What is computational pathology?

The application of computational methods, including machine learning and deep learning, to digitized pathology images for tasks such as diagnosis support, biomarker prediction, and increasingly the integration of morphological and molecular spatial data. Recent work has shown that routine H&E images encode recoverable molecular information, which computational pathology methods can learn to extract.

How does spatial biology relate to digital pathology?

Digital pathology digitizes tissue slides for computational analysis, while spatial biology measures molecular data, such as gene expression or protein abundance, with tissue position preserved. The two are converging because morphology visible in a digitized slide and molecular spatial data from the same tissue are increasingly analyzed together, using methods that either generate one modality from the other or learn a shared representation linking them.

Can you combine morphology and molecular spatial data?

Yes, through at least two distinct computational strategies. One generates a synthetic version of one modality directly from the other, such as producing a virtual stained image from spatial transcriptomics data. The other learns a shared representation of both modalities through contrastive training, predicting molecular patterns from morphology without requiring precise pixel-level registration between the two data types.

Spatial biology and computational pathology

Highlights in this article

Two spatial disciplines converging
Whole-slide imaging meets molecular data
Registering morphology and molecules
Computational pathology pipelines
Emerging applications

Practicing computational pathology increasingly means treating a stained tissue slide and a molecular spatial map of that same tissue as two representations of one underlying biological reality, rather than as two separate data types that happen to come from the same specimen. That reframing is what makes computational bridging between morphology and molecular data possible at all, and how that bridge is actually built, not just that it exists, is the subject of this guide.


Key takeaways Computational analysis of routine H&E images can predict complementary molecular readouts, establishing that tissue morphology itself encodes recoverable molecular information. Two genuinely different computational strategies bridge morphology and molecular data: generating one modality directly from the other, or learning a shared representation that sidesteps precise registration entirely. A method avoiding tissue registration achieved weighted F1 scores of 0.735 to 0.830 predicting three different immunohistochemistry stains directly from H&E images. Current pathology foundation models are trained mainly on morphology, while many clinical endpoints are governed by molecular state the model never directly observed, a named structural limitation emerging methods are starting to address. The direction of travel is toward molecular awareness built into a model during training, so that routine clinical deployment does not require running a separate molecular assay alongside every scan.

Key takeaways

Computational analysis of routine H&E images can predict complementary molecular readouts, establishing that tissue morphology itself encodes recoverable molecular information.
Two genuinely different computational strategies bridge morphology and molecular data: generating one modality directly from the other, or learning a shared representation that sidesteps precise registration entirely.
A method avoiding tissue registration achieved weighted F1 scores of 0.735 to 0.830 predicting three different immunohistochemistry stains directly from H&E images.
Current pathology foundation models are trained mainly on morphology, while many clinical endpoints are governed by molecular state the model never directly observed, a named structural limitation emerging methods are starting to address.
The direction of travel is toward molecular awareness built into a model during training, so that routine clinical deployment does not require running a separate molecular assay alongside every scan.

Two spatial disciplines converging

Pathology has been a spatial discipline since long before the term "spatial biology" existed. A pathologist reading a stained tissue section under a microscope is already making a diagnosis from spatial arrangement, which cells sit where, relative to which structures, and what tissue architecture looks like across a region. What has changed is that pathology is becoming molecular and computational at the same time, as whole-slide digital imaging and molecular spatial profiling increasingly examine the same physical tissue section, or adjacent sections, and increasingly ask to be analyzed together rather than separately.

That convergence rests on a specific, foundational finding worth stating precisely. Recent work has shown that computational analysis of routine H&E images can predict complementary pathology modalities, including immunohistochemistry, multiplex immunofluorescence, and spatially resolved transcriptomics. These results establish that tissue morphology encodes latent molecular information recoverable through learned mappings, meaning the visual patterns a pathologist has always read are not merely correlated with underlying molecular biology; they substantially encode it, in a form a computational model can learn to decode.

Whole-slide imaging meets molecular data

That encoding relationship is what turns a whole-slide image from a purely visual record into a potential proxy for molecular data a lab did not directly measure. If morphology genuinely encodes recoverable molecular signal, a sufficiently well-trained computational model can, within limits, estimate a molecular readout from a stain that was never run, rather than requiring every molecular assay to be physically performed on every sample.

That capability does not eliminate the value of actually running a molecular assay when precision matters, but it does change the economics of screening and triage: a computational estimate from a routine, inexpensive H&E stain can help decide which samples merit the cost and turnaround time of an actual molecular assay, rather than running that assay uniformly across every specimen regardless of whether the computational estimate already suggests a clear answer.

Registering morphology and molecules

Bridging morphology and molecular spatial data computationally can be approached in two genuinely different ways, and contrasting them directly clarifies what each is actually solving.


Generate the missing modality, or learn around the alignment problem entirely One strategy generates one modality directly from the other. A method called ST2HE is a cross-platform generative framework that synthesizes virtual hematoxylin and eosin images directly from high-resolution spatial transcriptomics data, integrating nuclei morphology and spatial transcript coordinates using a one-step diffusion model. In effect, it reconstructs what the tissue would look like under a microscope purely from its molecular data, which requires the model to have learned a detailed mapping between spatial transcript patterns and the morphological structures that typically accompany them. A second, structurally different strategy avoids generating either modality from the other at all. A method called HistoStainAlign predicts immunohistochemistry staining patterns directly from H&E whole-slide images by learning joint representations of paired H&E and IHC embeddings through a contrastive training strategy, explicitly avoiding the need for patch-level annotations or tissue registration between the two modalities. Evaluated on three commonly used IHC stains, P53, PD-L1, and Ki-67, the method achieved weighted F1 scores of 0.735, 0.830, and 0.723 respectively. Rather than aligning two images pixel by pixel, it learns a shared representation space in which corresponding morphological and molecular patterns sit close together, sidestepping the registration problem rather than solving it directly.

Generate the missing modality, or learn around the alignment problem entirely

One strategy generates one modality directly from the other. A method called ST2HE is a cross-platform generative framework that synthesizes virtual hematoxylin and eosin images directly from high-resolution spatial transcriptomics data, integrating nuclei morphology and spatial transcript coordinates using a one-step diffusion model. In effect, it reconstructs what the tissue would look like under a microscope purely from its molecular data, which requires the model to have learned a detailed mapping between spatial transcript patterns and the morphological structures that typically accompany them.

A second, structurally different strategy avoids generating either modality from the other at all. A method called HistoStainAlign predicts immunohistochemistry staining patterns directly from H&E whole-slide images by learning joint representations of paired H&E and IHC embeddings through a contrastive training strategy, explicitly avoiding the need for patch-level annotations or tissue registration between the two modalities. Evaluated on three commonly used IHC stains, P53, PD-L1, and Ki-67, the method achieved weighted F1 scores of 0.735, 0.830, and 0.723 respectively. Rather than aligning two images pixel by pixel, it learns a shared representation space in which corresponding morphological and molecular patterns sit close together, sidestepping the registration problem rather than solving it directly.

Neither strategy is universally superior; they solve different problems. Generating a matched image, as ST2HE does, produces an actual synthetic morphological output a pathologist could in principle inspect directly. Learning a shared embedding space, as HistoStainAlign does, produces a prediction without ever needing precise spatial correspondence between training pairs, which is valuable specifically when reliable pixel-level registration between modalities is difficult or unavailable to begin with.

Computational pathology pipelines

A pipeline built around either bridging strategy shares the same basic structure: a whole-slide image is divided into smaller patches for computational tractability, since a full gigapixel-scale slide cannot be processed as a single unit, each patch is embedded using a model that has learned relevant morphological or molecular features, and patch-level outputs are aggregated into a slide-level prediction or, in a generative approach, into a complete reconstructed image.

The specific technique used to bridge morphology and molecular data at the patch or embedding level, generation versus contrastive alignment, is a design choice made within that shared pipeline structure rather than a wholly separate architecture, which is why the two strategies described above are properly understood as different solutions to the same sub-problem within a common computational pathology workflow, not as competing overall pipeline designs.

Emerging applications

The clearest signal of where this technical area is heading comes from work naming a specific, structural limitation in current computational pathology models directly, rather than treating morphology-molecule bridging as already solved.

A recent preprint states the limitation precisely: computational pathology increasingly relies on whole-slide image analysis using pathology foundation models, but many clinical endpoints, including biomarker status, treatment response, and survival, are governed by molecular state, while these foundation models are pretrained mainly from H&E morphology and pathology reports. That is a genuine mismatch between what the model was trained to represent and what the clinical question actually depends on. A described framework, MIST, addresses this by introducing molecular priors into a pathology foundation model’s pipeline without requiring transcriptomic measurements at inference time, so the model incorporates molecular awareness learned during training using paired data, then can be deployed later using only a standard H&E image with no additional molecular assay required at the point of use.

That design pattern, molecular awareness built in during training so that routine deployment requires only the cheap, ubiquitous stain, is a plausible template for how spatial molecular insight could reach routine clinical practice at scale, without requiring every clinical sample to undergo its own spatial transcriptomics or multiplex imaging run. Deep learning approaches applied more broadly to spatial tissue data, beyond this specific morphology-molecule bridging question, are developed in AI in Spatial Pathology: Deep Learning on Tissue Data.

For the broader convergence argument and clinical infrastructure this spoke builds on, see Translating Spatial Biology into the Clinic: Digital Pathology and Beyond, and for where this fits within the full spatial biology pipeline, Spatial Biology in Drug Discovery: From Target Discovery to Translational Medicine.

This article was produced under Drug Discovery News’s AI editorial policies.