AI for glioblastoma hackathon - a deep dive into the winning solutions

In a groundbreaking two-day hackathon held at Paris's Palais Brongniart on Feb 3-4, Owkin and Servier brought together 130 global computational and translational research experts to tackle one of the most aggressive forms of brain cancer.

Glioblastoma is highly complex and heterogeneous—scientists and doctors still struggle to fully understand this disease’s biology. Despite more than 400 clinical trials conducted since 2005, treatment advances have remained limited. However, today's combination of advanced datasets and AI tools presents a historic opportunity for breakthrough discoveries.

At the heart of this hackathon lay exclusive access to two extraordinary datasets. Participants worked with Owkin's MOSAIC initiative data, comprising a multi-omics atlas for glioblastoma that includes spatial omics data generated using best-in-class spatial omics technology from 10X Genomics. Complementing this, the Parker Institute for Cancer Immunotherapy (PICI) has provided their cutting-edge spatial proteomics data from the BRUCE cohort. The two datasets combined offer unprecedented insights into the tumor microenvironment of glioblastoma patients.

Through a partnership with Bioptimus, participants had access to cutting-edge foundation model technology. Completing this exceptional ecosystem, Amazon Web Services (AWS) provided best-in-class infrastructure and computational power to drive this multidisciplinary effort, positioning this hackathon as a catalyst for transformative innovations in glioblastoma research.

Here's a deep dive into the three winning projects that emerged from this collaborative effort.

Clinical Award Winner: Team Glioblasters - Cellular niches in GBM

Team Glioblasters earned the Clinical Award for their innovative approach to analyzing spatial omics data. Their project focused on understanding the complex tumor microenvironment (TME) in glioblastoma, specifically investigating how different cell types interact within the tumor landscape.

The team's methodology involved utilizing spatial transcriptomics data to identify and analyze recurring cellular niches within glioblastoma tumors. By deconvoluting the spatial transcriptomics and single-cell data from the MOSAIC cohort, they identified distinct niches dominated by different cell types - some primarily consisting of tumor cells, others of immune cells, and some showing a mixed composition. The team then focused on the mixed, interface niches - areas where tumor cells meet the TME, first validating them using spatial proteomics data from the PICI cohort.

Over the next three months the team will focus on uncovering the prognostic relevance of these interface niches to identify those which contribute to the worst prognosis, and using sequencing data to dive deeper into the biology of tumour TME interactions. Through a preliminary receptor-ligand interaction analysis using single-cell RNA-Seq data they were able to reveal interesting insights into a potential role of a particular chemokine signaling pathway, which showed potential prognostic significance.

The team's future goals include developing predictive models and deeper investigation of tumor-TME interactions, with the ultimate aim of translating their findings into applicable biomarkers and candidate targets.

AI Methodology Award Winner: Team J - Integrating multimodal data with foundation models

Team J secured the AI Award for their innovative approach to integrating multiple data modalities using foundation models. Their project aimed to characterize patients in a newly created multimodal embedding space, where patients in the same group will likely share similar disease biology and response to treatments. They then propose to characterize these subgroups further using classical bioinformatic analysis, machine learning methods for imaging and single-cell RNA-sequencing analsyis, and tools like CellOT for prediction of drug response.

The team proposed an architecture that was inspired by the recently published THREADS architecture [Vaidya et al, 2025], employing different models for each data type: Novaer and Phikon-v2 foundation models for H&E images and spatial transcriptomics, scGPT transformer encoder architecture for single-cell and bulk RNA-sequencing data , and custom preprocessing steps for clinical data and whole-exome sequencing. The idea is that this multi-modal approach will allow them integrate these diverse data spaces into unified embeddings and to extract robust features from each data type.

The team suggested that using this model they can go beyond simple cluster analysis, using the representations learned to enable more sophisticated predictions for patient survival, tumor growth, and recurrence classification. The team's framework showed particular promise in its ability to create a comprehensive glioblastoma map that could serve both diagnostic and therapeutic purposes.

Special Jury Award Winner: Team SpaceAix - Combining multiple approaches

Team SpaceAix impressed the jury with their comprehensive approach that combined prior knowledge, multimodal data, and complementary AI tools. Their methodology involved several sophisticated techniques, including MOFA, differential expression analysis, optimal transport, and transformer models. Using these methods they were able to identify clinically relevant factors, factors that were relevant for patient stratification, features that were relevant for disease progression, and preliminary target and biomarker ideas.

The team developed a robust model initially trained on histopathology slides, which successfully stratified patients into high and low-risk categories with high accuracy (C-index = 0.69). They further refined this model to work efficiently with just 100 random tiles, maintaining similar accuracy levels (C-index = 0.64) while reducing computational overhead. Further analysis from the team revealed correlations between poor prognosis and specific clinical parameters, such as necrosis and the number of tumor sites.

By integrating single-cell and bulk RNA data with multiomic analysis, they identified 40 distinct cell states. They then used this analysis and an optimal transport model to create patient trajectories that showed the progression from low to high-risk states, and identify cell types and genes that were associated with this transition.

This comprehensive approach allowed them to identify potential biomarkers and therapeutic targets.

Looking forward

These winning projects demonstrate the power of combining diverse expertise and advanced technologies in the fight against difficult-to-understand disease, such as glioblastoma. Each team brought unique insights and methodologies, from spatial analysis of tumor microenvironment to sophisticated AI methodology and multi-omic data integration.

With the continued support of Owkin, Servier, Amazon Web Services (AWS), the Parker Institute for Cancer Immunotherapy, over the next three months these teams will have the opportunity to further develop their promising approaches. Their work represents significant steps forward in our understanding of glioblastoma and could potentially lead to improved treatment strategies for patients.