Bridging cells and space: refining Tangram through neighborhood-informed mapping

Abstract

Treball de fi de grau en Bioinformàtica. Curs 2024-2025

Tutors: Richard Rötger, Tomás Bordoy, Lucas Torp Dyssel

Spatial transcriptomics (ST) overcomes the spatial limitations of single-cell RNA sequencing but faces resolution trade-offs. Integrating ST with single-cell (SC) data leverages their complementary strengths, yet current mapping methods like Tangram lack spatial context and exhibit instability. This project analyzes Tangram’s limitations such as mapping inconsistency, bias toward abundant SC reference cell types, and lack of spatial neighborhoods and develops a post-processing framework to enhance biological plausibility. I introduce an Expectation-Maximization (EM) framework that refines Tangram’s probabilistic cell-to-voxel mappings by integrating gene expression similarity and spatial neighborhood information. Evaluated against voxelized MERFISH ground truth, the method improves precision (mean: 0.56) but shows conservative recall (mean: 0.46), highlighting a trade-off between spatial coherence and cell recovery. Our work demonstrates that spatial context refines integration and provides a foundation for more complex neighbor-aware mapping tools.