For decades, biological classification has relied on a fundamental assumption: that cells can be neatly sorted into discrete buckets. We talk about T-cells, neurons, or epithelial cells as if they are fixed, isolated identities. This discrete state assumption is computationally convenient—it allows us to use standard classification algorithms, draw clean boundaries, and simplify highly complex datasets into digestible summaries.
However, nature rarely operates in discrete categories. Processes like embryonic development, cell division, immune responses, and cancer progression are inherently fluid. Cells exist on a continuous spectrum of states, constantly shifting through intermediate phenotypes. When we force these continuous trajectories into rigid, discrete clusters, our computational models begin to break down.
An emerging paradigm shift in computational biology moves away from hard categorization, focusing instead on modeling the continuous geometry of cellular manifolds.
The Cost of Discretization
When traditional algorithms encounter a continuous biological gradient, their default behavior is to slice it into pieces. As [Guo‐Cheng Yuan] et al. point out in Challenges and emerging directions in single-cell analysis, this lack of clear boundaries creates severe hurdles for classification models, as cells frequently exist on a continuous spectrum rather than within fixed categories.
Forcing a continuous process into discrete buckets introduces two primary issues: labeling noise and trajectory fragmentation.
1. Labeling Noise and Artificial Thresholds
Consider a cell activating a specific pathway. In a discrete framework, researchers often categorize expression levels into binary "high" or "low" states. As [Margaret P. Chapman] et al. demonstrate in Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer, this binary simplification introduces significant noise and model inconsistencies. It completely ignores intermediate phenotypes that govern critical behaviors, such as how cancer cells gradually shift identities to escape targeted therapies.
In a single-author study, [Abel Szkalisity] expands on this in Regression plane concept for analysing continuous cellular processes with machine learning, explaining that forcing fluid phenomena like cell division into discrete classes obscures the subtle transitions between phases, creating artificial boundaries where none exist biologically. This issue is not unique to transcriptomics; [Niksa Praljak] et al. show in Integrating deep learning with microfluidics for biophysical classification of sickle red blood cells adhered to laminin that forcing continuous morphological changes (like cell sickling) into rigid classes misrepresents the true spectrum of disease progression.
2. Trajectory Fragmentation
When dimensionality reduction or clustering methods prioritize finding distinct groups, they inadvertently tear apart the underlying biological timeline. [Caleb Weinreb] et al. observe in SPRING: a kinetic interface for visualizing high dimensional single-cell expression data that standard visualization tools fragment continuous developmental trajectories into disconnected groups. This makes it incredibly difficult to see how cell types relate to one another over time.
Similarly, [Renee S. Hoekzema] et al. argue in Multiscale methods for signal selection in single-cell data that traditional differential expression analysis misses key genes driving smooth gradients because it relies on comparing pre-grouped, discrete clusters.
Discrete Clustering (Slices the Manifold):
[ Cell State A ] ---> | Boundary | ---> [ Cell State B ]
(Intermediate transitional details are lost or treated as noise)
Continuous Trajectory Modeling (Preserves the Manifold):
[ Cell State A ] ~~~~~~ (Intermediate Spectrum) ~~~~~~> [ Cell State B ]
(Gradients, densities, and fluid transitions are preserved)
Why Hard Boundaries Fail Mathematically
To understand why discrete clustering fails, we must look at the geometry of high-dimensional biological data. Cells are typically represented as points in a high-dimensional space where each dimension corresponds to a measurable feature (such as gene expression or protein abundance).
When cells are transitioning from one state to another, they form a dense, continuous path—a manifold—in this space. Standard clustering algorithms (like $k$-means or Louvain community detection) operate by partitioning this space, effectively drawing hyperplanes that cut through the data.
As [Ioana Bica] et al. explain in Unsupervised generative and graph representation learning for modelling cell differentiation, models optimized purely for identifying discrete clusters struggle to represent data that exists as a continuous spectrum. When boundaries are fluid or overlapping, standard computational approaches fail to capture the specific sub-states within those populations.
Furthermore, [Zhe Sun] et al. note in A Bayesian mixture model for clustering droplet-based single-cell transcriptomic data from population studies that assuming distinct categories prevents models from identifying pseudotime trajectories where cells exist entirely on a spectrum. This limitation is echoed by [Tongxin Wang] et al. in BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes, who point out that data alignment frameworks assuming discrete groups fail to accurately map continuous biological transitions across different experimental batches.
The problem even transcends cellular biology. In A hierarchical machine learning framework for the analysis of large scale animal movement data, [Colin J. Torney] et al. show that traditional movement models fail when they assume animals transition between a small number of discrete behavioral states (like foraging vs. traveling), because real-world activity levels shift along a gradual, continuous spectrum.
Preserving Geometry: How Continuous Modeling Works
To capture the true fluid nature of biological systems, researchers are developing frameworks that preserve manifold continuity. Instead of asking "Which cluster does this cell belong to?", these methods ask "Where does this cell sit on the continuous data manifold?"
Several distinct mathematical mechanisms have emerged to achieve this:
1. Path Metrics and Density-Based Distances
Traditional Euclidean distance measures a straight line between two points through empty space. If a cell population curves along a developmental trajectory, Euclidean distance will cut across the curve, incorrectly calculating the distance between early and late cells.
To solve this, [Andriana Manousidaki] et al. introduced scPMP in Clustering and visualization of single-cell RNA-seq data using path metrics. By using power-weighted path metrics, the algorithm calculates distances through high-density regions of the data manifold rather than straight lines. This preserves local cluster details while simultaneously maintaining the global geometry of the developmental trajectory, preventing transition states from being incorrectly merged or isolated.
2. Continuous Latent Spaces via Deep Generative Models
Rather than clustering cells explicitly, deep generative models like Variational Autoencoders (VAEs) compress high-dimensional data into a low-dimensional continuous latent space.
- Ensemble VAEs: [Foo Wei Ten] et al. developed the resVAE ensemble framework in resVAE ensemble: Unsupervised identification of gene sets in multi-modal single-cell sequencing data using deep ensembles to map the "fuzzy" identities and continuous transitionary states inherent in development and disease progression without forcing cells into single categories.
- Multi-Resolution Inference: In Deep generative modeling of sample-level heterogeneity in single-cell genomics, [Pierre Boyeau] et al. introduce MrVI, a model that bypasses predefined cell types entirely. By modeling continuous variations across a spectrum of cellular activation, it detects subtle, counterfactual variations across patient cohorts that rigid clustering would obscure.
- Continuous Deconvolution: This approach is particularly powerful in spatial transcriptomics. Standard methods try to assign discrete cell types to physical tissue locations. In contrast, [Romain Lopez] et al. developed DestVI in Multi-resolution deconvolution of spatial transcriptomics data reveals continuous patterns of inflammation. By modeling cellular variation as a continuous spectrum of functional states rather than discrete sub-clusters, DestVI avoids the mathematical instability of high-granularity clustering and successfully maps continuous environmental gradients, such as inflammation or oxygen deprivation.
3. Supervised Regression Planes and Multiscale Laplacians
Instead of unsupervised clustering, some frameworks map phenotypes directly to continuous geometric shapes. [Abel Szkalisity] introduced the Regression Plane (RP) framework, which maps cellular phenotypes onto a continuous 2D space, allowing machine learning models to track fluid transitions without categorical labeling noise. On the mathematical signal side, the multiscale methods proposed by [Renee S. Hoekzema] et al. utilize the Multiscale Laplacian Score to capture smooth gradients across multiple structural scales of the data graph, identifying critical genes along trajectories without requiring discrete cell groupings.
Nuances and Hybrid Frameworks
While continuous modeling captures biological reality more accurately than discrete clustering, it introduces its own set of challenges.
First, biological processes are not always perfectly smooth. As [Nevena Zivanovic] et al. note in Learning time-varying information flow from single-cell epithelial to mesenchymal transition data, systems can exhibit discrete state jumps or highly divergent, non-linear branching patterns. Models that strictly enforce a smooth, linear path can struggle when cells undergo abrupt, discontinuous transitions.
Second, discrete labels remain highly useful for human interpretation and communication; doctors need to know if a cell is a T-cell or a B-cell.
The cutting edge of this research direction therefore focuses on hybrid frameworks that bridge the gap. For example, [Xiyue Cao] et al. developed scPriorGraph in scPriorGraph: constructing biosemantic cell–cell graphs with prior gene set selection for cell type identification from scRNA-seq data. Their approach integrates discrete identity labels with continuous trajectory analysis by embedding biological prior knowledge (like gene pathways) directly into cell-cell graphs.
By combining the interpretability of discrete categories with the mathematical flexibility of continuous manifolds, these modern frameworks allow researchers to name the destinations while fully mapping the fluid journeys cells take to get there.
Key Foundational Work
Challenges and emerging directions in single-cell analysis (2017) — contributed 1 concept elements to this direction.
Single-cell RNA-sequencing uncovers transcriptional states and fate decisions in haematopoiesis (2017) — contributed 1 concept elements to this direction.
SPRING: a kinetic interface for visualizing high dimensional single-cell expression data (2016) — contributed 1 concept elements to this direction.