Clustering

Unsupervised

The default method is Leiden clustering. Leiden is a graph community detection algorithm which optimizes the measure of modularity. Simply put, Leiden is looking for dense communities in a graph. Since most of the single cell data is not structured as a graph, an intermediate method is needed to transform a sample-by-feature matrix into one. A common approach is to add an edge between each sample (cell) and its neighbors (as determined by Euclidean distances). For small datasets, we compute the exact \(k\) nearest neighbors in the reduced space and use them to construct an unweighted graph. For larger datasets (>5,000 cells), we switch to an approximate nearest neighbors approach that is based on the faiss package. Faiss greatly reduces the computation runtime at little cost in accuracy. This default behavior can be changed at anytime by expanding Leiden’s settings and changing the Graph Construction Method. Leiden does not require prior knowledge about the number of clusters, however, this number is affected by the resolution parameter (default: \(1\)). Resolution controls the tradeoff between inter and intra cluster densities, with higher resolution resulting in more clusters. This is the most important hyperparameter to tune when running Leiden. For CODEX datasets, we obtained good results when the resolution was set between \(0.1\) and \(0.3\).

Other algorithms include KMeans, Spectral Clustering, and Agglomerative Clustering. Unlike Leiden, for these three you need to specify the number of clusters in advance. To ease the tuning process, we allow the specification of multiple #clusters. This can be in the form of a list, e.g. \([4, 8, 16]\) will spawn three instances of the algorithm, but it can also be a range \((4, 9, 1)\) which will run the algorithm five times, one for each number between \(4\) and \(8\) (inclusive). The best cluster configuration is then determined by using the Silhouette Score.

Uncertainty clustering can only be run after labels have already been obtained by any of the other algorithms. Uncertainty clustering generates a new cluster with ID -1, for cells that have a high uncertainty score. The uncertainty score is computed according to the cells’ distance from the cluster centroids in the reduced space. Re-clustering the newly generated uncertain cluster using Constrained Leiden oftens leads to better results. Intuitively, the improvements come from incorporating information about the cluster centers into the algorithm, which is not considered in vanilla Leiden.

Semi-Supervised

A simple extension of the Leiden algorithm has been included in Cellar to allow re-clustering of the data while keeping certain clusters fixed. In other words, given a set of points \(\mathcal{A}\) and a “fixed” subset \(\mathcal{F}\in\mathcal{A}\), the cluster assignments for all the points in \(\mathcal{F}\) will not change during the iterations of Leiden. This is particularly useful for re-distributing any poor quality clusters that were obtained due to noise in the process of label transfer, or even in combination with uncertainty clustering where low-scoring cells may need to be re-clustered but homogeneous clusters need to be kept intact. Clusters can be fixed by expanding the settings tab for Semi-Supervised Leiden and ticking the checkbox next to that cluster’s or subset’s name.

Label Transfer

Often it is desirable to perform annotation of cell types in a supervised manner where the supervision comes in the form of “reference data”. This has the advantage of saving time, but also more importantly avoiding any human biases or errors during the annotatation process. Cellar implements two such methods for transferring labels in a supervised manner: 1) The Ingest function from Scanpy, and 2) SingleR. Ingest works by first projecting the query and reference data into a shared PCA space and then using e nearest neighbors classifier to assign labels. SingleR, on the other hand, is a correlation based method that queries the reference dataset and iteratively removes low scoring cell types until only one remains. This procedure is applied to each cell individually so SingleR is typically slower than Ingest. For both algorithms, we only consider the features which are in common for the query and reference data, so please make sure the naming convention for the two is the same.

NOTE: In Cellar, label transfer should be used under dual mode. The “active” dataset serves the role of a query dataset, while the “inactive” one will be used as a reference.

“Cell ID Based” label transfer simply matches the cell barcodes between the two datasets and copies cluster assignments from one to the other. This is useful when analyzing two data modalities that were profiled from the same cells, such as in the case of SNARE-seq.