HDBSCAN

The core innovation of HDBSCAN is its ability to identify clusters with significantly different densities within the same dataset. Unlike DBSCAN, which uses a global density threshold (eps), HDBSCAN builds a hierarchy over all possible eps values. It then uses a cluster stability metric to select the most persistent clusters across the hierarchy, naturally accommodating both dense and sparse regions without parameter tuning.

HDBSCAN provides a membership probability for each point, quantifying how strongly a point belongs to its assigned cluster. This soft clustering is derived from the point's distance to the cluster's core. Points with low probability are robustly classified as noise or outliers, a critical feature for identifying anomalous patient profiles or data artifacts in biomarker discovery pipelines.

The primary hyperparameter is min_cluster_size, which specifies the smallest grouping to be considered a cluster. This is highly intuitive for patient stratification, where a clinician can specify the minimum number of patients required to define a clinically meaningful subgroup. This single parameter replaces DBSCAN's eps and min_samples, dramatically simplifying model selection.

HDBSCAN constructs a mutual reachability graph using a novel distance metric. The mutual reachability distance between two points is the maximum of: the core distance of point A, the core distance of point B, and the original distance between them. This effectively pushes apart points in sparse regions, making the separation between dense clusters and noise more pronounced and robust.

After building a full single-linkage hierarchy, HDBSCAN computes a minimum spanning tree on the mutual reachability graph and converts it into a dendrogram. It then condenses this tree by pruning small, unstable splits. The algorithm traverses the condensed tree and selects flat clusters that maximize the sum of relative excess of mass (stability), providing an optimal, data-driven cut.

HDBSCAN is frequently paired with UMAP for dimensionality reduction before clustering, as UMAP preserves more global structure than t-SNE. In single-cell RNA sequencing workflows, the combination of UMAP for visualization and HDBSCAN for clustering has become a standard pipeline, often outperforming Louvain or Leiden community detection in identifying rare cell populations and transitional states.

HDBSCAN excels at identifying clinically meaningful disease subtypes (endotypes) from complex patient data without requiring a pre-specified number of clusters. Unlike centroid-based methods, it naturally handles the varying densities of patient subgroups—common in conditions like asthma or sepsis where a large, diffuse group of mild patients coexists with a small, dense cluster of severe patients. The algorithm's outlier detection capability is critical here, as it prevents forcing every patient into a cluster, instead flagging individuals with unique molecular profiles for further clinical review. This aligns with the biological reality that not all patients fit neatly into a single subtype.

In single-cell transcriptomics, HDBSCAN is a powerful tool for identifying discrete cell types and transient cell states. Its density-based logic naturally separates rare cell populations from abundant ones without the need for post-hoc filtering. Key advantages include:

• Variable density handling: Distinguishes a small, dense cluster of activated T-cells from a large, diffuse cluster of resting lymphocytes.
• Hierarchical output: The condensed tree allows researchers to explore cell populations at multiple levels of granularity, from broad lineages to fine subtypes.
• Noise robustness: Technical dropout events in scRNA-seq data create sparse regions; HDBSCAN correctly labels these as noise rather than forcing them into spurious clusters.

HDBSCAN is applied to spatial transcriptomics and multiplexed imaging data to identify spatially coherent tissue domains and cellular neighborhoods. By clustering cells based on both their molecular profiles and spatial coordinates, HDBSCAN reveals the tissue architecture underlying disease. For example, in oncology, it can delineate the tumor core, invasive margin, and immune-excluded stroma as distinct spatial clusters. The algorithm's ability to find clusters of arbitrary shape is essential here, as biological structures like tumor buds or tertiary lymphoid structures are rarely spherical in feature space.

HDBSCAN can be used to cluster nodes in patient similarity networks, where edges represent clinical or molecular similarity between individuals. After constructing a graph using methods like Similarity Network Fusion (SNF), HDBSCAN operates on the network's distance matrix to identify communities of phenotypically similar patients. This approach is particularly effective for:

• Multi-omics integration: Clustering patients based on fused similarity across genomics, proteomics, and imaging data.
• Rare disease cohort identification: Finding small, tight-knit groups of patients sharing a cryptic molecular signature within a large, heterogeneous registry. The minimum cluster size parameter provides direct control over the granularity of the stratification.

The outlier persistence concept in HDBSCAN makes it a natural choice for anomaly detection in clinical flow cytometry and mass cytometry (CyTOF) data. Instead of requiring a separate one-class classifier, HDBSCAN simultaneously clusters the major cell populations and flags rare events. This is critical for:

• Minimal Residual Disease (MRD) monitoring: Identifying a tiny cluster of leukemic cells among millions of normal hematopoietic cells.
• Biomarker discovery: Detecting a small, dense cluster of patients with an extreme response to therapy, whose shared molecular features can be investigated as predictive biomarkers. The GLOSH outlier score provides a continuous measure of how anomalous each data point is.

By applying HDBSCAN to time-series clinical data, researchers can identify distinct disease progression trajectories. After extracting features that capture the rate of change in lab values or symptoms, HDBSCAN clusters patients into groups such as 'rapid progressors,' 'slow progressors,' and 'stable.' The algorithm's robustness to noise is vital here, as transient fluctuations in biomarkers do not derail the clustering. The condensed tree visualization allows clinicians to see how trajectory clusters merge at different stability thresholds, providing insight into the hierarchical relationship between different progression patterns.