared to controls, providing statistical support for the observed repellency.
The team reports that the assays were concluded when bees ceased to be attracted to the untreated controls, ensuring a clear distinction between repellent and attractive stimuli. Each batch utilized workers from the same colony, maintaining consistency within individual tests, and for promising candidates, bees from multiple colonies were tested, with a minimum of six participating plates each.
Pesticide Contact Reduction as Solution to Pollinator Decline
Beyond simply documenting the detrimental effects of pesticides on honey bee colonies, researchers are now focusing on preemptive strategies to minimize contact between pollinators and harmful chemicals. Investigations reveal that manipulating olfactory cues offers a promising avenue for reducing bee visitation to treated crops, thereby lessening pesticide exposure and safeguarding colony health. This approach acknowledges that bees rely heavily on scent to navigate their environment and locate resources, presenting an opportunity to guide them away from danger.
The identification of effective repellent compounds is not solely about finding substances bees dislike; it is also about ensuring those compounds do not inadvertently attract pest insects. To assess this, the team employed trap-based assays using Drosophila melanogaster, commonly known as the fruit fly. Several tested compounds exhibited less than a 50 percent reduction in trap entry compared to controls.
This selectivity is important, as a successful repellent strategy must protect beneficial insects without exacerbating pest problems in agricultural settings. This research builds on a foundation of understanding the complex olfactory system of honey bees, which comprises over 170 odorant receptor genes and 21 IR genes responsible for detecting volatile chemicals. The team’s machine learning pipeline accelerates the discovery of insect repellents and expands knowledge of the underlying physicochemical principles.
The overall species-selective approach presented will not only identify additional novel repellent chemicals for honey bees but also inspire exploration of chemical space for repellent strategies applicable to other animals. The implications of this work extend beyond immediate pest control, impacting multiple scientific disciplines.
Publications cited within the study highlight the interconnectedness of pollinator health, nutrient levels in food supplies, and the synergistic effects of multiple stressors, including pathogens and pesticides, on bee survival. The team’s work, therefore, contributes to a growing body of knowledge essential for addressing the global decline in pollinator populations, a decline linked to factors including pesticide use, habitat loss, and disease.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
