Mots-Clés
Deep learning
Cancer
Imaging
Immunology
Description
SCREAM: Learning Spatial Cell Representations with Explainable and Adaptive Models
Project Overview
This project is part of the SCREAM efforts to integrate and model spatial proteomics data from diverse experiments using explainable deep learning (DL) models. Modern spatial imaging platforms such as MIBI, IMC, and CODEX allow the simultaneous measurement of dozens of proteins directly in tissue samples. However, each experiment typically targets a different set of proteins (known as markers), depending on the biological question and available reagents. As a result, datasets often share only a subset of markers, making direct comparisons difficult.
SCREAM addresses this challenge by making use of prior biological knowledge to learn robust cell-level representations that capture biological pathway activity and spatial context, even when the set of observed markers differs across experiments. You will implement a Vision Transformer (ViT) model, a popular neural network architecture accounting for spatially distributed motifs, to encode cell images, and benchmark its performance against existing approaches. This work will help improve tissue comparison across datasets and support applications such as disease classification and spatial imputation.
Student Tasks
- Implement and adapt a Vision Transformer to represent cells in multiplexed imaging data
- Benchmark ViT models against convolutional and hybrid approaches using classification
and clustering metrics
- Evaluate robustness to variation in marker panels, spatial context and image
preprocessing
- Analyze the interpretability of learned representations and compare them across
conditions
Learning Opportunities
- Apply cutting-edge ViT models to spatial single-cell data
- Gain hands-on experience in DL model training, validation, and benchmarking
- Learn to work with complex, high-dimensional biomedical imaging datasets
- Contribute to the development of interpretable tools for integrative spatial biology
Students should be familiar with Python and have basic knowledge of deep learning (e.g., TensorFlow). Interest in molecular cancer biology or spatial omics is welcome. Theses and internships of 4-6 months are preferred. We welcome students from interdisciplinary backgrounds eager to bridge machine learning and biomedical research.
Hosts: Systems Immunology and Single Cell Biology group at the German Cancer Research Center (DKFZ). Please send your application (including CV and cover letter) to felix.hartmann@dkfz-heidelberg.de and loan.vulliard@dkfz-heidelberg.de