Scientist Stories: Sarah Teichmann, Global Principles of Protein Interactions and Gene Expression

Good. Well, thank you very much, Sten, and thank you, Aviv, for making this happen, and for everybody turning out in such huge numbers. It's a really fantastic community, and I'm extremely excited about this third HCA meeting dedicated to computation, because computation is really at the heart of the Human Cell Atlas. And I'll talk a little bit about why during my talk.

It's at the heart of the human cells at a number of levels.

So it's super exciting for someone like myself with a computational background. So why, how did we come about, you know, how did we get this idea of starting a consortium or sort of rallying the community around the genomics of individual cells?

Well, the advent of whole genome methods to study individual cells is tremendously exciting because, of course, cells are the basic units of life, and they're the units of gene regulation. And so in the sort of hierarchy of an organism, the human body, the cells are really the sort of atomic units. The ability to comprehensively map them using, you know, quantitative genome-wide methods means that we can really, you know, start to be ambitious and attempt to fulfill this goal of creating a comprehensive reference map of human cells. And, you know, this really could potentially have huge implications, not only for basic biology, but also for many translational applications. So, Viv already alluded to these different levels of insights that we'll gain. So on the one hand, there are the cells themselves, but there's also their context within tissues, their dynamics, the difference forming a reference for global changes in disease, and cell-cell communication.

And really what's so powerful and attractive about the concept of a Human Cell Atlas, I think, is that we can now attempt to derive data-driven definitions for these fundamental concepts. And so why do this with an atlas? Why not do this in our own labs individually and start to answer these questions with our own favorite pet systems?

The reason is that in order to achieve this for an entire organism in a comprehensive sense, it's really a mammoth task.

Coordinating between ourselves, coordinating across countries completely makes sense.

And on top of that, having an integrated single data resource will allow us to gain insights that we can't from individual studies. There'll be meta-analyses across tissues and donors. It will—

having the data portal where everything is collected will be an invaluable reference for the community for the future.

And both at the medic— and that will have also medical and translational impact.

As Aviv said, we—

there are four pillars that have evolved, essentially the biological networks, technology development, and what we're here to discuss specifically is the data platform and the computational analysis approaches. And, you know, these are four equal pillars, but as I said, it's really exciting that we've come together to discuss this part because these are so crucial and so central to the entire effort.

And can really make or break this project. So what are the fundamental questions that require theory to address them?

Well, I've named 6 here. Donna and John have thought about this a lot, and we'll discuss these during the breakouts over the next days. There's, of course, cell type and cell state. How much information is actually encoded in the molecular profile? How big is the atlas?

How are we going to decide to define the bounds and what the focus is.

How are we going to sample cells? Knowledge-based or random or combination of both?

How do we normalize and integrate in order to have this reference that amalgamates data from different centers and methods?

And as I said, this is really key for the Cell Atlas to add value basically as a global consortium. What are the strategies to integrate single-cell RNA sequencing and the spatial data? And so I want to concentrate now for a few minutes on how much information is encoded in the molecular profile.

How can computational methods basically suck information or knowledge out of the data? You know, let's look at that a little bit.

And then what are the strategies?

I'll talk a little bit about the interface between single-cell RNA sequencing and spatial data. So the data— in terms of single-cell transcriptomics has increased exponentially.

It sort of followed a Moore's law of single-cell genomics over the last 5 to 10 years. And so this is basically— so charts the number of single cells in studies and the time scale in the publications. And you can see this is a log scale.

So essentially, there's been one log increase in the number of cells in published studies over the past few years. So this has really been a step change in terms of what we can learn from the data because the richness of these data sets allows us to make much deeper inferences.

And I'm going to use an example specifically from something, a paper that we published a few months ago with a time course, a dynamic time course of an infection challenge in mouse where we profiled T cells and the antigen-presenting cells that talk to them. And so one of the things that we can infer from cells that are collected at different time points, so each color is a time point, day 0, 2, 3, 4, 7.

What we can learn from these few hundred cells is that they actually align in this way developmentally.

So the, as Aviv showed, the pseudotime inference basically tells us about the developmental time course.

And so in essence, this means that, you know, these snapshot datasets are giving us developmental time, at least in an average sense, at a very high resolution.

And this is what I want to emphasize here in the context of this computational meeting, is that this is using a machine learning method based on Gaussian process latent variable modeling that Fabian has done some work on these types of methods, and this was with Oliver, who'll be talking later, and also Neil Lawrence in Sheffield.

And so the, you know, what's really been an eye-opener to me over the past 5, 6, 7 years is the power of machine learning and this whole class of methods.

And not only did this help us trace the pseudo-time inference, the developmental time and the two different major latent variables or major processes at the heart of this.

But it also allowed us to reconstruct the cellular decision-making going on in this immune challenge.

So Aviv has mentioned T helper cells, these are also T helper cells. And here in this context, the T cells have to decide whether they're turning left or whether they're turning right in this cellular decision-making process. And what we've used is overlapping mixtures of Gaussian processes to reconstruct that decision-making in an unbiased way. And identify not only the two final populations in red and in blue, but also how the cells are having the molecular circuits that control the fate decision into these two different subsets. Thirdly, the exact same data together with the antigen-presenting cells that we know communicate with the T cells tell us also the circuits of cell-cell communication. So we can basically use our knowledge of protein families, protein classes, and sequence properties of proteins to predict or have a database of which are the secreted ligands and which are the cell surface receptors on the plasma membrane. And we can then dynamically correlate the ligands and their expression in one cell type with the receptors in another cell type and predict an axis of communication that's taking place. So this ligand is secreted by the monocytes, and the receptor is increasing expression on the T cells and controlling this interaction.

Of course, we need to validate this, but you can see the principle of computational prediction also in this context. So it's not just individual cells, it's how they relate to each other. So there's a lot that we can learn, and, and we can infer from these rich datasets. Now, the second pillar of the Human Cell Atlas that Aviv has also referred to is the spatial relationships between the cells, the spatial technologies.

And these are arguably at a slightly earlier sort of point in their evolution and development. But you can also see basically log increases in the number of reported cells, the number of genes that are profiled in parallel with these methods. And there's basically a, you know, a sort of state space, essentially, of different methods that profile more cells or more genes at the same time. And also the field of view, the tissue area, which is reflected in the size of these little points. So you basically have SeqFISH, MERFISH, and many methods developed here in Stockholm. Such as the spatial transcriptomics and the other methods. And I'll go—

I'll talk a little bit about how we can analyze them.

So this is an example from the Lundberg paper last year where they have essentially a tissue slice on an array with a poly-T capture of the mRNA and then little voxels that are barcoded. So how can we identify interesting genes in these patterns in an unbiased way, again, using the power of machine learning? So we want to see which—

how genes are expressed in a spatially biased way and not evenly expressed, as we would in different sort of cells, as we would from a disaggregated experiment such as the single-cell RNA sequencing experiments that I described earlier. But here we're looking in the context of this two-dimensional coordinate system.

And we're again using Gaussian processes now in this spatial context to identify genes that have either a spatial bias, so a function that's dependent on the coordinates, or a periodic bias, and it can be linear or sort of described by a radial function. So these are the kinds of examples. Rickard mentioned a method similar to this at the Keystone meeting earlier this week. And here we can use it to infer the genes that are expressed in a spatially biased way in this tissue slice.

What this— taking this a little bit further, we could then use this information also to cluster genes and find optimal or optimized clusters of groups of genes that are expressed in these regions that then correspond to cells or cell states that are present in the different regions. And you can see how these two different kinds of information relate to each other to give us a fuller picture of tissue architectures in the human tissue histology, essentially, anatomy. And this is—

the beauty of this is that these are unbiased methods, and the inferences are quantitative and valuable.

So I'll just finish off by saying a few words about the growth of the HCA community, looking back into the history for the people who weren't there at the two previous meetings. So we started off, we kicked off with a meeting in London last October.

And really launched with a group of people that was a cross-section of different communities. And then we had a more focused technologies meeting in Stanford in February, and now of course we are here at the computational meeting in Stockholm. And looking forwards, there's another meeting in the Weizmann, or in Israel, in the middle of October, and then next March on the 8th and 9th—

this is the—

date is mistaken— This will be back in the UK.

And so there'll be information about all these meetings on the website, and you can go look at it there. And this is really, you know, hundreds of people have now joined the Human Cell Atlas community, and this growth of the community and this interest and excitement, I think, is fair to say that it's tremendously gratifying to Aviv, myself, and all of us who are involved. And extremely exciting.

We've also been working together on a few publications, a review that was written together by all the meeting participants at the October meeting. This should be last year. It's on bioRxiv, and it's under revision now. And writing the white paper is an ongoing effort that many of us are involved in, and I think it's— that's a valuable sort of reference document for the community. And then Aviv and I are writing a commentary invited by Nature. There have been calls for the Human Cell Atlas, and the Chan Zuckerberg Initiative has been— had an international technology call that closed in April, and that will sort of act as a catalyst, I think, going forward for pilots over the next year.

And at that juncture, I will basically summarize and leave the questions here, and Aviv and I will answer questions. Thank you.