Paper Review | Toward a Common Coordinate Framework for the Human Body

For quite some while now, I’ve carried a slight sensation of guilt within me. Guilt for being so caught up in my own projects and activities that I’ve come to neglect one of the most essential duties as a researcher: to survey, read, and contemplate upon new (and old) material that have been published. It’s so easy to put that interesting paper aside in order to finish a project deadline, as no immediate gains are observed from reading through the former in contrast to completing the latter. Once a habit, this behavior efficiently imprints the idea that “making” is more important than learning; this might give a false sense of efficiency at first, but (I believe) has a severe negative impact on the quality and creativity of one’s work in the long run. How am one supposed to grow and expand the knowledge sphere if one never travels beyond its borders? Hence, I’ve somehow managed to conjure a vision of myself being a more active reader gatherer of information this year by committing to the task of doing bi-weekly paper reviews, starting today. I should say here, that these reviews will mainly be focused on summarizing the material presented, rather than evaluating them; but of course, if I have opinions or comments which I deem relevant, these will be shared as well. My hope is to continue this at least throughout the year, to then evaluate and reassess whether its an endeavor worth continuing. Alas, let’s get started.

Paper Title : Toward a Common Coordinate Framework for the Human Body

Authors : Jennifer E. Rood, Tim Stuart, Shila Ghazanfar, Tommaso Biancalani, Eyal Fisher, Andrew Butler, Anna Hupalowska, Leslie Gaffney, William Mauck, Gökçen Eraslan, John C. Marioni, Aviv Regev, Rahul Satija

Published : 12-12-2019

doi : https://doi.org/10.1016/j.cell.2019.11.019

Review : In ambitious projects like the Human Cell Atlas we often rely on joint efforts from multiple labs, as well as collection of data across multiple individuals, ages, phenotypes, both sexes and plenty of other features. This strategy is very apt for producing large and diverse datasets; but unless we somehow harmonize and integrate all of the data into a common framework it will be nothing more than a big mess that is ought to bring more headaches than joy. The authors of this paper discuss these issues in terms of a common coordinate framework which they - very neatly - define as a reference map that can : “[..] assign a reproducible address to every location”. CCFs can be constructed at different scales (macro, meso, micro and fine are used as examples here), but its purpose always remains the same, to relate regions in independent samples in a shared context.

Personally, I think it’s easiest to motivate the need of a CCF through an analogy. Imagine you have a set of portraits, painted by different artists. Some paintings depict the same individual, but not all of them. An example of this scenario is illustrated below:

Now, if we wanted to assess how skilled each painter was at recreating certain facial features (e.g., eyes or noses), we must somehow identify these different parts in each of the pictures before we could speak of comparing them. However, it’s insufficient to only compare the images pixel-by-pixel as we have several different motifs. Somehow, we need to use the common features of a face, to identify the unique elements of each artist. Similarly, when building an atlas of the human body, organs or sub-region of an organ using data from multiple sources, we need to know where each piece fits into the larger picture. A CCF would further enable a very precise exploration of variation across individuals, and is obviously a fundamental tool in any quest to study heterogeneity.

In the paper they discuss some of the challenges associated with the creation of a CCF, and consider the inherent anatomical diversity across individuals as perhaps the most prominent one. They also bring up a good point about how consistent annotation (used to assemble the CCF) is not as straightforward of a task as it might seem, since biological compartments and structures do not always have well-defined boundaries. Another challenge in devising methods to construct CCFs arise from the fact that a sample’s character by and large dictates how appropriate a certain method is. To illustrate this, two extremes are given as examples: (1) when anatomical regions are highly similar across all samples (e.g., early stages of embryogenesis), and (2) when cells are (seemingly) randomly organized (for example in a tumor). Of course, these two extremes represent end-points on a continuum, and most samples place somewhere in-between the two. For samples more like (1) leveraging the consistency and a priori knowledge is preferable, while for samples like (2) one would instead do better by employing data-driven methods where the locations are learnt from the data. Of course, in both scenarios it is appealing to envision a form of Bayesian approach, where we update our beliefs regarding a sample’s location as more data is gathered.

Different variants of CCFs exist, three broad classes are given in the paper: using Anatomical Plane Coordinates, Landmark based construction and complex non-standard approaches. The first (anatomical plane coordinates) more or less aligns samples by registration to a reference (to account for inter-specimen variation). It uses a standard coordinate axis as the reference point to which distances are measured. One huge benefit is that distances between objects in this space, represent true distances. The landmark based approach is more or less an extension of the anatomical plane coordinates, but where one uses known landmarks (e.g., a vein bifurcation or a certain axon bundle) rather than anatomical planes. These landmarks allow one to anchor points in different samples to a reference. Having identified the landmarks, one may then apply a linear transformation to map a query image to the reference. For both the aforementioned strategies, the use of a reference is essential - the template itself may however be updated iteratively as the process progress. The complex non-standard approaches are used when there’s a lack of anatomical structure or when no clear landmarks can be identified. In complex approaches, local non-linear (in contrast to the other methods) warping is applied to account for large anatomical variation.

Now of course, as you might have noted, both of the two first methods relies on a reference template. Hence the question of how such a template is obtained, to which the authors answers that it depends largely on the sample. For highly stereotypical samples, a successful approach has been to use an iterative process, starting with a seeding set to which the whole dataset is aligned, a new reference is then extracted and used as a seed in the next iteration; a procedure repeated until convergence. To overcome the inter-individual variability in samples with “similar inter-individual organ structure but differences in cell type location and organ dimensions” a certain degree of supervision was added to the procedure; by manual annotation of landmarks etc. Semi-supervised processes are also predicted to be of great use in the construction of a human atlas. For highly non-stereotypical samples, there were no good strategies for template construction at the date of their writing, and they mention how this will have to be remedied by new methodological developments.

Interestingly the authors highlight a fourth orthogonal, and fairly unconventional, approach that discards the use of a pre-defined coordinate system, landmarks and templates, in favor of learning it from the data itself. Multiple examples of methods for spatial reconstruction (see Seurat v3 and novoSpaRc) can be seen as testimonies to spatial information being contained within the expression profiles of cells. To me, this is very similar to the SLAM problem in robot localization, where one tries to create the map while also finding ones current position within it.

They also discuss how, once a CCF is established, samples of the same and different modalities may be mapped to it. If we are operating with data of the same modality, the process should be nothing but straightforward. We simply use the same strategy as when creating the CCF itself. For other cross-modality-mapping the case is slightly different, and integration will only be possible if the assumption that both modalities shares a latent representation (e.g., chromatin accessibility) holds.

Now, one idea that the authors introduce, and which according to them is novel (I can find nothing that would contradict the claim) is that given all these challenges and how much influence a sample’s character has on the choice of method, is to dismiss the objective of creating a single CCF. Instead, multiple CCFs should be created in a hierarchical (by scale) fashion, which would allow both horizontal and vertical movements across the atlas, but where each CCF is adapted to best fit the data. This would represent a sort of “atlas of atlases” as they express it. To me, this makes perfect sense, although it would have been interesting to see some more discussion of how the vertical movements (between scales) were envisioned.

Their conclusions are brief, and so shall mine be. They make a strong argument for the value of CCFs, and why they deserve our attention. Inter-individual variability is the biggest hurdle to tackle, but there are strategies for it, though highly dependent on the data; this is not a “one-size fits all” type of problem. When the data is complex and we can’t establish natural anatomical templates or find good landmarks, perhaps it’s better to let the data dictate its own reference, using approaches inspired from methods of spatial reconstruction. All in all, CCFs brings structure to a very chaotic and seemingly unstructured space - and they are imperative to the process of creating a unified atlas of the human body.