Endotype Discovery

The foundational machine learning technique for endotype discovery, which groups patients based on inherent molecular similarities without using predefined diagnostic labels. Unlike supervised classification, clustering reveals natural disease subtypes by partitioning high-dimensional omics data into distinct, homogeneous groups. Key algorithms include k-means, hierarchical clustering, and density-based methods like DBSCAN, each suited to different data distributions and cluster shapes. The output is a set of candidate endotypes that must then be validated against functional mechanisms and clinical outcomes.

A critical preprocessing step that transforms high-dimensional omics data into a lower-dimensional space while preserving its essential structure. In endotype discovery, where datasets often contain tens of thousands of features (genes, proteins, metabolites) for relatively few patients, dimensionality reduction combats the curse of dimensionality. Techniques like Principal Component Analysis (PCA) extract linear variance, while non-linear methods like t-SNE and UMAP excel at visualizing complex patient manifolds, often revealing distinct endotype clusters that are obscured in the raw feature space.

A statistical framework specifically designed for the unsupervised integration of multiple heterogeneous omics data types—such as genomics, transcriptomics, proteomics, and metabolomics—from the same patient cohort. MOFA infers a set of latent factors that capture the principal sources of biological variation across all data modalities. These factors directly represent the underlying molecular drivers of endotypes, revealing which specific pathways and biological processes define each patient subgroup, rather than relying on a single data type.

A computational method that constructs a separate patient similarity network for each available data type (e.g., mRNA expression, DNA methylation, microRNA). These individual networks are then iteratively fused into a single, comprehensive network that captures both shared and complementary information. Community detection algorithms applied to the fused network identify robust patient clusters. SNF is particularly powerful for endotype discovery because it avoids the assumption that all data types contribute equally, allowing the strongest biological signal to emerge naturally.

A resampling-based methodology that addresses the inherent instability of clustering algorithms when applied to noisy biological data. Consensus clustering runs a chosen algorithm (e.g., k-means) on multiple subsamples of the patient cohort and aggregates the results into a consensus matrix, which quantifies the frequency with which each pair of patients is assigned to the same cluster. This process identifies robust and reproducible endotypes that are not artifacts of a single random initialization or parameter choice, providing statistical confidence in the discovered subgroups.

Also known as pseudotime analysis, this computational method orders patients or single cells along a continuous path based on their molecular profiles. Unlike discrete clustering, trajectory inference models dynamic disease progression and identifies transitional states between endotypes. This is critical for understanding diseases that evolve over time, such as cancer metastasis or neurodegenerative conditions, where a patient may not belong to a static cluster but rather exists along a continuum of molecular change driven by a specific pathobiological mechanism.