Faire le lien entre morphologie cellulaire et transcriptomique grâce à l’apprentissage automatique.

Cells are the fundamental units of life, and their function are intrinsically linked to their spatial organization, a piece of information lost in standard genomic, transcriptomic, or proteomic profiling. Spatial biology aims to study this organization with Spatial Transcriptomics (ST) technologies measuring gene expression while preserving spatial context. These methods typically represent a trade-off between spatial resolution and transcriptome coverage. Sequencing-based ST offers transcriptome-wide coverage for multi-cellular spots, whereas imaging-based ST provides sub-cellular resolution for a limited panel of genes. In parallel, Hematoxylin and Eosin (H&E) histology remains the cornerstone of clinical pathology, offering rich morphological details at a massive scale and low cost. This PhD thesis presents machine learning methods to bridge these modalities, inferring single-cell gene expression directly from nuclear morphology in routine H&E slides using ST as supervision.Our first project addressed the challenge of sequencing-based ST by developing a Multiple Instance Learning framework tailored to the multi-cellular nature of the data spots. Next, we leveraged the precision of imaging-based ST, which enables the pairing of individual nuclear images with their transcriptomic profiles after cell segmentation. To investigate the predictive power of spatial context, we represented tissues as cellular graphs based on spatial coordinates. Finally, to make cell-level morphological features broadly reusable, we developed LEMON (Learning Embeddings from Morphology Of Nuclei), a single-cell foundation model for pathology. Trained with self-supervised learning on millions of nuclei spanning multiple tissues and cancer types, LEMON produces robust and transferable morphology embeddings.Together, these contributions allows to link high-resolution spatial genomics with widely available digital pathology. They establish that nuclear morphology and the local cellular neighborhood are powerful predictors of transcriptional state. The resulting tools provide a practical and scalable framework to map gene expression at single-cell resolution from routine H&E slides, promising to accelerate research in tissue organization, cancer biology, and precision medicine.