## Introduction/Background Histology, a century-old technique that uses dyes to reveal the structure of biological tissue under the microscope, has been the cornerstone for clinical diagnosis and characterization of diseases like cancer. While histology images of tissue samples can be routinely obtained in the lab, their accurate interpretation relies on clinical expertise of the pathologist, is time-consuming and subjective, and allows for only coarse-grained interpretations based on visually detectable features. On the other hand, technologies that can tag and measure the levels of molecules in a tissue sample enable a much more detailed characterization of tissue biology, and are capable of revealing clinical biomarkers informing disease prognosis as well as diagnosis. Such techniques have undergone a progressive evolution, from immunohistochemistry that can help visualize a few targeted features (protein markers) at a time, to more advanced genomic deep sequencing technologies that can simultaneously capture the abundances of tens of hundreds of molecular features. The recent advent of spatially resolved sequencing like spatial transcriptomics (SRT) has further expanded the possibilities. Today, it is possible to capture spatial patterns of expression of large numbers of genes at single-cell resolution, offering a highly nuanced view into the spatial cellular architecture of healthy tissue and its distortion in disease. For instance, the spatial cellular organization of the immune microenvironment of a tumor (TME) as revealed by SRT can inform fundamental questions of clinical significance, such as the severity of the disease and its expected progression in the patient, and what features might correlate with the effectiveness of a particular treatment. Despite its promise, the widespread adoption of genomic deep sequencing in the clinical setting remains challenging, given the relatively high costs, and practical difficulties in preparing and preserving tissue samples obtained from a patient for the sequencing assay, especially in low-resource settings. This has spurred much work over the past decade to effectively leverage machine learning to bridge the gap between histology images, which are easier to obtain and share, and the granular markers revealed by high-end molecular sequencing. Although there has been substantial progress in computationally predicting molecular markers from HE images in bulk tissue samples, the limited spatial resolution of this data obscures the information in the underlying spatial organization that could support much more detailed characterization of the disease phenotype. As such, recent work combining histology imaging with spatially resolved transcriptomics to build predictive models, endeavors to push the envelope even further. Despite the increasing availability of paired histology and SRT data for training models, the field faces significant challenges in making these predictive approaches more accurate, clinically meaningful, and practically usable to have a shot at real-world adoption (e.g.[ this Review](https://www.nature.com/articles/s41467-025-56618-y.epdf?sharing_token=Lgqxbmk2Lub9_hRV9NHLENRgN0jAjWel9jnR3ZoTv0NtQlPNfoXX1H7ym5LeSJOBDfgwzmg3rtHFuPqGL5fHLpUEI_aPhN3ID5SEouoTsiY921lq3osQEHBYMJWaiwDnLjwCU9y5njHdSeAhEXH7XSON8eGeJXfYfD6YmMB6ErQ%3D)). We present this Challenge with the intent of highlighting this fundamental problem at the cutting edge of digital pathology to a wider community, and we hope that it will catalyze novel advances in AI-driven generation of insights from complex multimodal biological data. ## Why Cell Composition: Although most existing approaches in the scientific literature tackle the prediction of individual genes' spatial expression profile from histology image features, this Challenge invites submissions to predict the spatial pattern of cell composition. This is a deliberate choice and our rationale is two-fold: 1. First, it remains a challenge for models to make consistently accurate predictions over all measured genes. This can distort the interpretation of models from a clinical utility perspective. Not all genes are likely to be of clinical relevance in a particular setting. Moreover, the inherently noisy nature of single-cell sequencing data and SRT measurements imposes a limitation on fair assessment of models by comparison against this imperfect ‘ground truth’. 2. The prediction of cell composition presents a more meaningful goal, and differences in cell composition can by itself reveal interesting insights into tissue biology and disease-associated changes (for e.g., the spatial distribution of immune cell types in the TME, or the diversified organization of cell types across brain layers, can be very informative). By focusing on these higher-level aggregated biological signals (represented by cell types) revealed by the image inputs, rather than the lower-level noisier measurement signals at the gene-level, we are trading off granularity while still requiring that the model be capable of predicting robust spatially resolved patterns of biological interest. ## Core Challenge ### Design a computational pipeline to predict the spatially resolved cell-type composition of a tissue sample, given its Hematoxylin-and-Eosin-stained (HE) slide image. You are provided the positional coordinates of an array of ~2000 spots on a two-dimensional tissue slice from a tumor sample, accompanied by the corresponding aligned HE slide image. For each spot (~55 um diameter), you will be predicting the abundances of a set of predefined cell types, given the image information alone. Teams are expected to develop predictive models for this purpose. We provide training data in the form of high-resolution histology images for 6 slides (tissue samples) along with the paired ground truth labels for the cellular composition of ~10k spots with ~55 um resolution. These ground truth labels represent the abundances of 35 biologically distinct cell types at every spot, and were assigned by a computational algorithm. The cell types have been inferred from the corresponding 10x Visium spatial transcriptomics data for the tissue sections. Given its relatively low resolution, each Visium spot can be associated with ~1-10 cells, requiring the application of deconvolution techniques to fine-map the composition of cell types at every spot. To encourage generalizable approaches that make use of the image features alone, the tissue identity and disease subtype of the samples are not disclosed. The original sample identifiers and the original cell type names have been masked (sample ids are replaced by generic labels S1-S7 and cell names by C1, C2,… C35). Entries will be evaluated on a hold-out test slide, the data for which comprises the positional coordinates of a grid of spots on a two-dimensional tissue slice, accompanied by the corresponding whole-slide high-resolution HE color image. The corresponding ground truth (spot-level cell type composition information) is withheld until the end of the Challenge deadline. Submissions will be compared in terms of their cell composition predictions against the ground truth cell composition across all the spots marked on the test slide. This will happen on two levels - a public leaderboard available through the course of the Challenge wherein teams will be ranked by the performance on a public test set, and a final private leaderboard based on a private test set, which will be released upon the completion of the Challenge.