Machine learning of honey bee olfactory behavior identifies repellent odorants in free-flying bees in the field
- Department of Molecular, Cell and Systems Biology, University of California, Riverside, United States
- Department of Entomology, Centre for Integrative Bee Research (CIBER), University of California, Riverside, United States
- Interdepartmental Neuroscience Program, University of California, Riverside, United States
Figures
Figure 1
A chemical informatics method to predict honey bee repellence.
Overview of the steps for the machine learning-based cheminformatics pipeline used to model honey bee repellence from the 3D structure of a training set of known repellents, and subsequent in silico screening for novel repellents from a large chemical space.
Figure 2
Predicted honey bee repellents.
(A) Testing chamber containing a one-choice trap to determine whether an odorant will repel fruit flies, and the mean percentage of fruit flies caught in a trap treated with predicted repellent odorants (10% vol/vol in paraffin oil) and baited with 10% apple cider vinegar. N=5–10 trials (~20 flies/trial). Error bars = s.e.m. *p<0.05, **p<0.005, ***p<0.001, ****p<0.0001. (B) Photograph of the two-choice Petri dish arenas used to test the behavior of honey bees to repellent candidates individually. (C) Table with preference indexes in the two-choice assay for the first choices of honey bee workers offered two containers of honey water (50 μl of 50%) placed on filter paper disks treated with a test chemical or treated with solvent alone. The index was calculated for each compound as (total number of repellent choices minus total number of solvent choices) divided by the sum of all tests. We used McNemar tests for pairwise comparisons.
-
Figure 2—source data 1
Data for D. melanogaster trap assays.
- https://cdn.elifesciences.org/articles/104831/elife-104831-fig2-data1-v1.xlsx
Figure 3 with 1 supplement
Reiterative training of machine learning (ML) predictive models and testing predicted honey bee repellents.
(A) Schematic pipeline for the overall computational-behavior hybrid pipeline showing the two rounds. (B) Mean preference index of honey bees in making the first choice to move to the repellent treated side in a two-choice plate assay. 70 μl of pure honey was used in the two containers to attract the bees. (C) Average consumption of honey from containers placed on each side in the two-choice plate assays (N=460 plates). The index is calculated for each compound as which honey pot has greater consumption (total number of repellent side with greater consumption minus total number of solvent side with greater consumption) divided by the sum of all choices. (D) Mean percentage of fruit flies caught in a trap treated with predicted repellent odorants at application rates to test in field tests (0.1 mg/cm2) and baited with 10% apple cider vinegar. N=10–14 trials (~20 flies/trial). Error bars = s.e.m.
-
Figure 3—source data 1
Fly trap entry data.
- https://cdn.elifesciences.org/articles/104831/elife-104831-fig3-data1-v1.xlsx
Figure 3—figure supplement 1
Average percentage of two-choice plate assays in which honey bees made their first choice to either move to the repellent treated side or solvent treated side across different honey bee colonies, N=3-6 colonies (~6 plates/colony).
Error bars = 1 s.e.m., *p<0.05, **p<0.001.
Figure 4
Field testing of top repellents using robbing assays.
(A) Representative photo of a robbing assay. (B) Time course of the mean numbers of foraging honey bees visiting for each indicated compound tested (0.1 mg/cm2). DEET was used as a positive control and acetone as a negative control in each assay. N=6–18. Error bars = s.e.m. (C) Mean number of honey bees for each compound summed over the time points shown in B. *p<0.05, **p<0.005, ***p<0.001. We used Kruskal-Wallis tests for pairwise comparisons. (D) Mean number of honey bees for each compound tested at the regular dose (0.1 mg/cm2) and half dose (0.05 mg/cm2), summed over the time points over the duration of the experiment as above. N=4, error bars = s.e.m.
-
Figure 4—source data 1
Data on honey bee robbing assays.
- https://cdn.elifesciences.org/articles/104831/elife-104831-fig4-data1-v1.xlsx
Author response image 1
Author response image 2
Additional files
-
Supplementary file 1
Optimized DRAGON physicochemical descriptor set from iterative training in Figure 3A used to predict bee repellent compounds.
- https://cdn.elifesciences.org/articles/104831/elife-104831-supp1-v1.docx
-
Supplementary file 2
Repellency testing in the two-choice plate assay from the second iteratively trained round of predictions.
- https://cdn.elifesciences.org/articles/104831/elife-104831-supp2-v1.docx
-
Supplementary file 3
Training chemicals used for the ML.
(a) Initial data for Round 1 machine learning. (b) Data added for Round 2 after testing for bee repellency.
- https://cdn.elifesciences.org/articles/104831/elife-104831-supp3-v1.docx
-
MDAR checklist
- https://cdn.elifesciences.org/articles/104831/elife-104831-mdarchecklist1-v1.docx
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Machine learning of honey bee olfactory behavior identifies repellent odorants in free-flying bees in the field
eLife 14:RP104831.
https://doi.org/10.7554/eLife.104831.3
