AI compares tumour ‘floor plans’ to explore immunotherapy response

Two tumours can contain similar kinds of cells and arrange them very differently. Immune cells at an edge may occupy a different biological environment from cells dispersed through the tumour. University of Chicago researchers are investigating whether these spatial patterns—the tissue's ‘floor plans’—can help explain how cancer responds to treatment.[1]

The team developed statistical and AI tools using published spatial-transcriptomics data from 262 solid tumours, the university reported on 22 October. An exploratory test then examined 16 samples from patients with non-small cell lung cancer and known responses to immunotherapy. The university says the spatial approach predicted those responses better than the standard-of-care biomarker used for comparison.[1]

From a cell inventory to a neighbourhood map

Spatial transcriptomics retains information about where gene activity occurs in a tissue sample. That makes it possible to study both the molecular signals and their arrangement. Conventional tissue examination can also reveal architecture; the computational challenge here is to compare complex molecular and cellular patterns systematically across different tumours.[1]

Led by systems biologist Arjun Raman, the researchers looked for recurring cell communities that the university calls spatial groups. A group carries information about how neighbouring populations are organised together. That can add a different dimension to a description based on the overall abundance of each cell type.[1]

The tools represent those structures in a form that allows samples to be compared. A latent-space analysis places tumours near one another when their derived spatial characteristics are similar. Researchers can then ask whether this resemblance is associated with a shared clinical behaviour, even when the samples differ in their size or individual measurements.[1]

What the lung-cancer test adds

The collaboration with lung-cancer specialist Marina Garassino connects a mathematical description of tissue to an outcome that matters to patients. Instead of stopping at a visually interesting map, the researchers asked whether the map contained information about treatment response.[1]

Sixteen samples provide an early signal, with substantial uncertainty about how the approach will perform in a broader population. The university account does not provide a numerical accuracy estimate suitable for generalising to future patients. The next tests need larger and independent groups, clear separation of model development from evaluation, and comparisons with existing clinical predictors.[1]

Consistency across tissue preparation and measurement is another practical issue. A useful similarity measure should reflect tumour biology rather than differences between laboratories or data-processing methods. Establishing that robustness is part of turning a research representation into a dependable biomarker.[1]

A broader use for spatial biology

The group plans to explore ovarian-cancer response to chemotherapy. That extends the research question to a different disease and treatment, where the relevant patterns may differ. It remains a planned investigation rather than an additional validated result.[1]

The study offers a common language for comparing tissue organisation. Its appeal is that the arrangement of cells may explain something their individual identities cannot: which populations meet, which neighbourhoods form and how those environments relate to therapy. The initial lung-cancer result gives researchers a concrete reason to test that idea at a larger scale.[1]

Report an error or suggest a correction