Trajectory Inference

Assigns each patient or cell a pseudotime value representing its position along a continuous biological process. Unlike real time, pseudotime is a latent dimension inferred from molecular similarity. The algorithm identifies a root state and orders samples by their progressive transcriptional divergence. This reveals transitional states invisible to static clustering, such as cells undergoing differentiation or patients progressing from pre-disease to advanced pathology.

Builds a minimum spanning tree or k-nearest neighbor graph on the high-dimensional molecular space. Algorithms like Monocle and Slingshot then identify the longest path through this graph as the primary trajectory. Branch points in the graph represent fate decisions—moments where a biological process diverges into distinct outcomes, such as a progenitor cell committing to one lineage over another.

Identifies genes whose expression changes smoothly as a function of pseudotime using generalized additive models (GAMs). Rather than comparing discrete groups, this analysis reveals cascades of transcriptional regulation. Key outputs include:

• Switch genes: abruptly activated at specific pseudotime points
• Temporal modules: co-expressed gene clusters peaking at distinct phases
• Branch-dependent genes: differentially expressed between trajectory branches

Combines trajectory inference with RNA velocity, which estimates the rate of change in gene expression by comparing unspliced and spliced mRNA ratios. This adds a directional arrow to pseudotime, resolving the arrow of time ambiguity inherent in static snapshots. The resulting vector field predicts future cellular states, distinguishing between a process moving forward versus reversing.

Abstracts trajectories as paths through an epigenetic landscape, a metaphor for developmental potential. Cells roll downhill through valleys representing stable attractor states, with bifurcations representing lineage choices. Computational implementations use optimal transport and diffusion maps to reconstruct this landscape from single-cell data, quantifying the probability of transitioning between any two states.

Applies trajectory inference to clinical cohorts by ordering patients along a disease severity continuum using multi-omics profiles. This replaces binary case-control designs with a continuous view of pathogenesis. Applications include:

• Modeling cancer evolution from pre-malignant to metastatic states
• Ordering neurodegenerative progression from prodromal to advanced stages
• Identifying critical transition points preceding irreversible organ damage

The computational assignment of a quantitative measure of progress to each cell or patient sample along a biological process. Unlike real time, pseudotime is a latent dimension inferred from molecular similarity. Cells are ordered based on gradual transcriptomic changes, positioning them along a trajectory from a starting state to an end state. The unit of pseudotime is arbitrary and reflects transcriptional distance rather than clock time. Algorithms like Monocle and Slingshot use this concept to reconstruct differentiation hierarchies or disease progression arcs.

A graph-theoretic structure connecting all data points with the minimum possible total edge weight, forming the backbone of many trajectory inference algorithms. In single-cell analysis, an MST is constructed on clusters or individual cells in a reduced-dimensional space. The longest path through this tree is often designated as the main trajectory. This approach, central to tools like Monocle 2, assumes that biological transitions follow a parsimonious path of minimal transcriptional change, avoiding spurious connections between distant cellular states.

A method that predicts the future state of individual cells by distinguishing between unspliced and spliced mRNA transcripts. The ratio of nascent (unspliced) to mature (spliced) mRNA acts as a proxy for the rate of gene expression change. By modeling transcriptional kinetics, RNA velocity vectors point toward a cell's predicted future state in gene expression space, providing a directional arrow on a trajectory. This transforms a static snapshot into a dynamic map, resolving the directionality ambiguity inherent in standard pseudotime methods.

A non-linear dimensionality reduction technique that models the probabilistic diffusion process of transitioning between data points. Unlike PCA or t-SNE, diffusion maps explicitly capture the multi-scale structure of branching differentiation pathways. The method defines a random walk on the data, where transition probabilities reflect local similarity. The resulting diffusion components represent the slowest modes of change in the system, naturally ordering cells along developmental trajectories. This is a foundational component of the Destiny R package for trajectory inference.

A smooth, one-dimensional curve that passes through the middle of a data cloud, generalizing the concept of principal components to non-linear manifolds. In trajectory inference, principal curves are fitted to high-dimensional single-cell data to model continuous biological processes. The Slingshot algorithm extends this to principal graphs—branching structures that simultaneously fit multiple curves to capture lineage bifurcations. Each cell is projected onto the nearest point on the curve, assigning both a lineage identity and a pseudotime value.

A conceptual metaphor visualizing cell differentiation as a ball rolling down a landscape of bifurcating valleys. Each valley represents a stable cell fate, and ridges separate distinct lineages. Trajectory inference algorithms computationally reconstruct this landscape from high-dimensional molecular data. Methods like Waddington-OT use optimal transport theory to model the probabilistic flow of cells through this landscape, capturing the stochastic nature of fate decisions. This framework unifies the concepts of pseudotime, branching, and fate probability.