Intra-tumor heterogeneity (ITH) refers to the fact that not all cancer cells within a single tumor are genetically identical. Over time, cancer cells evolve and develop into distinct subgroups called subclones - each with slightly different genetic makeups. This diversity makes it harder to treat cancer effectively, because a drug that kills one subclone may leave others unharmed.
In kidney cancer - particularly clear cell renal cell carcinoma (ccRCC) - this genetic heterogeneity is especially pronounced. Key cancer-driving genes like BAP1, PBRM1, and SETD2 are often mutated in only some regions of the tumor, not all of it. This means a biopsy from one area may give an incomplete picture of the whole tumor.
The best way to study this diversity has been multi-regional sequencing - taking DNA samples from multiple areas of the tumor to map its genetic landscape. However, this approach is expensive, labor-intensive, and difficult to scale to routine clinical care.
This study explored a more practical alternative: using the physical appearance of tumor tissue under a microscope - its morphology - as a proxy for underlying genetic diversity. Because visible tissue features reflect the molecular state of cells, differences in appearance across a tumor may signal differences in genetic makeup.
MorphoITH is a new computational framework developed to measure morphological diversity in tumor tissue using standard H&E-stained slides - the same pathology slides already routinely used in cancer diagnosis. By analyzing how tissue appearance varies across a tumor, it aims to infer underlying genetic heterogeneity without requiring additional testing.
The system uses a type of AI called self-supervised deep learning to generate a mathematical summary of tissue appearance for each small patch of a slide. These summaries capture multiple features simultaneously - cell size and shape, how cells are organized into structures, and the surrounding microenvironment - without being programmed with explicit rules about what to look for.
A key challenge in this kind of analysis is separating real biological differences from misleading patterns - for example, two tissue areas can look similar just because they are physically close together, not because they are genetically similar. MorphoITH includes rigorous statistical controls to account for this spatial autocorrelation.
The system was developed and validated using several datasets from kidney cancer patients, including tissue microarrays, whole-slide images, and multi-region sequencing data, allowing comprehensive testing of its ability to detect and interpret tumor heterogeneity.
The AI at the core of MorphoITH was trained on tissue microarray (TMA) data from 421 kidney cancer patients, comprising 785 tissue cores across 10 TMA blocks. The training used a self-supervised strategy: the model learned that tissue patches from the same TMA core should be recognized as similar to each other, without needing human-labeled examples of what 'similar' means biologically.
Several different AI architectures were tested during development, including BYOL, Triplet, and MoCov2 approaches, with different backbone networks (Vision Transformer and ResNet50). The final selected model used a BYOL training strategy with a Vision Transformer encoder, which produced the best results on a benchmark task measuring how often a tissue patch's nearest neighbors came from the same TMA core.
To apply the model to full tumor slides, a sliding window scanned each slide and generated a 1,000-dimensional feature vector for every tissue patch. A separate classifier was used to distinguish tumor from non-tumor tissue, ensuring that only relevant regions were included in subsequent analyses.
To measure how well MorphoITH distinguished between biologically different tissue areas, a separability score was calculated using a linear classifier. Two control baselines - one preserving spatial structure, one scrambled - were used to interpret whether the observed differences were statistically meaningful or simply due to physical proximity of similar-looking tissue.
The first validation test confirmed that MorphoITH could reliably distinguish tumor tissue from non-tumor tissue - correctly separating these two regions in 92% of 444 kidney cancer slides from The Cancer Genome Atlas. In 7% of cases, separation was partial, often because tumor and non-tumor tissue were physically intermixed.
MorphoITH was then tested on clinically recognized tissue features within ccRCC. It successfully separated three classes of vascular architecture - patterns of blood vessel organization that carry prognostic significance - in 95% of cases. Low-grade and high-grade tumor regions, classified by nuclear size and nucleolar prominence, were also fully distinguishable in nearly all tested slides.
Importantly, MorphoITH was not specifically programmed to recognize these features - it learned them implicitly. When compared to traditional hand-crafted measurements (nuclear size, vascular density, eosin intensity), MorphoITH identified similar tissue areas as similar, but captured additional variation that single measurements missed. It effectively combined multiple morphological signals into a single unified similarity score.
A color-coded visualization tool was developed alongside MorphoITH, mapping the AI's feature outputs onto slide images as pseudo-colors so that regions with similar morphology appear in similar shades. This makes the patterns immediately visible to pathologists and researchers reviewing the images.
A key test was whether MorphoITH could detect genetic differences within a single tumor slide. Using matched slides stained for three important kidney cancer genes - BAP1, SETD2, and PBRM1 - researchers compared regions where these genes were functioning normally to regions where they had been lost.
MorphoITH showed full morphological separation between normal and loss regions in 58% of cases, partial separation in 33%, and no separation in 8%. The strongest morphological signals were associated with BAP1 loss, followed by SETD2 and PBRM1 - likely reflecting the degree to which each gene's loss visibly changes how tumor cells look and behave.
When the AI divided tumor slides into 10 spatially defined morphological clusters, these clusters largely respected the genetic boundaries between normal and gene-loss regions - meaning a morphological cluster rarely spanned both a normal and a mutant area. This finding supports the idea that morphologically distinct regions correspond to genetically distinct subclones.
A per-slide heterogeneity score - measuring how much morphological diversity exists within a single tumor slide - increased with higher nuclear grade and metastatic stage. In survival analysis, patients with high morphological heterogeneity scores had significantly worse progression-free survival, and this score was an independent predictor even after accounting for nuclear grade.
A practical application of MorphoITH is to guide where tissue should be sampled from a tumor for genetic testing. Currently, sampling is often done at a single location or randomly - which may miss genetically distinct subclones in other areas of the tumor.
The researchers tested a morphology-guided sampling strategy: first identify the most morphologically distinct regions of the tumor (using MorphoITH clusters), then select one sample from each. This approach more than doubled the likelihood of capturing both a normal and a gene-loss region compared to random sampling, in the case of BAP1 and SETD2 genes.
This finding has important practical implications. If tumor biopsies are guided by morphological diversity rather than chosen arbitrarily, genetic testing would be more likely to uncover the full spectrum of mutations in a patient's tumor - providing a more complete picture for treatment decisions.
This strategy could be particularly valuable in clinical trials and research studies where understanding a tumor's full genetic complexity is critical, and could eventually improve how biopsies are planned for routine clinical care of kidney cancer patients.
In a deeper investigation, the researchers analyzed three kidney cancer patients who had multi-region whole-exome sequencing data from 26 tumor samples matched to their tissue slides. This allowed direct comparison between the genetic family tree of tumor subclones (a phylogenetic tree) and morphological similarity patterns detected by MorphoITH.
Across all three patients, morphological similarity tracked closely with genetic relatedness. Tumor regions that were genetically similar - sharing more mutations - also tended to look more similar under the microscope. The global correlation between morphological distance and genetic distance was r=0.56 (p less than 0.0001), a strong relationship given the inherently noisy nature of both measurements.
The analysis also revealed interesting exceptions. In a few cases, two samples had very similar morphology despite being genetically quite different - a pattern consistent with convergent evolution, where different genetic paths lead to the same biological outcome. In other cases, a small number of mutations caused dramatic morphological changes, particularly involving the TSC1 gene.
These findings suggest that as cancer evolves and subclones diverge genetically, their visible appearance generally changes in parallel - but significant genetic events can accelerate or redirect this process. MorphoITH captures this relationship without requiring genetic data, making it a powerful standalone tool for studying tumor evolution.
MorphoITH represents a significant step toward making the study of tumor heterogeneity more accessible. Because it works on standard H&E slides that are already collected routinely in clinical care, it does not require additional tissue, special staining protocols, or expensive genetic sequencing to provide information about tumor diversity.
The framework is explicitly unbiased - it does not require researchers to know in advance what to look for. Unlike earlier computational tools that were trained to detect specific mutations or tissue types, MorphoITH learns morphological similarity in an open-ended way, capturing variation across many biological dimensions simultaneously.
Importantly, the core findings held up even when different AI encoders were used in place of the custom MorphoITH model, including large general-purpose pathology foundation models. This robustness suggests the approach is not dependent on any particular technical implementation, making it easier to adopt and adapt in different research settings.
Looking ahead, MorphoITH could serve as a foundation for understanding tumor aggressiveness, guiding biopsy strategies, and informing treatment planning in kidney cancer and potentially other cancer types. Connecting visible tissue patterns to underlying genetic complexity may ultimately help clinicians deliver more precise and personalized care to their patients.