Compete and Collaborate with Top Minds to Decode Tissue Biology using AI
%201.svg){kind=icon}
The primary objective of this challenge is to develop AI models that can accurately predict the spatial distribution of biological cell types in a tissue slice using routine H&E slides.
This will enable deeper insights into tissue biology, bridging the gap between traditional histology and advanced spatial transcriptomics.
Develop a computational pipeline to predict the spatially resolved cell-type composition of a tissue sample using only its H&E slide image. Given a histology image with a grid of spots, participants must build models to estimate relative cell-type proportions for each spot. Training data includes high-resolution images and paired ground truth labels, while evaluation is based on mean squared error (MSE) against a hold-out test slide. Performance will be tracked on a public leaderboard (ongoing rankings) and a private leaderboard (final rankings revealed post-challenge). This challenge pushes the frontiers of AI-driven digital pathology and spatial biology.
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:
Predicting all measured genes accurately remains challenging, potentially distorting clinical interpretation. Many genes may not be clinically relevant, and the inherent noise in single-cell sequencing and SRT data limits fair model assessment against an imperfect ‘ground truth.’
Predicting cell composition offers deeper biological insights, such as immune cell distribution in the TME or brain organization. By focusing on higher-level cell-type signals from images rather than noisy gene-level data, we trade granularity for more robust spatial pattern prediction.
Open to all individuals!
Participants can work in diverse teams of up to 10 experts.
Examples of relevant profiles:
- Researcher
- AI/ML engineer
- Data scientist
- Computational biologist
- Bioinformatics Scientist
- Others
Participants will develop a computational pipeline that takes an H&E-stained histology image as input and outputs the estimated cell-type composition for each spot on a grid overlaid on the image. Training data will consist of high-resolution images paired with corresponding "ground truth" labels indicating the actual cell-type proportions at each spot.
Model performance will be evaluated using Mean Squared Error (MSE) between the predicted and actual cell-type proportions on a held-out test set. Throughout the challenge, participants can track their progress on a public leaderboard. Final rankings, however, will be determined based on a private leaderboard revealed after the challenge concludes. This challenge aims to advance the fields of AI-driven digital pathology and spatial biology.
Traditional histology is crucial for disease diagnosis but limited by subjective interpretation and time-intensive expert analysis. Spatial transcriptomics (SRT) gives a clearer picture of tissue biology, helping identify biomarkers for treatment and prognosis, but its clinical use is limited by cost and logistics. Machine learning bridges this gap by predicting molecular markers from histology images, but challenges remain in achieving clinical-grade accuracy. Despite this potential, significant challenges remain in achieving the necessary accuracy and clinical relevance.
This challenge brings experts together to advance AI-driven digital pathology for better histology and spatial transcriptomics predictions.
We have hosted similar challenges before! Check out our previous hackathon: MLRW 2022: AI-Driven Biomedical Hackathon.
El-Hackathon 2025 is a competition where participants develop AI models to predict clinically relevant outcomes from H&E whole-slide images (WSIs).
Simply click Register Now button and follow the instructions to sign up.
.svg)
.svg)
.svg)
%201%20(1).svg)
%20(1).svg)
%201%20(1).svg)
.svg)
%201%20(1).svg){kind=icon}
