Metabolic activity organizes olfactory representations
Curation statements for this article:-
Curated by eLife
eLife assessment
This is an important study that asks why odors smell similar even though their chemical structures appear quite different. The authors use machine-learning to make a compelling case to map the odor-relatedness of compounds to their place in metabolic pathways and propose that this is a general feature of odor perception across the animal kingdom. The conclusions could be strengthened by considering published physiological data.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Neuroscience (eLife)
Abstract
Hearing and vision sensory systems are tuned to the natural statistics of acoustic and electromagnetic energy on earth and are evolved to be sensitive in ethologically relevant ranges. But what are the natural statistics of odors , and how do olfactory systems exploit them? Dissecting an accurate machine learning model (Lee et al., 2022) for human odor perception, we find a computable representation for odor at the molecular level that can predict the odor-evoked receptor, neural, and behavioral responses of nearly all terrestrial organisms studied in olfactory neuroscience. Using this olfactory representation (principal odor map [POM]), we find that odorous compounds with similar POM representations are more likely to co-occur within a substance and be metabolically closely related; metabolic reaction sequences (Caspi et al., 2014) also follow smooth paths in POM despite large jumps in molecular structure. Just as the brain’s visual representations have evolved around the natural statistics of light and shapes, the natural statistics of metabolism appear to shape the brain’s representation of the olfactory world.
Article activity feed
-
-
Author Response
Reviewer #2 (Public Review):
In this manuscript, the authors use an embedding of human olfactory perceptual data within a graph neural network (which they term principal odor map, or POM). This embedding is a better predictor of a diverse set of olfactory neural and behavior data than methods that use chemical features as a starting point to create embeddings. The embedding is also seen to be better for comparison of pairwise similarities (distances of various sorts) - the claim is that proximity of pairs of odors in the POM is predictive of their similarity in neural data from olfactory receptor neurons.
A major strength of the paper is the conceptualization of the problem. The authors have previously described a graph neural net (GNN) to predict verbal odor descriptors from molecular features (here, a 2019 …
Author Response
Reviewer #2 (Public Review):
In this manuscript, the authors use an embedding of human olfactory perceptual data within a graph neural network (which they term principal odor map, or POM). This embedding is a better predictor of a diverse set of olfactory neural and behavior data than methods that use chemical features as a starting point to create embeddings. The embedding is also seen to be better for comparison of pairwise similarities (distances of various sorts) - the claim is that proximity of pairs of odors in the POM is predictive of their similarity in neural data from olfactory receptor neurons.
A major strength of the paper is the conceptualization of the problem. The authors have previously described a graph neural net (GNN) to predict verbal odor descriptors from molecular features (here, a 2019 preprint is cited, but a newer related one in 2022 describing the POM is not cited). They now use the embedding created by that GNN to predict similarities in large and diverse datasets in olfactory neuroscience (which the authors have curated from published work). They show that predictions from POM are better than just generic chemical features. The authors also present an interesting hypothesis that the underlying latent structure discovered by the GNN relates to metabolic pathway proximity, which they claim accounts for the success in the prediction of a wide range of data (insect sensory neuron responses to human behavior). In addition to the creativity of the project, the technical aspects, are sound and thorough.
There are some questions about the ideas, and the size of the effects observed.
- The authors frame the manuscript by invoking an analogy to other senses, and how naturalstatistics affect what's represented (and how similarity is defined). However, in vision or audition, the part of the world that different animals "look at" can be very different (different wavelengths, different textures and spatial frequencies, etc). It is still unresolved why any given animal has the particular range of reception it has. Each animal is presumably adapted for its ecological niche, which can have different salient sensory features. In vision, different animals pick different sound bandwidths or EM spectra. Therefore, it is puzzling to think that all animals will somehow treat chemicals the same way.
Our assumption (an assumption of the broader interpretation, not of the analyses themselves) that all terrestrial animals have a correlated odor environment is certainly only true for some values of “correlated”. One could imagine, for example, that some animals are able to exploit food energy sources that humans cannot (for example, plants with high cellulose content), and that they might therefore be adapted to smell metabolic signatures of such plants, whereas humans would not be so adapted. This seems quite reasonable and there are probably many such examples. In future work they might be used to test the theory directly: representations might be more likely to differ across species on tasks when the relevant ecological niches are non-overlapping. We have updated the discussion to propose such future tests. However, it is also apparent that the odor environment overall is nonetheless highly correlated across species. Recent work (Mayhew et al, PNAS) showed that nearly all molecules that pass simple mass transport requirements (that should apply to all mammals, at the least) are likely to have an odor to humans, so it seems unlikely that the “olfactory blind spots” are intrinsically large.
- The performance index could be made clearer, and perhaps raw numbers shown beforeshowing the differences from the benchmark (Mordred molecular descriptor). For example, can we get a sense of how much variance in the data does it explain, what percent of the hold-out tests does it fit well, etc.?
The performance index in Figure 1 is required to compare across different types of tasks, which are in turn dictated by the nature of the data (e.g. continuous vs categorical). Regression tasks yields an R2 value and categorical tasks yield an AUROC. We normalized and placed these on a single scale in order to show all of the tasks clearly together. We have added a table to the shared code (from link in Methods section, go to predictive_performance/data/dataset_performance_index_raw.csv) that shows the original (non-normalized) values, for both the POM and the benchmark(s) across multiple seeds and various metrics with the model hyper-parameters that generate the best performance.
- The "fitting" and predictions are in line with how ML is used for classification and regression inlots of applications. The end result is a better fit (prediction), but it's not actually clear whether there are any fundamental regularities or orders identified. The metabolic angle is very intriguing, but it looks like Mordred descriptor does a very good job as well (extended figure 5 [now Figure 2-figure supplement 5]). Is it possible to show the relation between metabolic distance and Mordred distance in Figure 2c? In fact, even there, cFP distance looks very well correlated with metabolic distance (we are talking about r= 0.9 vs r = 0.8). This could simply be due to a slightly nonlinear mapping between chemical similarity and perceptual similarity (which was used to get POM distance).
We show additional “showdown” comparisons between metabolic distance, POM distance, and alternative distance metrics in the new Figure 2-figure supplement 3 and Figure 2-figure supplement 4. Indeed, the Mordred descriptors perform well; after all, metabolic reactants and products must be at least somewhat structurally related. But POM (derived only from human perceptual data) outperforms it significantly. Visual inspection of Figure 2c also reveals that the dispersion of structural distances (at a given metabolic distance) is just much higher than the dispersion of POM distances. This won’t change if one uses a non-linear curve fit, as it is a property of the data itself.
It’s also worth noting while r=0.8 and r=0.9 might seem close, in terms of variance unexplained (1 - r2) they are approximately two-fold different. Reducing the unexplained variance by half seems like a meaningful difference. Alternatively, if one simulates scatter plots with correlation r=0.8 vs r=0.9, it is apparent that the latter is simply a much tighter relationship.
- How frequent are such examples shown in Fig 2d? Pentenal and pentenol are actually verysimilar in many ways, and it may be that Tanimoto distance is not a great descriptor of chemical similarity. cFP edit distance is quite small, just like metabolic distance. The thiol example on the right is much better. Also, even in Fig 2C POM vs metabolic distance, the lowest metabolic distances have large variations in the POM values - so there too, metabolic reactions that create very different molecules in 1 step can vary widely in POM distance as well.
We agree that Tanimoto distance is not perfect. We were unable to find a measure of structural distance that agreed with human intuitions about “structural distance” in all cases; indeed that intuition is often generated by an understanding of odor/flavor characteristics of function in metabolic networks, which would beg the question! To answer the question about the frequency of examples like the ones shown in Figure 2d, we created a new density map (Figure 2-figure supplement 4) showing the number of one-step metabolite pairs for a given range of POM vs cFP edit/Tanimoto distance. We found >25 pairs of metabolites in the same “small POM distance” and “large structural distance” quadrant from which we found the original examples shown in Figure 2d..
- A major worry is that Mordred descriptors are doing fine, and POM offers only a smallimprovement (but statistically significant of course). Another way to ask this question is this: if you plot pairwise correlation/distance of pairs of odors from POM against that for Mordred, how correlated does this look? My suspicion is that it will be highly correlated.
It will look highly correlated (as shown in the new Figure 2-figure supplement 3). The reason is that metabolic reactions cannot make arbitrary transformations to molecules (the reactants must have some structural relationship to the products) or similarly that olfactory receptors (in any species) cannot have arbitrary tuning – at the end of the day receptors mostly bind to similar-looking classes of molecules. As stated above, we believe that the improvement here is not just statistically significant but meaningful – a 2-fold drop in unexplained variance is large – and that it is important to identify principles by which the nervous system can be tuned, above and beyond the physical constraints imposed by basic rules of chemistry.
Also, the metabolic distances that we constructed from available data are themselves noisy, since not all metabolic pathways and the compounds that compose them are known, which places an upper bound on the correlation that we could have obtained. Despite that, we still found a correlation of r>0.9.
- The co-occurrence in mixtures and close POM distance may arise from the way theembedding was done - with perceptual descriptors used as a key variable. Humans may just classify molecules that occur in a mixture as similar just from experiencing them together. Can the authors show that these same molecules in Fig 4d,e have very similar representations in neural data from insects or mice?
We have added a new Figure 4-figure supplement 1 to show this. One constraint is that the neural datasets must contain molecules that are also in the natural substance datasets used in Figure 4. In all cases where the data is sufficient to be powered to test the hypothesis (i.e. more than five co-occuring pairs of molecules in essential oil), we observe an effect in the predicted direction.
-
eLife assessment
This is an important study that asks why odors smell similar even though their chemical structures appear quite different. The authors use machine-learning to make a compelling case to map the odor-relatedness of compounds to their place in metabolic pathways and propose that this is a general feature of odor perception across the animal kingdom. The conclusions could be strengthened by considering published physiological data.
-
Reviewer #1 (Public Review):
This study builds an odorant organization map as estimated by a neural network trained on several odor perceptual classification databases. The authors come up with an attractive hypothesis about the link of odor perception to metabolic connectedness, as opposed to a range of other ways of classifying odorant compounds. There are several interesting implications of this, which the authors touch upon, but could perhaps frame as specific predictions.
The authors clearly have generated a powerful methodology, a useful classifying network, and a well-organized database. The study would be much stronger if the methodology were more thoroughly explained, with open code and data availability as expected for a computational study, and as a resource for further research on the topic.
It would also be valuable to …
Reviewer #1 (Public Review):
This study builds an odorant organization map as estimated by a neural network trained on several odor perceptual classification databases. The authors come up with an attractive hypothesis about the link of odor perception to metabolic connectedness, as opposed to a range of other ways of classifying odorant compounds. There are several interesting implications of this, which the authors touch upon, but could perhaps frame as specific predictions.
The authors clearly have generated a powerful methodology, a useful classifying network, and a well-organized database. The study would be much stronger if the methodology were more thoroughly explained, with open code and data availability as expected for a computational study, and as a resource for further research on the topic.
It would also be valuable to place the current findings in the context of considerable earlier work that has sought to map odor perception and place it in the context of structural and chemical features.
-
Reviewer #2 (Public Review):
In this manuscript, the authors use an embedding of human olfactory perceptual data within a graph neural network (which they term principal odor map, or POM). This embedding is a better predictor of a diverse set of olfactory neural and behavior data than methods that use chemical features as a starting point to create embeddings. The embedding is also seen to be better for comparison of pairwise similarities (distances of various sorts) - the claim is that proximity of pairs of odors in the POM is predictive of their similarity in neural data from olfactory receptor neurons.
A major strength of the paper is the conceptualization of the problem. The authors have previously described a graph neural net (GNN) to predict verbal odor descriptors from molecular features (here, a 2019 preprint is cited, but a …
Reviewer #2 (Public Review):
In this manuscript, the authors use an embedding of human olfactory perceptual data within a graph neural network (which they term principal odor map, or POM). This embedding is a better predictor of a diverse set of olfactory neural and behavior data than methods that use chemical features as a starting point to create embeddings. The embedding is also seen to be better for comparison of pairwise similarities (distances of various sorts) - the claim is that proximity of pairs of odors in the POM is predictive of their similarity in neural data from olfactory receptor neurons.
A major strength of the paper is the conceptualization of the problem. The authors have previously described a graph neural net (GNN) to predict verbal odor descriptors from molecular features (here, a 2019 preprint is cited, but a newer related one in 2022 describing the POM is not cited). They now use the embedding created by that GNN to predict similarities in large and diverse datasets in olfactory neuroscience (which the authors have curated from published work). They show that predictions from POM are better than just generic chemical features. The authors also present an interesting hypothesis that the underlying latent structure discovered by the GNN relates to metabolic pathway proximity, which they claim accounts for the success in the prediction of a wide range of data (insect sensory neuron responses to human behavior). In addition to the creativity of the project, the technical aspects, are sound and thorough.
There are some questions about the ideas, and the size of the effects observed.
1. The authors frame the manuscript by invoking an analogy to other senses, and how natural statistics affect what's represented (and how similarity is defined). However, in vision or audition, the part of the world that different animals "look at" can be very different (different wavelengths, different textures and spatial frequencies, etc). It is still unresolved why any given animal has the particular range of reception it has. Each animal is presumably adapted for its ecological niche, which can have different salient sensory features. In vision, different animals pick different sound bandwidths or EM spectra. Therefore, it is puzzling to think that all animals will somehow treat chemicals the same way.
2. The performance index could be made clearer, and perhaps raw numbers shown before showing the differences from the benchmark (Mordred molecular descriptor). For example, can we get a sense of how much variance in the data does it explain, what percent of the hold-out tests does it fit well, etc.?
3. The "fitting" and predictions are in line with how ML is used for classification and regression in lots of applications. The end result is a better fit (prediction), but it's not actually clear whether there are any fundamental regularities or orders identified. The metabolic angle is very intriguing, but it looks like Mordred descriptor does a very good job as well (extended figure 5). Is it possible to show the relation between metabolic distance and Mordred distance in Figure 2c? In fact, even there, cFP distance looks very well correlated with metabolic distance (we are talking about r= 0.9 vs r = 0.8). This could simply be due to a slightly nonlinear mapping between chemical similarity and perceptual similarity (which was used to get POM distance).
4. How frequent are such examples shown in Fig 2d? Pentenal and pentenol are actually very similar in many ways, and it may be that Tanimoto distance is not a great descriptor of chemical similarity. cFFP edit distance is quite small, just like metabolic distance. The thiol example on the right is much better. Also, even in Fig 2C POM vs metabolic distance, the lowest metabolic distances have large variations in the POM values - so there too, metabolic reactions that create very different molecules in 1 step can vary widely in POM distance as well.
5. A major worry is that Mordred descriptors are doing fine, and POM offers only a small improvement (but statistically significant of course). Another way to ask this question is this: if you plot pairwise correlation/distance of pairs of odors from POM against that for Mordred, how correlated does this look? My suspicion is that it will be highly correlated.
6. The co-occurrence in mixtures and close POM distance may arise from the way the embedding was done - with perceptual descriptors used as a key variable. Humans may just classify molecules that occur in a mixture as similar just from experiencing them together. Can the authors show that these same molecules in Fig 4d,e have very similar representations in neural data from insects or mice?
-
-
-