Gang Li, a postdoctoral researcher in Bill Noble's lab at the University of Washington, presents SKEPTIC (Single Cell Soot Time Classifier), a novel framework for inferring pseudotime in time‑series single‑cell sequencing and imaging datasets. The talk begins by illustrating the challenge of asynchronous cell differentiation, which leads to multimodal pseudotime distributions that unsupervised methods (PCA, Monocle) fail to capture accurately. Li reviews SuperTime, a penalized logistic‑regression approach that improves over earlier methods but still suffers from limited predictive power, rank‑based outputs, and lack of proper train‑test validation. SKEPTIC addresses these issues by employing a multiclass support‑vector‑machine (SVM) that learns class‑conditional probabilities for each time point, uses multiple score functions (one per class), incorporates nonlinear decision boundaries via an RBF kernel, and adopts nested cross‑validation to avoid overfitting. The model’s pseudotime is obtained as the conditional expectation of the class labels, yielding a continuous trajectory. Validation on mouse embryonic stem‑cell differentiation data shows higher classification accuracy (94% vs. SuperTime) and more realistic temporal predictions, confirmed by leave‑one‑time‑point analyses that demonstrate the model captures true temporal structure rather than batch effects. Benchmarking across six public scRNA‑seq datasets demonstrates uniformly superior performance (statistically significant via Wilcoxon tests). The framework is further extended to single‑nucleus imaging data (using ResNet embeddings), to sex‑specific analyses in scATAC‑seq and scRNA‑seq (revealing delayed differentiation in female cells), and to single‑cell Hi‑C data (using contact decay profiles). The talk concludes with acknowledgments, code and manuscript availability on bioRxiv, and a brief Q&A discussing methodological intuition, continuous versus discrete pseudotime, and applicability to datasets with few time points.
Questions this video answers
SKEPTIC uses multiple score functions—one per class—rather than a single logistic regression with four thresholds, greatly increasing predictive power. Additionally, the RBF kernel introduces non‑linearity, allowing more flexible decision boundaries compared to the linear logistic model.
SKEPTIC provides both discrete class labels and a continuous pseudotime obtained via conditional expectation of the class probabilities. The work focused on comparison with SuperTime because SuperTime already outperforms other methods like PCN and Monocle.
The framework requires at least three time points for robust training. With only two time points a classification model could still be trained, but the authors have not evaluated performance in that scenario and consider it an interesting direction for future work.
Chapters
Hello, everyone. My name is Gang Li. I'm a postdoc at Bill Noble's lab here at UW. Today, it's my great honor to present our recent work, "Situational Analysis for Time Series Single Cell Sequencing and Imaging Data." As yesterday, we have seen many talks talking about cells are growing asynchronously and also cells differentiate at variable states. Here, I show you a very simple example for a mouse embryonic stem cell differentiation process. Where cells differentiate to embryonic body cells and all the way to the neuroprogenitor cells. And imagine for a cell, maybe at day 7, you have growth and transit to embryonic body cell, but even another cell at the same time point may grow slightly faster than this cell, and another cell maybe grow slightly slower than this cell. When you combine all the cells at the same time point, look at their cell time, you might see a unimodal distribution. And when you have a time series single-cell sequencing dataset, you might have a multimodal pseudotime distribution. And the goal of our project is to robustly infer the pseudotime for such kind of time series sequencing datasets. And the microscopy images are coming from the Bonora et al. Genome Biology paper. So once you have pseudotime for a time series sequencing dataset, here is a gene by cell matrix. Each cell has a time label. You can use them to identify some time-varying genes, the genes that are changing along the pseudotime trajectory. And also you can order your cells based on the pseudotimes. We are not the first one to tackle this problem. People in the field first tried the principal components. Usually you'll find your first principal components roughly correlated with the time label you have. Here in this pancreas cells, There are 8 donors. The age of the donor ranges from 1 to 54. You find the first component roughly correlated, but you also find a lot of the models are kind of overlap with each other. And more, there's another method called Moloco. If you're in first-year time, you might heard of this before, but as this example, you will see one of the monocle components is correlated with time, but they put the H21 on the rightmost, which puts the— basically, the first component doesn't quite reflect the true time label we have of the single cells. So we found that for those unsupervised methods, their first components usually are driven by other global variability. The temporal variability might at the second or the later components might be different from the global variability. Based on that, recently there is a method called Supertime, which adopts ordinary logistic regression, which explicitly use a time label for each cell to train a penalized ordinary logistic regression to select a subset of the genes and to infer the pseudotimes using the time labels. Here is a counter-interactive situation where you have a simple score function. Here you have 5 time labels. SuperTime will infer 4 thresholds for this score function and classify the first group versus the rest, the first 2 groups versus the rest, so on and so forth. However, when they applied to the same dataset, although the performance is much improved compared with PCN model, we still see a lot of overlap in this pseudo-time model distribution. We multiply it by this limited classification accuracy, and also the pseudo-time inferred by Supertime is rank-based. You see the range of the pseudo-time on the x-axis is from -10 to 15, doesn't quite reflect the true donor ages. So based on that, we propose a new framework called SKEPTIC, Single Cell Soot Time Classifier, which basically adopts a multiclass support vector machine model to first infer the class conditional probabilities for each cell and then infer soot time by taking the conditional expectation. And compared with the previous model, we improve our predictive power using not just one score function for all different classes. We will adopt k different score functions. And also our SKEPTIC model would allow for nonlinearity when you consider, say, radial-based function, kernel function to generate nonlinear decision boundaries. The third difference is for the SuperTime is showing your training performance on the whole dataset where we propose to do nested cross-validation to provide a more robust and prevent overfitting on the training dataset. What we do is for each dataset, for each fold of the data, for example, the yellow test fold, we will train a model on the purple set and predict the zero-time on the yellow set and repeat this for all the datasets we have. And we first test our model and compare with supertime and mouse embryonic stem cell differentiation data. On the left is a confusion matrix for our SCAPI model. On the right is a confusion matrix for supertime. A higher number on the diagonal indicate better performance of our classification model. And in this comparison, our model achieved a 94% accuracy compared with Supertime, slightly less than 90%. And the improvement is significant according to a Fisher exact test. And we also want to point out Supertime does not have a train-test split. It's simply the training performance. So despite the bias towards Supertime, we're still outperforming them. And you might be curious if our SKEPTIC model are simply learning batch effects or the real temporal structure, because imagine in a scenario for each time label, if you have a strong batch effect specific for that time label, you can still train SKEPTIC, achieve a perfect confusion matrix look like this. To confirm that, we perform a leave-one-time-out analysis. So every time we mask, for example, the T1, time and train a model on t2 to t5 and predict all the cells from t1. So in a scenario, if the model are simply learning batch effects, we expect the model to predict uniform errors on the off-diagonal. So in this case, since we masked the time for each time point, so the diagonal are all 0. And if the model are truly learning some temporal structure, we will expect a successful model to predict time points to be adjacent time points. So if we mask T1, we'll expect the model to predict most of the cells to be T2 instead of the rest of the later 3 time points. And when we do this leave-one-time-point analysis, we found indeed our model are predicting the most of the cells close to the diagonal. And this is a 3-time And what is a little bit different is that we don't— for the day 11 cells, we don't predict too much on the day 21. This is because the middle 3 time points are embryonic body cells. So they look— for the day 11, they look more similar to day 7 instead of the day 21 neural progenitor cells. And on the right-hand side is pseudotime visualization instead of confusion matrix. After test out on this dataset, we next benchmark our SCAPI model on 6 public single-cell RNA-seq datasets. 5 of them evaluated in the original Supertime paper, which they have shown to outperform PCN-Monocoat. So we directly compare their performance on those 5 datasets. And on all the 6 datasets, our SCAPI model achieved uniformly better performance in terms of the classification accuracy. And the improvement is also significant according to Wilcoxon rank test. After showing that the SCAPRI model can work well on single-cell RNA-seq data, we next test, can we generalize this model to single nucleus imaging data? Our collaborator, Renjing, from Dr. Fuller's lab here at UW, they help us to collect the same 5 time point differentiation mouse embryonic stem cell differentiation data, but have the single nucleus image. So here I show you some random crops of the image from 5 time points. And specifically, instead of running on 2D image itself, we adopt an image processing pipeline which using the pre-trained ResidualNet to embed all the 2D images to a 1D smaller than 24-dimensional representation. And after we have the 1D vector for each image, we run the SKEPTIC model and compare with SuperTime. Again, our model can achieve improved classification accuracy, and the improvement is significant according to the Fisher exact test. And more importantly, we find the SKEPTIC model can provide a more fine-grained label compared with the original time point label. For this, we compute the absolute correlation of the pseudotime for each cell with some morphologic features. Here I show two of them, nuclear area and nuclear perimeter. We find skeptic pseudotime have a higher absolute correlation. This actually are supported by the literature that during the mouse embryonic stem cell differentiation, process, the embryonic stem cell will have large nuclei, and for MBC, they will have smaller nuclei. After showing the SCACT model can work well on the image data, we further test on the sex differences on both single-cell ATAC and single-cell RNA-seq dataset. So here I show you the male single-cell ATAC sequencing result from SCACT model. So they don't have the 5 time points, only 4 for this one. But when we compare it with the female cell performance on the single-cell attack, for all the cells of day 7, if we look at the attacks to the time from skeptic model, we find the female cells have a high value tail on the left-hand side. This actually is because during the embryonic stem cell differentiation, the female cells shows some delayed differentiation pattern, which is well known in the literature. We did a similar analysis for the single-cell RNA-seq dataset by comparing all the RNAs pseudotypes for the day 7 cells. We observe a similar delay pattern in the RNAs pseudotype. And last but not least, we have also tried to extend our framework to the single-cell Hi-C data. For this purpose, we extract the contact decay profile for each single-cell Hi-C data and then run our SCEPI model. We get a pretty decent accuracy. And here are the pseudotimes for the 5 time points. You see those models are not very overlapped with each other. Hopefully, I have convinced you our SCEPI model can outperform the state-of-the-art method SuperTime in both time series single-cell sequencing and imaging data. And we have shown that SCEP-TE can work well for various types of the dataset. And it can detect sex-specific differentiation from single-cell ATAC and sequencing data. In our manuscript, we also show that we can identify some methylation delay using some co-assay data. But with time limit, I will skip this. And our code is available. The manuscript is also available on the bioRxiv. With that, I would like to thank my advisor, Bill Noble, and Christine and Xinchen for providing us small sample samples and Doug Fuller's lab for generating the images for the analysis, and John Philip for the machine learning expertise, and all the Noble Lab members. And thank you for listening. I'm happy to take any questions. Questions for Gong? Hi, thank you for the super cool method. I was wondering if you could provide any intuition based on the formulations why your method performs better than Supertime? Yeah. Yeah, so there are 3 major motivations. The first one, the logistic regression is only one score function with 4 thresholds. And for our multiclass SBI model, for each classification, you will have a score function. So that greatly improves predictive power. And also logistic regression model is a linear-based model. And with RBF kernel, we can have nonlinear decision boundary, which also outperforms. The traditional ordinary logic suppression. That's a major motivation. Thank you.
I just wanted to ask, how would this method compare to other pseudo-time inference methods that try to map out a continuous time label rather than these discrete time labels that you use for your model? I was wondering if there's an advantage over using Septiq over those kinds of methods.
So just want to clarify, our model have both discrete label and also the continuous label. For example, I have shown a plot like this. We have a continuous pseudo-time inference by conditional expectation. So we did not compare with other methods like PCN and Monaco because in the previous evaluation, SuperTime have been shown that outperform other supertime inference method. So we directly compare with, focus our comparison with supertime in this manuscript.
Nice talk. So I was wondering how your method does if you have very few time points. So if you have, obviously you can't do anything if there's only one.
Yeah.
But if you just have two, then I guess life gets more difficult and like then how about a comparison to monocal and that sort of approach?
Yeah, that's a great question. So for the time series dataset, at least we request at least three time points, but when there are only two time points, you can still train a classification model, but I'm not sure. We haven't evaluated on such dataset, but that's very interesting to look at.
Feels like in that setting, there should be something that's like in between, like your thing, or we can look at the trajectory.
Thanks for the suggestion. All right, let's give Gong another round of applause.
