In this talk, Karin Isaev (also referred to as Karen Asaev) from David Knowles's lab presents Leaflet, a probabilistic model designed to jointly infer latent cell states and differential splice-junction usage (psi) from single-cell RNA split‑read data. The presentation begins with an overview of alternative splicing as a driver of protein and cellular diversity, illustrated by the CALCA gene’s tissue‑specific isoforms. Existing differential splicing methods (Bre2, SciX) are discussed, highlighting their reliance on pre‑annotated cell types and limited ability to discover novel events. Isaev then describes how splice junctions can be quantified from short‑read split‑reads, modeling junction counts with a binomial distribution and usage ratios with a beta distribution. Leaflet extends this framework by introducing latent cell‑state assignments and a global proportion variable, learning all parameters via mean‑field variational inference and coordinate ascent variational inference (CAVI), leveraging sparse matrix operations for efficiency. Validation on 8,000 Smart‑Seq2 mouse brain cells from the Tabula Muris atlas shows that Leaflet recovers cell‑type clusters consistent with known annotations and uncovers splicing‑driven subpopulations within microglia. The model also provides a differential splicing score (approximate log Bayes factor, LBF) that highlights junctions such as those in the PSAP gene, which serve as splicing‑specific markers. Future directions include moving from discrete to continuous cell‑state representations, adapting Leaflet to high‑throughput platforms like SplitSeq, and exploring cross‑species conservation of splicing patterns. The Q&A addresses how Leaflet handles extreme sparsity (by modeling only observed junctions and relying on overall read distributions) and its relation to mixed‑membership models like Latent Dirichlet Allocation.
Questions this video answers
We only model cell‑junction pairs where the junction is observed; for many cells we make no updates for a given junction. We rely heavily on the overall distribution of observed junction reads, and performance depends on the sequencing platform—Smart‑Seq2 provides sufficient coverage, whereas high‑throughput platforms like SplitSeq are more challenging due to greater sparsity.
Initially we tried a mixed‑membership model inspired by LDA to allow a cell to exhibit multiple topics (cell states), but it proved too challenging with current data. The final Leaflet model is a mixture model, though conceptually related to LDA.
Chapters
Hi, everyone. My name is Karen Asaev. I'm from David Knowles's group. I'm really excited to be here and share with you some of the work we've been doing focused on studying RNA splicing as a source of cellular diversity. And so ultimately, we're interested in modeling how variations in alternative splicing, essentially this process of exons being stitched together to form mature RNAs, may contribute to phenotypic differences that we observe among cells. So to start off, we know that one of the core principles of alternative splicing is to contribute to the protein diversity that we observe among species. A nice example of this is actually this gene here called CALCA, which can be alternatively spliced to skip the yellow exon and produce the protein calcitonin. And this protein is expressed basically only in thyroid cells, where it acts to reduce calcium levels in the blood. Alternatively, it can be spliced where it actually includes the yellow exon, and this ends up producing a totally different protein that is actually expressed only in neurons, where it is released to control vasodilation and pain pathways. So not only is this process leading to two different functioning proteins, those proteins are also expressed in a very cell-type-specific manner. And so we're interested to study this sort of phenomenon more globally to see how these trends contribute to cell diversity. More recently, recent works have shown preliminary investigation into this process and how it may vary across single cells during development, such as shown here, where we can look at individual cells across this type of trajectory and evaluate which isoforms are expressed at each stage of this process using differential splicing analysis, for example. And so this can help us better understand the underlying biology of these distinct cell populations And we can, of course, imagine expanding this approach to studying the brain, where we know that alternative splicing is very prevalent, but it's often understudied. So how do we actually go about doing this type of differential splicing analysis on the single-cell level? This work is still, you know, it's pretty recent. Over the last couple of years, these sort of technologies have even started allowing us to look at this problem. And so several methods already exist for doing this, such as Bre2. Which uses Bayesian regression to model read counts across two potential transcript isoforms. This approach, however, requires cell type annotations, and it can only sort of model two contrasts of cells at a time, such as disease cells or healthy cells. And it also requires an annotation of all the sort of splicing events that you're interested in looking at. Meanwhile, a newer method called SciX, or I think they like to say "six," actually, It is a probabilistic model that's inspired by autocorrelation. And so here, rather than requiring this sort of pre-annotated cell types, it focuses more on modeling whether an exon will be cell-state associated or cell-state independent across this more continuous space. So what if we wanted to discover potentially new splicing events, for example, those not annotated in the reference genome? And also, what if we didn't have well-sorted cells with nicely defined marker genes to define those cells in advance. So to answer the first question, let's first consider that most widely used sequencing platforms actually give us short reads, which are not very helpful when it comes to measuring the abundance of full transcript isoforms, such as the one shown on the left. So instead, we can quantify each splice junction, such as the one shown here between the pink and purple exons, for example. On the right. And so we can evaluate then the proportion of reads that map to these junctions relative to the other junctions within the same sort of clustering event. And so then we can essentially rely on these splicing observed reads to guide our annotations of the splicing events rather than relying on some reference genome or some reference annotation of these splicing patterns. The link on the left is in white, so you might not be able to see it. Sorry. But in our lab, we've been sort of developing several computational and experimental approaches to actually try and knock down different splice junctions with CRISPR-Cas13, for example, to actually observe the functional consequences of those changes. So how do we actually quantify these splice junctions? Like I said, we can use short reads. And so specifically, we're looking at the split reads. And so we can extract the number of split reads that correspond to these different splice junctions. And in this very nice simple illustration, we essentially have 3 exons in yellow, red, and teal. And so we can look at the different split reads that we obtained from our sequencing experiment and sort of that tells us what splicing event has occurred. And so here basically these splice junctions overlap, so you can have sort of the exon being skipped or the exon being included. And so these all correspond to one splicing event. Splicing event where we can model the observed junction counts with a binomial distribution, where the total number of trials in this case is just the total number of split reads in this overall splicing. We call them like intron clusters. And then the individual junction usage ratios we can model with this variable psi. Okay, and then furthermore, we can actually model this junction usage ratio itself using a beta distribution. So it ranges from 0 if the junction is never utilized, or 1 if it is always utilized. But how can we take this step further to model not only different junctions, but the same junction across different cell types? And so this is actually what would allow us to do differential splicing analysis, for example, across different conditions or cell types. And so if we had these different clusters of cells, we can imagine each color corresponding to a different cell type. We can imagine this hypothetical scenario where this one particular junction is differentially utilized among them. But again, this sort of approach would require some sort of predefined cell type annotations, which we often don't have. And also, as we saw in a previous talk, these annotations are often not actually discrete. And so what if we're interested in learning more about potential subpopulations of cells that might exist? In your sample. So to that end, we developed a model called Leaflet, sort of a next kin of Leaflet, or sorry, Leafcutter, the previous model in our lab. And so here we can visualize the setup of all our variables and their dependencies. And so the two that I have already mentioned are shown in gray, which are the observed junction counts and intron cluster counts. Then, sorry, here, Then we have the PSI value, which represents the learned junction usage ratio, which we can actually model as we're sort of learning all of these things. And then we are allowing this junction usage ratio to vary across cell states K. So uniquely, what we have here now is also the cell state assignment variable. We're basically simultaneously learning for each cell what cell assignment or cell state assignment it is coming from, while also learning which junctions are contributing to each cell state. And lastly, we also have here a latent variable just representing the overall proportions of all the cell states or cell types in the data. And I'm using cell state and cell type really interchangeably, but ultimately we're just imagining sort of different groups of cells existing in space that have something in common. So how do we actually learn these variables? We can make use of the mean field assumption, which allows us to nicely approximate the posterior distribution here with this factorized variational posterior, and we can assume that each group of the latent variables we saw are independent of each other. And so we can define the variational parameters for the individual latent variables in our model. And to actually optimize this, we employ coordinate ascent variational inference where we can iteratively optimize one parameter at a time while keeping the others fixed, basically trying to improve our approximation of the true posterior distribution. I just wanted to briefly touch here on what the actual cavi updates might look like, and really just to highlight that one of the advantages of this approach is efficiency. So because we're dealing with single-cell data, it is very sparse, and so we can make use of sparse matrix that can help us significantly speed up the inference. And the updates themselves are also quite simple, which is nice and straightforward and allows our model to be easy to implement and efficient. So how does this actually work, or can we rely on these random initializations for the parameters? To test that, we basically took 8,000 cells from the tabula muris ATLAS. So basically brain cells from several different mice, specifically sequenced with a SmartSeq2 platform. And so those are just— they have predefined annotated cell types, which are shown on the left. And then essentially we ran Leaflet 10 random times, or randomly initializing parameters 10 random times, running the whole thing, and then getting the cell state assignments that it produced. And then essentially looking at how often cells were co-assigned together. And so then clustering all of that, we can see if pairs of cells were consistently assigned to the same cluster. And it is quite nice to see that overall the cells do cluster together in these sort of big yellow blocks, which also correspond to the sort of predefined cell type that they're coming from. So the overall structure makes sense, and luckily we're not seeing like totally random assignments. Of cells. So that's good, but what else can we learn using Leaflet? So if we just focus on the same 8,000 cells from the mouse, in this case, this representation was made using total gene expression. So that's kind of what is most commonly done. And basically, we can see the different cell types in the mouse brain sort of nicely separated. But we can, on top of this, now also add the cell state assignments that we get from Leaflet basically on top of the total gene expression representation. And so we see overall these sort of two labels, you know, like correspond pretty well. But in addition, we can see that within the microglia, for example, which are shown here corresponding to those pink cells on the left, there are potentially subpopulations of cells in this cluster that are driven primarily by their splicing differences rather than total gene expression. And just to zoom in on it more, plotting can sometimes hide what's really going on. So now the orange and blue are flipped. So actually most of the cells here are in cell state 1, which is in blue, but still like several hundreds of cells corresponding to microglia are attributed to a different cell state, which again may be driven more by splicing differences that are there. So to touch on another useful thing that Leaflet can do is actually conduct differential splicing analysis sort of simultaneously, right? So while we're learning the cell state assignments, we can also find the junctions that are differentially spliced between them. And so to give these junctions a formal score, we call it the— or we use the approximate log base factor here, which is determined by comparing the probabilities that the parameters alpha and beta, which are again just those variational parameters for modeling our psi latent variable, across— so the probability that those vary across K cell states with the probability that they're essentially the same across K cell states. And so we can give each junction in our data a score. And basically when we did this across all the cells, one of the top set of scoring junctions came from this gene called PSAP. And so the gene on its own is not really a marker gene. This is just sort of normalized total gene expression for this gene. We can see that it's sort of ubiquitously nicely expressed among cells. But if we look at the two junctions that we recover with our LBF score or leaflet score, we can see now instead that on the splicing level for these two specific events, they really nicely mark the different— some of the different cell types that exist in the data. So overall, I think there's great potential to kind of use this to find novel, not just total gene markers, but splicing-specific factors. For different cell types and cell states. Yeah. So in terms of future directions and where we would like to take this work, basically, you know, so far these cell state assignments have really been discrete, and we would really like to move into modeling the cells on a more continuous level. And so we're actively working on that. We would also like to extend the model to be able to incorporate data from high-throughput experiments such as that from SplitSeq. From ParsBio. And a very interesting potential avenue for this is to also look at the— sorry, exploration of the conservation trends of splicing in cell types across species, especially the human and mouse, for example, for which we have a lot of data across different organs. And so in summary, we present Leaflet, which is the probabilistic model that can infer differential splicing while also modeling the latent cell states in the data. And so we have the code up, and I'm also happy to chat anytime about anything splicing or single-cell related. And most importantly, I would just like to thank everyone in my lab, and several folks from our lab are also presenting posters later. And so if you're interested in splicing and other things in our lab, then please check them out. And that is all for me today. Thank you.
Great. And we have time for some questions. So, I don't think I'm allowed to ask them. So, it strikes me that, I mean, obviously, this kind of data is usually sparse, right? Now, if you're looking at these splicing junction events, and if you have, like, 9-to-1 ratio, now you're order of magnitude sparser, right? So I have a hard time getting a sense for how your model deals with that extreme sparsity. Are you borrowing information or just the fact that you have a distributional assumption helps with the sparsity?
Yeah, so basically for— we're really only modeling cell junction pairs where that junction is observed. And so for many instances, we basically wouldn't make any updates for that junction for a given group of cells. So I think we rely heavily on the sort of overall distribution of junction reads that are observed. The sparsity issue also really depends on which sequencing platform you're using the data from. So, you know, with this SmartSeq approach, we're actually able to get pretty convincing results that make biological sense. But with high-throughput sequencing platforms such as SplitSeq, where, you know, we initially thought, well, the reads are actually coming from sort of across the transcript, maybe we'll have better coverage there. Like, the sparsity is just— you just can't manage it. So I think part of it is just relying on this high sort of high-coverage data. So we still have enough signal in overall across cells for given splicing events.
Thanks for a really interesting talk. When your graphical model slide, it went past a bit fast for me to really posit, but could you comment? I mean, is it similar at all to latent Dirichlet allocation, those kinds of models? Document topic models. LDA.
Yes. Yeah. So, one of the first things we actually tried to do was a mixed membership model rather than just a mixture model. And so the idea there was to kind of be inspired by LDA and to also try and model within a cell if there are sort of multiple topics or multiple cell states that it's exhibiting. That was really the main goal, but it turned out to be a little bit too challenging to achieve. With the data that we have. But yes, it's very much related.
Thank you. Okay, if there's no more questions, we'll go on to the next speaker. And let's thank Karen again. So next up we have Alexander Sace. Um, who will be telling us about multimodal biophysical convolutional neural networks for cis-regulation.
