Low-dimensional olfactory signatures of fruit ripening and fermentation
Curation statements for this article:-
Curated by eLife
eLife Assessment
This important study presents results for the theory of odor coding in hyperbolic spaces by revealing spiral trajectories in the dynamics of odors during natural, ethologically relevant processes such as ripening. In the current manuscript, the strength of the evidence is solid and would be strengthened by answering several technical points raised by reviewers.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Odors provide an important communication channel between plants and animals. Fruits, vital nutrient sources for animals, emit a complex array of monomolecular volatiles. Animals can use the structure of these mixtures to assess properties of fruit predictive of their nutritive and reproductive value. We analyzed the statistics of fruit odor mixtures sampled across stages of ripening and fermentation to find that they fall on a low-dimensional hyperbolic map. Hyperbolic maps, with their negative curvature and an exponentially expanding state options, are adept at describing hierarchical relationships in the data such as those arising from metabolic processes within fruits. In the hyperbolic map, samples followed a striking spiral trajectory. The spiral initiated near the map’s core, representing the under-ripe phase with specific profiles of monomolecular volatiles. Progressively mapping along the unfolding spiral trajectory were scent mixtures corresponding to ripening, and then rotting or fermentation. The unfolding process depended on the specific fermentation processes that dominated in the samples, determined largely by the microbes (e.g. bacteria or yeast) present in the sample. These results generalized across fruit types and describe trajectories in the natural odorant space with significant behavioral relevance for insects.
Article activity feed
-
eLife Assessment
This important study presents results for the theory of odor coding in hyperbolic spaces by revealing spiral trajectories in the dynamics of odors during natural, ethologically relevant processes such as ripening. In the current manuscript, the strength of the evidence is solid and would be strengthened by answering several technical points raised by reviewers.
-
Reviewer #1 (Public review):
Summary:
This work represents a new development in the theory of odor coding and recognition, based on mapping odor mixtures in low-dimensional hyperbolic spaces. The authors describe the dynamics of odor mapping, across stages of ripening and fermentation (trajectories in odor space), which, surprisingly, generalize across fruit types.
Strengths:
The approach provides a remarkably concise and clear description of the odor dynamics. As a model, the approach is mathematically exhaustive and generalizable. The analyses are technically correct and statistically robust.
Weaknesses:
None.
-
Reviewer #2 (Public review):
This article presents an analysis of the chemical composition of head-space generated by fruit at differing stages of ripeness. The authors used gas chromatography-mass spectrometry (GC-MS) to record the chemical makeup of the respective head-space samples. The authors process the data and present it in a low dimensional space. They then draw conclusions from the geometry of that representation about the process of fermentation.
I have a number of major concerns with some of the stages in the argument advanced by the authors:
(1) As far as I understand, the authors restrict their analysis to 13 molecules which appear in samples of all three levels of ripeness. This choice causes the analysis to overlook the very likely (and meaningful) possibility that different molecules present at different levels of …
Reviewer #2 (Public review):
This article presents an analysis of the chemical composition of head-space generated by fruit at differing stages of ripeness. The authors used gas chromatography-mass spectrometry (GC-MS) to record the chemical makeup of the respective head-space samples. The authors process the data and present it in a low dimensional space. They then draw conclusions from the geometry of that representation about the process of fermentation.
I have a number of major concerns with some of the stages in the argument advanced by the authors:
(1) As far as I understand, the authors restrict their analysis to 13 molecules which appear in samples of all three levels of ripeness. This choice causes the analysis to overlook the very likely (and meaningful) possibility that different molecules present at different levels of ripeness are informative and might support different results.
(2) It is unclear what was used as control? Empty bag? Please include the control results in your supplementary table, or indicate in the text if you eliminated compounds that were found in the control.
(3) It is not clear that Figure 2-H _looks_ like a spiral. The authors should provide a quantifiable measure of the quality of the fit of a spiral rather than other paths. Furthermore, in the section "collective spiral ..." the end of paragraph one, "the points were best fitted by a two parameter archemedian spiral" best out of what? best out of all two parameter spirals? Please explain
(4) In the section "estimating odor source phenotype ... " the authors write: "we first calculated the association of odorant compounds with different phenotypes in this dataset" how was that done?
(5) Even if hyperbolic space MDS is slightly better, an R^2 value for Euclidean MDS of 0.797 is very good and one could say that Euclidean MDS is also an option.
(6) In the section "collective spiral ..." near end of paragraph two: " we removed outlier samples for days 10 and 17 for two reasons...". Why does a smaller number of samples should make a certain day an outlier.
(7) In section titles "collective spiral progression of multiple..." the authors write: the hyperbolic t-sne embedding exhibited batch effects across runs that amounted to rotation of the data. To compensate for these effects and combine data across runs we performed Procrustes analysis to align data across runs".
Can we be sure that this process does itself not manufacture an alignment of data? The authors should apply the same process to random or shuffled data and see if the result is different from the actual data.
-
-
-
