Scientist Stories: Yutong Wang, Decoding cell cell communication through deep learning

pairs from transcriptomics data. So I'll start with the motivation. So in order for the multicellular organisms to function well, then the cells, they definitely are not isolated. They have to be constantly talking with each other. So then the ligand-receptor interaction is indeed central to the cell-cell communication. So what do I mean by ligand-receptor interaction? So as we can see here in the figure, Essentially, the sending cell, they will be able to secrete the small molecules in green here. They're the ligand, and they will bind with the receptors on the surface of the receptor cells. Yeah, so this is how cells would communicate through the ligand-receptor. However, it still remains to be a computational challenge to identify or predict the ligand-receptor interactions. That actually underlie the cell-cell communication mechanism. And this is especially important for G-protein-coupled receptor, GPCR. It's basically essentially just as the figure shows here, it's a 7-transmembrane domain receptor. And it's actually also the largest family of the cell surface hormone receptors. This is a very important family of receptors because GPCR, they're actually targeted by 35% of the FDA-approved drugs. However, actually, we have not fully explored the potential of GPCR yet. The reason is that approximately 20% of the human GPCRs still remain orphan. By orphan, I mean we do not know the endogenous ligands that will bind to these GPCR. So it's actually very important to be able to de-orphanize for these GPCR in order to better understand the signaling pathway and also potentially to identify the novel drugs. So, yeah, that's here the biological question we want to answer, how to identify the unknown ligands for the receptors. Okay, so in order to answer this question, here we turn to the principle of spatial proximity. So the cell-cell communication is the signaling mechanisms of cell-cell communication can be classified based on the distance. So here on the top, we can see that's autocrine action. Essentially, the small molecules secreted by one cell will actually act onto the same cell itself. And then the second type is called paracrine action, where the ligands will act onto the nearby cells. Last but not least, there is also the endocrine action, Essentially, the ligands will act on the distant cell. And normally, the circulation is through bloodstream or something like that. So due to the localized autocrine and paracrine effects, so normally, or most of the time right now, inference cellular communication often relies on leveraging the spatial proximity from spatial transcriptomics dataset. So thanks— to the very nice illustration from the two keynote speakers about spatial trans-omics. I think, yeah, hopefully we all have a very good understanding about spatial trans-omics right now. And here, I mean, just to briefly recap. So here I'm showing the data of spatial trans-omics data of four different genes. Essentially, for each gene, the x and y-axis that correspond to the spatial coordinate in space.. And for each pixel here, for each dot that corresponds to one pixel, where each pixel contains like normally like 1 to 10 cells, multiple cells, and the color for each pixel that corresponds to the spatial gene expression for this specific gene in that specific spot or pixel. So Yeah, this is spatial transcriptomics. And then so essentially, why can we leverage on the principle of spatial proximity? This is because, as what we're talking about, because of the autocrine and paracrine effect. So the cells that are close to each other are more likely to interact. So this implies that if two genes, for instance, like the ligand-receptor pairs, if they are indeed interacting,, then it's more likely that they will share some very similar spatial gene expression pattern. So for instance, as we can see on the top here, for these two genes, their spatial gene expression pattern or the spatial distribution is very similar. And indeed, in the actual biology, they are indeed interacting. Notice that here the interaction, that's not inferred based on the spatial transcriptomics. That's the underlying ground truth. And yeah, so this is not like the bottom pair where for this pair of ligand-receptor, they're actually indeed not interacting and they do not share a similar spatial distribution. Okay, so what are some of the existing methods that are trying to solve the problem? So the first big category include, for instance, CellphoneDB or CellChat. Essentially, they're They're trying to curate the known ligand-receptor interactions from the literature. However, the problem with that is that they leverage on the prior knowledge, and they're not suitable for predicting the novel interaction. Actually, yeah, I'll also show some of the results if we try to predict the novel interaction using them. The other is SpatialDM that was published earlier this year. Essentially, yeah, they're actually having similar methods like this. Essentially, what they're trying to do is they're trying to leverage the spatial statistics, for instance, such as spatial cross-correlation like Moran's I, to try to identify the coexpressed ligand-receptor pairs. However, the problem with the spatial statistics is that they lack capture like the complex and also nonlinear spatial patterns. Last but not least, there is also GCNN, which is a kind of graph neural network that's trying to solve the problem. However, the problem with that is that, so even though there are like 20,000 protein encoding genes, but there are only like hundreds of ligand and receptor genes, like encoding the ligand receptor. And this only includes about 1,000 to 2,000 true interactions. So for this kind of graph neural network, because they rely on supervised learning, then, and, but they also have very limited label data. So this is actually very hard for them to truly make use of the spatial transcriptomics dataset from like 20,000 genes. So, okay, now we have our question. How could we try to come up with a method that can predict the novel interaction and also can capture the nonlinear spatial relationship, and most importantly, how could we learn the useful representation from the massive unlabeled spatial trans-omics data? Okay, yeah, so here what we're proposing is basically a two-step method where the first step is that we want to treat the gene spatial expression as an image for unsupervised feature extraction. What I mean here is basically for the spatial gene expression of each gene, we take that as one, we consider that as one image, and then we just pass the 20,000 protein coding gene images through this convolutional autoencoder. And then the task here is just trying to reconstruct the images. This will help us to learn the latent embedding space for each gene. So, okay, after we get the pre-trained latent embedding for each gene, then we can move on to the second step. That's for supervised classification. So essentially, okay, as we can see from the left, so we can feed the model with a pair of the pre-trained latent representations of one ligand and one receptor. So we have each of them to go through one subnetwork of this Siamese neural network. Essentially, what Siamese neural network is trying to do is that, so basically, we have this twin, a pair of twin subnetwork with shared model architecture and model weights. This is trying to encourage our model to be able to do similarity modeling. Okay, so after we have both the ligand receptor go through this subnetwork, And then we aggregate the output, and then we just run that through some simple binary classifier in terms of whether for this specific pair of ligand-receptor, whether they interact or not. Okay, yeah, so that is our two-step approach where we can make use of all the 20,000 protein coding genes, not only including the ligand-receptor corresponding genes. So our method is called GGPair, and indeed it's performing much better compared to other state-of-the-art. So what are we trying to do here, or what is the experiment? So the data that we're using here is a mouse brain spatial transomics data. And so essentially we're performing a 5-fold cross-validation. So for all the receptor genes, we split that into 5 different folds, and then we take Each time we take the 20% of the receptor genes as testing, and then the rest of the receptors as training. And then for the ligands, we can just— yeah, this is because— so here we're splitting the receptors for training, testing, splitting because we want to prevent data leakage. And then for the ligand, we can just randomly match with the receptors. And we repeat this 5-fold cross-validation using 20 random seeds to ensure robustness. And so, as we can see here, so the, basically the metric that we're using here is the area under the precision-recall curve. And on the x-axis, the leftmost correspond to all the receptor genes. And then we stratify based on different receptor families, including this GPCR family that we mentioned at the very beginning. So our method is called GGPAIR. It's to the very left. You can see that our model is performing much better compared to other state-of-the-art methods. Okay, and so focusing on gPCR, then we can do the next experiment, as in leave-one-out validation. So essentially what we're doing here is in each of the leave-one-out iterations, extract one single GPCR receptor out as testing, and then we just pair this single receptor gene with all the possible ligands. So this will be our testing set. And then essentially, so then we'll be able to get the predicted probability for each of those pairs, and then we can rank all the ligands based on the predicted probability, so which is the x-axis. That you can see here. So in expectation, we hope that the actual, the ground truth, true interacting ligands will always be in the top-ranked genes based on the predicted probability. And indeed, that's the case for our method GGPair. Most of the true interacting ligands are in the top 30 genes. But this is not true for other state-of-the-art Okay, so in summary, we introduced this new method called GGPair, which is a novel framework for ligand-receptor interaction prediction using spatial transomics. And so essentially what we're doing here is like a dual deep learning approach with unsupervised convolutional autoencoder for spatial feature extraction, and then followed by supervised meso network for interaction classification. So our method outperforms the existing methods, enhancing the understanding, especially for GPCR signaling, which is a very critical venue for the drug discovery. And I also want to acknowledge my colleagues, Fending and Antoine, on this project, and also my advisor. Okay, yeah, with that, I'll be happy to take any questions.

Thank you.

Do you have any rationale of why the ligand-receptor interaction would be nonlinear? And maybe one or few examples that the simple linear geostatistic methods fail to discover those pairs, but it could be your deep learning model will be able to predict well.

Yeah, well, okay, essentially in terms of the linearity, That's just based on the Euclidean distance, right? But I mean, in terms of the actual tissue, in terms of the communication or in terms of how the different cell states within the tissue, they might be spatially distributed, that might not be linear, which is why it's better to consider that as like a complex image.

For predicting cell-ligand-receptor interaction, right? So there is something other than co-expression. Then I was wondering what would be that thing? Like, what's the biology of those nonlinear, like, not co-expressed ligand-receptor interactions?

Yeah. Okay. So I guess, yeah, your question is in terms of this part. So, um, so first of all, like in terms of the actual ligand-receptor interaction, I think it's, there could be like actually multiple different steps in terms of the whole signaling pathway that might be very complicated. And also in terms of here, I think, oh, we're not saying that spatial co-expression is not helpful at all. We're just saying that in terms of the co-expression, it's not sufficient to only consider the linear co-expression. As in, if we only use the spatial statistics, more like you can consider that as Pearson correlation in the spatial sense, then that's not sufficient. That's what I mean here. Okay.

Okay. Oh, thanks. Thanks. Nice talk. I wanted to understand the autoencoder setup a little bit more. Is it, you have multiple images or you have, or each gene effectively counts as like one sample and then it's one network for every, like shared across all genes.

I see. Yeah. It's like one network shared across all the genes.

Okay. So you just need one image because then the genes are kind of like your samples in how we would normally think about training.

Are you saying that each gene can be a channel?

Is that what you're saying? Well, it could be, but that's not what you're doing, right? Because this is just one channel. And so your sample size for training this is like 20,000 or whatever the number of genes is.

Right. Yes. Okay. Thank you very much. So the next speaker is Mu Qiao from LinkedIn. He's going to talk about deciphering the genetic code of neuronal type connectivity. Okay.

So yeah, first, thanks to the organizers for this wonderful conference, and it's exciting