Automated Classification Across Biological Continua

This direction is merged into another direction: categorical diagnostic limitations. Its last version stays readable and citable here.

Version 1 of 113 papers20 conceptsfrozen 2026-10-04

In clinical medicine and biology, we often rely on discrete categories: a tumor is either benign or malignant; a cell is either a blast or a mature lymphocyte; a skeletal structure is at stage 2 or stage 3 of development. However, nature rarely operates in jumps. Across a wide range of recent research, a consistent theme has emerged: the most difficult challenges in automated diagnosis stem from the fact that biological states exist on a continuous spectrum, while our classification systems demand hard boundaries.

The Discretization Trap

The fundamental problem is that human-defined labels often impose "bins" on a smooth gradient of change. When a biological process is a continuum, the boundaries between these bins become arbitrary and "noisy."

For example, Hairui Li and colleagues, in their work on cervical vertebral maturation, note that growth is a fluid morphological transition. When clinicians try to force these transitions into discrete stages, it leads to high variability. A vertebra doesn't "snap" from one stage to the next; it drifts. Similarly, Shiaofen Fang and Mats Larsson found that human iris patterns, such as Fuchs’ crypts, lack objective morphological thresholds, creating a "gray zone" where even experts disagree.

This ambiguity isn't just a human failing—it is a mathematical hurdle for machine learning. If the "ground truth" labels provided by experts are inconsistent because the underlying biology is a continuum, the model receives contradictory signals during training.

Molecular Overlap and the "Indeterminate" Case

The continuum problem is perhaps most visible at the molecular level. Traditional diagnostics often struggle with "indeterminate" cases—samples that sit exactly between two known states.

In the study of thyroid nodules, Yaoting Sun and Hao Zhang demonstrated that proteomic profiles of benign and malignant tissues significantly overlap. This molecular "blurring" explains why traditional needle biopsies often fail to provide a clear answer. By using deep learning to analyze a 19-protein panel, they were able to map where a sample sits on this spectrum, providing a more precise classification than binary morphology alone.

This pattern repeats in oncology. Marcus Renner and Wolfgang Hartmann showed that high-grade soft tissue sarcomas exhibit overlapping epigenetic signatures, suggesting these cancers exist on a biological continuum rather than as isolated types. Even at the single-cell level, Félicie Giraud-Sauveur and Stergios Christodoulidis argue in the STHELAR dataset that the transition from a healthy to a cancerous cell state is a transcriptomic gradient, making "definitive" ground truth labels for malignancy inherently difficult to establish.

Why Modern Models Can Bridge the Gap

If biology is a continuum, why are new AI models succeeding where previous methods failed? The answer lies in how these models represent data and how researchers are grounding them in multi-modal evidence.

  1. Learning the Latent Space of Development: In hematology, cells mature through a series of intermediate forms. Gregory Goldgof and Shenghuan Sun developed DeepHeme, a classifier that effectively "learns" the maturation spectrum of bone marrow cells. Instead of just memorizing labels, the model's internal representations mirror the actual biological relationships of blood cell development. Similarly, Andrea Loddo and Carsten Marr used a self-supervised foundation model, RedDino, to capture the subtle morphological shifts in red blood cells that define their health and shape.

  2. Multi-Stage Pipelines to Mimic Clinical Logic: Rather than a single "black box" jump from image to diagnosis, researchers like Mahnaz Mohammadi are using two-stage AI systems for cervical biopsies. By first identifying regions of interest and then classifying the severity, these models can better handle the "histological overlap" of precancerous lesions.

  3. Graph-Based Structural Representations: In drug discovery, Yunzhuo Zhou and David B. Ascher found that kinase binding modes are structural continuums. Their tool, kinCSM, uses graph-based signatures to represent molecules. This allows the model to account for "hybrid" binding modes—molecules that share features with multiple classes—rather than forcing them into a single, potentially inaccurate category.

Conclusion

The shift in this research direction is a move away from treating "noise" or "ambiguity" as a data error to be cleaned. Instead, researchers are treating the biological continuum as a fundamental feature of the data. By building models that can navigate these gradients—whether through high-resolution spatial transcriptomics, proteomic signatures, or deep morphological features—the field is moving toward a more nuanced, "spectrum-aware" form of diagnostics that matches the complexity of the biology it seeks to categorize.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_automated_classification_across_biologic_b14270eb,
  title = {Automated Classification Across Biological Continua},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-classification-across-biological-continua/rdr_b14270eba76a4c20ae8c1027f7815469},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_automated_classification_across_biologic_b14270eb_v1,
  title = {Automated Classification Across Biological Continua},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-classification-across-biological-continua/rdr_b14270eba76a4c20ae8c1027f7815469/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}