In this talk, Aviv Regev outlines the vision and rationale for the Human Cell Atlas (HCA), describing how a comprehensive molecular map of every human cell can transform biology and medicine. She frames cells as the fundamental units of the body, analogous to elements in the periodic table, and argues that recent single‑cell and spatial profiling technologies now make it possible to capture the multiple facets of cell identity—type, state, lineage, location, and interactions—in a unified molecular coordinate system. Regev walks through four major analytical pillars of the HCA: (1) a data‑driven taxonomy of cell types, emphasizing the need for precise, scalable classification methods and power analyses; (2) spatial histology, leveraging in‑situ molecular signatures to map cells within tissue architecture and requiring spatial power analysis and cross‑sample alignment; (3) developmental and physiological dynamics, using pseudo‑time and trajectory inference to reconstruct lineage, branching, and state transitions, while disentangling biological variation from technical noise; and (4) integration of multi‑omic, cross‑species, and perturbation data to uncover regulatory codes and enable comparative biology. Throughout, she highlights computational challenges—defining distance metrics, scaling algorithms to millions of cells, handling technical confounders, and building iterative statistical frameworks—that must be solved before, during, and after data collection. Regev also stresses the broader impact of the HCA on disease research, drug discovery, diagnostics, and regenerative medicine, illustrating how cell‑type‑specific expression can link GWAS hits to pathogenic cells, improve toxicity profiling, and enable precise targeting of disease‑relevant states. Finally, she outlines the collaborative, open‑access, and equity‑focused principles guiding the HCA effort, the need for diverse expertise, and the iterative, community‑driven approach required to build and maintain such a resource.
Chapters
The base unit of our body is the cell, and in that sense, cells are akin to other elements that we know about in science in general and in biology. You can think of them similar to the elements in a periodic table or to genes in the genome. And ever since cells were originally discovered, biologists have always been on an endeavor to discover cells and characterize them and identify their unique characteristics. And this was an effort or an endeavor that was always fueled by the available technologies. In this process, we learned that there are many different facets to a cell's identity, including their types, their states, their history of transitions and lineage, where they're located in the body and what other cells they interact with. And it was— apparently that needs to be fixed. And it was increasingly clear that molecules have to underlie all of these different facets. And if this is the case, then we would expect that genomic profiles could maybe form a unifying set of coordinates. In the last few years, because of, again, technical advances in our ability to profile individual cells, both as they are dissociated from tissue and increasingly in the context of tissue, we can start thinking about putting all of this together in order to build a reference map of the cells of the human body. And in doing so, what I want to take a few minutes at the beginning today is to think about the types of biology that we could learn from having this kind of reference map. What are the computational challenges that each of these levels would introduce, which will be the topic of discussion for the next two days, and also why we do it in terms of the impact that we would want to have not just on fundamental basic understanding of biology, but also on human health. And I'm going to do this going through examples one by one in each of these facets: the taxonomy, histology, developmental biology and physiology, which form some of the basic pillars of our understanding of multicellular organisms like humans. And so I'm going to start with taxonomy, or with the question of cell types. One of the things that we hope to get out of a human cell atlas is a taxonomy of cells based on their molecular characteristics and the ability to harmonize this taxonomy with other ways in which we characterize cells, like their morphology or their location in the body. And these are two recent examples of work looking, for example, at the retina, at neurons in the retina, or at cells in the blood, and in fact discovering, using single-cell profiling in this case, new cell types and validating them to be new cell types by other ways of characterizing cells, either in the nervous system or in the immune system. What are some of the big challenges that as a computational community we need to take on in order to characterize cell types correctly? Well, we can think of a cell type maybe as a region or a probability distribution in some complex multidimensional space. But as we start thinking about this, we realize that while a cell type is a very intuitively compelling concept, it is often quite difficult to pin down this concept, especially as we go further and further down to finer and finer distinction. And one of the hopes is that as a computational community, we will come up with data-driven ways in which to make this concept of cell type much more precise and well-defined. There are also a lot of challenges that are pragmatic about how to actually go about it. For example, can we come up with the right power analyses in order to determine how many cells we actually need to profile in order to collect cells and determine their identity that are present at a certain level of rarity? There are many technical considerations. We can classify cells by their profiles, but we also have to be very mindful of technical confounders. And we want to make sure that what we identify is biologically coherent and not a technical artifact that is confounding the same properties. We have to think about how we define the distance, a distance metric between cells based on their profiles. And as this example shows, taking an analysis that might work very well for one system— in this case, worked quite well for one type of neuron— and applying it for another and more complex class of neuron no longer cuts it. And you need to revise the distance metric in order to get a classification that actually works. Of course, you also need to be able to scale sufficiently. If there will be millions and tens of millions and possibly billions of cells in a cell atlas, many of the algorithms that we use today will simply break because of scalability considerations. We have to make sure that we don't over-categorize. And to know that we categorize correctly, We need to make our definitions coherent with those of other classifications and taxonomies and to have a notion of what it means to actually discover a cell type and validate it. Now, the payoff of doing cell types correctly is going to be enormous. Besides a basic understanding of the cells from which we're made, this will allow us to understand better the relationship between genes and disease. For example, a geneticist studying a genome-wide association study will be able to identify the cells in a complex tissue in which the gene is actually expressed and presumably acts, as I'm showing here for cell-type-specific expression in the retina in the context of genes from a genome-wide association study of acute macular degeneration, or for genes in the gut for GWAS of IBD, inflammatory bowel disease. A drug developer trying to understand inadvertent toxicities we'd have a better handle on other cell types, including incredibly rare ones, when the gene being targeted might be expressed. And we can even imagine a new level of diagnostics, where we take something like the complete blood count, which in fact is a single-cell assay, but a rather simple one, and replace it with CBC 2.0, based not just on some crude characterizations of the types of cells, but a very fine characterization going all the way to the experiences that those cells have had in the body. Moving on to the next layer, we can start thinking about where cells actually are, and not just their types, in the context of histology. And here we have a great opportunity to leverage the fact that expression profiles, that the molecules that cells use, actually carry in them the imprint of the local neighborhood of the cell, as well as some much more global coordinates in which the cell resides in the body, all the way to anatomic position. This is also a remarkable technical opportunity, because it is exactly at the sweet spot of combining two technologies that our community has in hand: the ability to profile dissociated cells and the ability to measure signatures of molecules in situ. And by taking both of these together, we can actually map cellular profiles into their position. This is, of course, a computational problem, and there are multiple challenges that it introduces. If we think about canonical tissue, we might want to take a series of specimens, look at profiles of cells in one context, and then find the right and most informative signatures to actually measure in situ. This leads to the concept of spatial power analysis, the ability to detect spatial patterns for which we need to be sufficiently powered. It also introduces the challenge of mapping many samples into a common coordinate framework, which is particularly difficult to humans that do not all have the same genetic background and are isolated at exactly the same age from the same animal facility. When we move to non-canonical tissue, be it lymph nodes or tumors, this challenge is even amplified. We don't expect the specimens all to be the same. But could there be a generative model that would take us away from precise Cartesian coordinates into some higher-order feature space of the features that are actually is actually reproducible across these many samples. And again, the payoff would be huge, because for example, we can imagine moving from the 150-year-old technology of H&E stains into a fully molecular view of histology and pathology in situ. And in the context of disease mechanism and drug discovery, the enormous progress that was made, for example, in immunotherapy shows how important it is to understand the identity of cells, not just in the context of who they are, but also in the context of who their neighbors are and how they talk to them. Now, cells are not static entities. They're dynamic, and they have histories, and in particular, they have a developmental history. And one of the remarkable successes of this community already has been to show that we can take a snapshot of right now And within this snapshot, we can pseudo-order cells based on the different points that they assume in a trajectory. Each of them is only in one point in time, but they are part of a continuum. In this way, we're actually leveraging asynchrony. Biologists have always fought against heterogeneity and asynchrony, but single-cell and in situ analysis allow us to leverage it. And this leads to examples from linear processes like neurogenesis in the adult brain, through looking at branching processes like hematopoiesis, and even taking a much finer look at programming and reprogramming of stem cells. And again, there are big computational challenges that still remain for our community to solve. We can think of these dynamic processes as continuous paths in space. There are practical questions. How many cells do we need to sample in order to get a good idea of every step in this? When some processes are really fast, the cells might be really rare. Could we fill in those blanks by some assumptions on how a continuous process might look like? We need to infer branching and lateral movement, cell divisions and cell death and migration events, all of which would require more sophisticated models. We might have to understand multiple paths that the same cell is taking at any given moment in time. We want to understand cell fate maps. We want to take a look at the cell right now and say, what might be its potential? What other cells could it become if sufficient time passes? If we had the ability to also look at genetic information, which is not outside of reach. We could even build cell lineage maps that look at cells today and tell us where they came from. And again, this has opportunities for real impact on disease areas that are very dynamic. For example, in regenerative medicine, it could allow us to identify cells and tissues that have regenerative potential and find molecular mechanisms that we can manipulate in order to increase regeneration, for example, after injury. In the context of making new cells, it could give us better codes for reprogramming and regeneration. In diseases like cancer that actually hijack the normal developmental hierarchies, as I'm showing here, for example, in glioma, it could give us a way to identify cancer stem cells and target them precisely. There are additional dynamical processes that cells go through, all those physiological processes from vacillations like the cell cycle to one-time directed responses, like the response to pathogen or nutrition. We need to be able to distinguish and order and characterize these processes as well. And again, here we're leveraging asynchrony instead of fighting against it. But there are challenges. There are interactions between multiple simultaneous processes. At any given moment in time, the cell might sense the food that it has, the pathogens that are around it, and where it is in the cell cycle, And these processes are partly independent from each other, but not wholly independent from each other. Cells can also assume multiple stable states across a spectrum, as is the case, for example, for T helper cells and the extent of pathogenicity that they might have for disease. But we have to be able to distinguish between variation that is biological, meaningful, and functional, and the variation that we just expect due to either biological stochasticity, or the noise in our measurement technologies, and that's a very difficult distinction to make. But if we can, then we can have substantial impact. By looking at the states of cells, we have a great opportunity of zeroing in on disease mechanism and doing it very precisely, for example targeting only the pathogenic cells while sparing the healthy ones. We also give great tools then to the drug discoverers, We give them signatures that they can screen against. We give them an identification of how drug resistance might arise and how they can fight it by changing the states of those cells. Now, all of these different facets combine together, and one of our challenges is never to forget that. We can't just look at one process and forget about all the others. The cell has a type, and it is located somewhere, And it is undergoing multiple transitions at any moment in time in which we look at it. And I would venture to say that we're only starting to parse out how to think about each of these separately and about how they come together. Now, the atlas will tell us a lot of things because of what is in it, but it will also empower us to take the same paradigms, the same experimental technologies and the same computational methods, and apply them much more broadly across human biology and beyond. Quite clearly, this applies to disease biology. The same that I have told you and showed you in example could be applied to any disease context. In any disease, there is a taxonomy of cell types in the ecosystem of the disease. It is organized into histological patterns. There might be developmental and physiological processes and cell-cell interactions. The Human Cell Atlas will not take on all diseases. But the tools and the infrastructures and the concepts and the computational algorithms and software that it will build will be applicable across all diseases. It will also give us an amazing starting point to go after molecular mechanisms in the human body. We can find the regulatory codes that actually control this phenomenal diversity and potential that cells have. We can do it by looking at covariation across cells. When cells are temporally ordered, we are even more powered to do this. When we have multiomics data that includes both mechanistic and phenotypic information on the same cell— whoops— on the same cell, things are even better for us. And if we have the ability to perturb the system genetically, as we can do in model organisms, then all the better. Again, the Human Cell Atlas will not do all of those things. But it is the starting point for all of them. And then finally, in biology, we can never only study the human. There are things we cannot do in people, but that we can do in model organisms, including many of those functional validations and studies that will follow on the Human Cell Atlas. By having a Human Cell Atlas and the ability to compare it to other model organisms, we can tell how good our models are, where they need improvement, and often get rid of a lot of confounders that have made interpretation extraordinarily difficult and also introduced a lot of ongoing arguments in the community that have not been resolved. This will come with a set of computational challenges. We will have to rethink comparative genomics and comparative biology and questions like alignment when we try to match cell types and states and continuous processes across organisms. Mechanisms in this case. So the Cell Atlas will enable a lot. I'll close in the last minute with a few early thoughts on how we will actually put it together. Like in many, many problems in life, when you actually don't know the answer ahead of time, you're going to have to iterate. You're going to have to take a guess and combine different methods because none of them is perfect and have both a both computational algorithms to analyze the data and statistical frameworks for power analysis in order to empower and make decisions throughout every step in the way. And so computation for the Human Cell Atlas is not just about what happens after the data has been collected. It is at every point, and in particular before the data has been collected and the experiment has been design. We also need to bring together diverse communities— biologists, technologists, data experts, and software engineers, and analysts and computer scientists. We have to do it in a mindful way so that as a community we come together, and we have an organizing committee for that, but also working groups of experts like the analysis working group led by Dana Pare and John Marioni that actually put this meeting together. And we need to have the right principles in place, and we'll never fail to remind ourselves of them. There has to be equity in the Cell Atlas. It has to be open access. Everything from methods to computational algorithms to data has to be accessible to everyone. It is an international collaboration across the scientific community, open to any and all who wish to participate. But we do adhere to strong standards for quality and reproducibility while maintaining flexibility in the face of a very rapidly changing technology and computational landscape. And with this, I will close. Thank you. Happy to take any questions and also introduce the next speaker. Thank you.
