The Data Hiding in Plain Sight: Why Tissue Architecture Matters More Than We Teach

By Dr. Ruseik Rahumath · 26 June 2026 · Colombo, Sri Lanka

There is a quiet revolution happening in pathology, and it does not require a million-dollar sequencer.

For decades, we have trained medical students to identify cells. We teach them to recognize nuclear pleomorphism in cancer cells, to count mitotic figures, to spot invasion patterns at the tumour edge. These skills form the bedrock of histopathology diagnosis. But we rarely ask a deceptively simple question: how are these cells arranged relative to each other, and what does that arrangement mean?

The spatial organization of cells — their neighbourhoods, their clustering patterns, their relationships with stroma and immune infiltrates — carries information that is invisible when we look at individual cells in isolation. A growing body of evidence suggests that this spatial context may be as clinically significant as the cells themselves.

The Evidence Accumulates

In melanoma, the spatial arrangement of tumour-infiltrating lymphocytes determines immunotherapy response more reliably than cell counts alone (Orenes-Piñero et al., 2026, Front Immunol). Pham and colleagues (2026, J Immunother Cancer) demonstrated that spatial ecosystems — neighbourhoods of tumour cells, immune cells, and macrophages — outperformed gene expression signatures in predicting anti-PD1 response. Boe et al. (2026, Genome Biology) used spatial transcriptomics to show that melanoma cell fates under therapy are shaped by their microenvironmental niche, not intrinsic transcriptional programs.

Perhaps most strikingly, Zhang et al. (2026, Nature) developed spatial ecotypes — computational signatures inferred from tissue architecture — that predict immune infiltration and survival across multiple cancer types. The data was already on the H&E slide. Machine learning simply helped us read it.

What This Means for Sri Lanka

Platforms like 10x Genomics Visium and multiplexed ion beam imaging (MIBI) capture spatial gene expression at single-cell resolution, but they are expensive and largely inaccessible in LMIC settings. The temptation is to see this as a limitation. I see it differently.

When you cannot buy the expensive test, you look harder at what you already have. Every Sri Lankan pathology department has decades of H&E slides accumulating in archives. These are not just diagnostic records — they are untapped spatial datasets waiting for the right analytical lens. My research partner Pasindu and I are beginning to explore this territory, starting with colorectal cancer research and moving toward digital pathology workflows.

A key insight from the recent literature is that the most impactful spatial features may not require transcriptomics at all. Zormpas-Petridis et al. (2018) achieved 95.7% accuracy in classifying melanoma tissue regions using only H&E images with a hierarchical superpixel method — no molecular data needed. Their work established a principle: spatial context is not auxiliary information but a prerequisite for accurate classification.

The Methodological Landscape

The computational approaches to capturing spatial context fall into three broad families. Hierarchical models mirror how pathologists perceive tissue — region-level classification before cellular analysis. Graph-based methods (graph neural networks, graph transformers) model cells as nodes and their spatial relationships as edges, enabling neighbourhood-aware classification. Agent-based models simulate how cell-cell interactions at the local scale generate the macroscopic heterogeneity observed in clinical specimens (Jamshad & Jackson, 2026, Bull Math Biol).

All three families converge on the same insight: spatial information is not a feature to be extracted but a structural constraint on classification. Cell identity is context-dependent.

A Mindset Shift, Not a Technology Problem

The barrier is not the absence of technology. The barrier is how we train doctors to see. Medical education emphasizes cellular features — nuclear grade, mitotic index, cytoplasmic characteristics — while paying far less attention to tissue architecture. But the way cells stack together may tell a story clearer than an expensive molecular test.

This is an invitation to every young doctor and researcher in Sri Lanka and beyond. Collaborate across specialities. Look at your H&E slides differently. Ask not just "what cell is this?" but "how is it arranged relative to its neighbours?" The answer might already be there, hiding in plain sight.

References

1. Zormpas-Petridis et al. (2018). Capturing global spatial context for accurate cell classification in skin cancer histology. arXiv:1808.02355.
2. Pham et al. (2026). Melanoma cell states shape spatial tumor-immune ecosystems. J Immunother Cancer. PMID: 42315253.
3. Boe et al. (2026). Spatial transcriptomics reveals influence of microenvironment on intrinsic fates in melanoma therapy resistance. Genome Biol. PMID: 42174697.
4. Zhang et al. (2026). Non-invasive profiling of the tumour microenvironment with spatial ecotypes. Nature. PMID: 42092150.
5. Orenes-Piñero et al. (2026). Immunoscore and beyond. Front Immunol. PMID: 42183189.
6. Dong et al. (2026). Single-cell profiling uncovers a hypoxia-vascular-immune axis. J Transl Med. PMID: 42098765.
7. Voglis et al. (2026). Spatially-resolved single-cell imaging of melanoma brain metastases. Neuro Oncol. PMID: 42209753.
8. Jamshad & Jackson (2026). A Hallmark-Integrated, Agent-Based Framework. Bull Math Biol. PMID: 42118484.
9. Long et al. (2026). Transcriptome graph transformer. BMC Bioinformatics. PMID: 42168843.
10. Nonchev et al. (2026). Representation learning for multi-modal spatially resolved transcriptomics. Bioinformatics. PMID: 42179160.