Our implementation of metric based iterative clustering pipeline starts with performing Leiden clustering on a pre-processed dataset at a very low resolution, followed by evaluating each cluster using a selected quality metric like Shannon entropy or silhouette scores in a specified feature space (e.g., PCA or Cell by Gene Matrix).

Once initial clustering is run, the pipeline identifies low-quality clusters from an ideally broad set of assignments. For example, clusters with high entropy or low silhouette scores get flagged and the algorithm re-clusters those subsets using the leiden algorithm, and replaces them with finer subclusters to improve overall cluster quality. Additionally, NS-Forest can be used to apply re-clustering on a reduced matrix of only binary genes.

This process could be repeated for a fixed number of iterations or when convergence is reached which we calculate by tracking global/cluster-wise silhouette score values and stopping the process when scores stop improving. We also integrate UMAP plots and bar plots of quality scores to monitor progress.