Advancing Spatial Transcriptomics with DeepSpot
Kalin’s approach is built upon DeepSpot, a deep learning method developed at ETH Zürich, Department of Computer Science (D-INFK). By extending DeepSpot’s capabilities to support 10x Genomics Xenium data, he improved the accuracy of single-cell gene expression predictions in patients with Inflammatory Bowel Disease (IBD). The competition provided an opportunity to independently validate DeepSpot’s performance and compare different approaches in the field.
The challenge brought together teams from around the world to tackle the complex task of predicting transcriptomic profiles from histopathology images. Kalin’s solution ranked at the top, demonstrating the effectiveness of his methodology.
DeepSpot has been developed to enhance the prediction of single-cell spatial transcriptomics from H&E-stained histology images, leveraging deep learning advancements. By integrating spatial tissue context and leveraging recent advancements in pathology foundation models, DeepSpot significantly improved gene correlation across multiple datasets from patients with metastatic melanoma, kidney, lung, or colon cancers. Notably, DeepSpot’s application led to the generation of 1,792 spatial transcriptomics samples, encompassing 37 million spots, from The Cancer Genome Atlas (TCGA) cohorts of melanoma and renal cell carcinoma.
This advancement addresses the resource-intensive nature of current spatial transcriptomics technologies, which limits their routine adoption in patient care. The researchers anticipate that DeepSpot and the accompanying dataset will serve as valuable resources for biological discovery and as benchmarks for evaluating spatial transcriptomics models, potentially stimulating further advancements in computational spatial transcriptomics analysis.
A Collaborative Effort
The DeepSpot project is the result of strong collaborative excellence between multiple research groups:
- Our team at ETH Zürich (Prof. Gunnar Rätsch)
- Computational and Translational Pathology Lab at the University of Zurich and University of Basel (Prof. Viktor H. Koelzer)
- Silina Group at the Institute of Pharmaceutical Sciences, ETH Zürich (Prof. Karīna Siliņa)
This interdisciplinary approach has been key to refining machine learning applications in pathology and biomedical research.
Acknowledging the Organizers and the Broader Research Community
We appreciate the efforts of the Broad Institute of MIT and Harvard, the Eric and Wendy Schmidt Center, CrunchDAO, MIT EECS, the Laboratory for Innovation Science at Harvard, MIT Institute for Data, Systems, and Society (IDSS), and Mass General Hospital for organizing this challenge and providing the necessary infrastructure.
Learn More
- Learn more about Kalin’s solution and DeepSpot: DeepSpot (medRxiV)
- Explore the challenge details: Broad Institute Challenge
- View the leaderboard: Competition Results
We celebrate Kalin’s success as a milestone in applying machine learning to spatial transcriptomics and look forward to further advancements in the field!
