How does the structure of data impact cell-cell similarity? Evaluating how structural properties influence the performance of proximity metrics in single cell RNA-seq data

ESCHR: a hyperparameter-randomized ensemble approach for robust clustering across diverse datasets

Clustering is widely used for single-cell analysis, but current methods are limited in accuracy, robustness, ease of use, and interpretability. To address these limitations, we developed an ensemble clustering method that outperforms other methods at hard clustering without the need for hyperparameter tuning. It also performs soft clustering to characterize continuum-like regions and quantify clustering uncertainty, demonstrated here by mapping the connectivity and intermediate transitions between MNIST handwritten digits and between hypothalamic tanycyte subpopulations. This hyperparameter-randomized ensemble approach improves the accuracy, robustness, ease of use, and interpretability of single-cell clustering, and may prove useful in other fields as well.

A hyperparameter-randomized ensemble approach for robust clustering across diverse datasets

Clustering analysis is widely used to group objects by similarity, but for complex datasets such as those produced by single-cell analysis, the currently available clustering methods are limited by accuracy, robustness, ease of use, and interpretability. To address these limitations, we developed an ensemble clustering method with hyperparameter randomization that outperforms other methods across a broad range of single-cell and synthetic datasets, without the need for manual hyperparameter selection. In addition to hard cluster labels, it also outputs soft cluster memberships to characterize continuum-like regions and per cell overlap scores to quantify the uncertainty in cluster assignment. We demonstrate the improved clustering interpretability from these features by tracing the intermediate stages between handwritten digits in the MNIST dataset, and between tanycyte subpopulations in the hypothalamus. This approach improves the quality of clustering and subsequent downstream analyses for single-cell datasets, and may also prove useful in other fields of data analysis.

The impact of similarity metrics on cell-type clustering in highly multiplexed in situ imaging cytometry data

Motivation

The advent of highly multiplexed in situ imaging cytometry assays has revolutionized the study of cellular systems, offering unparalleled detail in observing cellular activities and characteristics. These assays provide comprehensive insights by concurrently profiling the spatial distribution and molecular features of numerous cells. In navigating this complex data landscape, unsupervised machine learning techniques, particularly clustering algorithms, have become essential tools. They enable the identification and categorization of cell types and subsets based on their molecular characteristics. Despite their widespread adoption, most clustering algorithms in use were initially developed for cell suspension technologies, leading to a potential mismatch in application. There is a critical gap in the systematic evaluation of these methods, particularly in determining the properties that make them optimal for in situ imaging assays. Addressing this gap is vital for ensuring accurate, reliable analyses and fostering advancements in cellular biology research.

Results

In our extensive investigation, we evaluated a range of similarity metrics, which are crucial in determining the relationships between cells during the clustering process. Our findings reveal substantial variations in clustering performance, contingent on the similarity metric employed. These variations underscore the importance of selecting appropriate metrics to ensure accurate cell type and subset identification. In response to these challenges, we introduce FuseSOM, a novel ensemble clustering algorithm that integrates hierarchical multiview learning of similarity metrics with self-organizing maps. Through a rigorous stratified subsampling analysis framework, we demonstrate that FuseSOM outperforms existing best-practice clustering methods specifically tailored for in situ imaging cytometry data. Our work not only provides critical insights into the performance of clustering algorithms in this novel context but also offers a robust solution, paving the way for more accurate and reliable in situ imaging cytometry data analysis.

Availability and implementation

The FuseSOM R package is available on Bioconductor and is available under the GPL-3 license. All the codes for the analysis performed can be found at Github.

The specious art of single-cell genomics

Dimensionality reduction is standard practice for filtering noise and identifying relevant features in large-scale data analyses. In biology, single-cell genomics studies typically begin with reduction to 2 or 3 dimensions to produce "all-in-one" visuals of the data that are amenable to the human eye, and these are subsequently used for qualitative and quantitative exploratory analysis. However, there is little theoretical support for this practice, and we show that extreme dimension reduction, from hundreds or thousands of dimensions to 2, inevitably induces significant distortion of high-dimensional datasets. We therefore examine the practical implications of low-dimensional embedding of single-cell data and find that extensive distortions and inconsistent practices make such embeddings counter-productive for exploratory, biological analyses. In lieu of this, we discuss alternative approaches for conducting targeted embedding and feature exploration to enable hypothesis-driven biological discovery.

Effect of distance measures on confidences of t-SNE embeddings and its implications on clustering for scRNA-seq data

Arguably one of the most famous dimensionality reduction algorithms of today is t-distributed stochastic neighbor embedding (t-SNE). Although being widely used for the visualization of scRNA-seq data, it is prone to errors as any algorithm and may lead to inaccurate interpretations of the visualized data. A reasonable way to avoid misinterpretations is to quantify the reliability of the visualizations. The focus of this work is first to find the best possible way to predict sample-based confidence scores for t-SNE embeddings and next, to use these confidence scores to improve the clustering algorithms. We adopt an RF regression algorithm using seven distance measures as features for having the sample-based confidence scores with a variety of different distance measures. The best configuration is used to assess the clustering improvement using K-means and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) based on Adjusted Rank Index (ARI), Normalized Mutual Information (NMI), and accuracy (ACC) scores. The experimental results show that distance measures have a considerable effect on the precision of confidence scores and clustering performance can be improved substantially if these confidence scores are incorporated before the clustering algorithm. Our findings reveal the usefulness of these confidence scores on downstream analyses for scRNA-seq data.