Genomics and morphology: Are we there yet?
Generative deep learning models for the joint learning of morphology and genomic traits aim to find maximum correlations between different non-linear data modalities, while allowing reconstruction in sparse or expensive regimes. In this talk, we explore benchmarks for understanding whether these models are interpretable and informative from a biophysics point of view and propose extensions that improve performance when data sources are correlated.
Emergence of division of labor in tissues through cell interactions and spatial cues
Single-cell gene expression data reveals the diversity within and between cells of different types. Often, cells of the same type span a variety of expression profile structures, including continua, but the origin of such patterns is usually unknown. In this talk I will discuss how we use a theoretical approach to predict the patterns that emerge in expression and spatial space in the presence of spatial gradients and cell-cell interactions in the tissue.
Improved marker detection through label refinement in case-control single-cell RNA-seq studies
While scRNA-seq technologies have allowed us to measure gene expression at a single cell resolution, in case-control studies the condition label of each cell is assigned at the coarser sample level. Therefore, all cells dissected from case samples are labeled as affected by the condition, while in reality, some of them may be indistinguishable from control cells. In this talk we will introduce a novel method that addresses this problem and demonstrate how standard analysis pipelines can fail to discover case-specific subtypes and their respective marker genes.