Optimization of clustering parameters for single-cell RNA analysis using intrinsic goodness metrics

The accurate clustering of cell subpopulations is a crucial aspect of single-cell RNA sequencing. The ability to correctly subdivide cell subpopulations hinges on the efficacy of unsupervised clustering. Despite the advancements and numerous adaptations of clustering algorithms, the correct clustering of cells remains a challenging endeavor that is dependent on the data in question and on the parameters selected for the clustering process. In this context, the present study aimed to predict the accuracy of clustering methods when varying different parameters by exploiting the intrinsic goodness metrics.

This study utilized three datasets, each originating from a distinct anatomical district and with a ground truth cell annotation. Moreover, the investigation employed two clustering methods: the Leiden and the Deep Embedding for Single-cell Clustering (DESC) algorithm. Firstly, a robust linear mixed regression model has been implemented in order to analyze the impact of clustering parameters on the accuracy. Consequently, fifteen intrinsic measures have been calculated and used to train an ElasticNet regression model in both intra- and cross-dataset approaches to evaluate the possibility of predicting the clustering accuracy.

The first-order interactions demonstrated that the use of the UMAP method for the generation of the neighborhood graph and an increase in resolution has a beneficial impact on accuracy. The impact of the resolution parameter is accentuated by the reduced number of nearest neighbors, resulting in sparser and more locally sensitive graphs, which better preserve fine-grained cellular relationships. Furthermore, it is advisable to test different numbers of principal components, given that this parameter is highly affected by data complexity. This procedure has enabled the effective prediction of clustering accuracy through the utilization of intrinsic metrics. The findings demonstrated that the within-cluster dispersion and the Banfield-Raftery index could be effectively used as proxies for accuracy, for an immediate comparison of different clustering parameter configurations.