I will highlight motivation and use cases for using recent advances in single cell and spatial transcriptomics and proteomics technologies for studying tumour microenvironments and their relation to treatment responses and disease outcomes. I will then present new approaches in machine learning and biostatistical modelling to exploit and biologically interpret these data. 

A fundamental task is dissecting differences in cell type composition and cell type-intrinsic gene expression differences in tissue samples from patient cohorts. A main challenge is the context-dependence and approximative nature of the notion of “cell type” itself. To address this, I will present latent embedding multivariate regression (LEMUR), a machine learning approach that integrates data from different samples, provides a predictive generative model for each cell’s gene expression in all conditions and all cell states, and identifies groups of cells with consistent differential expression patterns. 

In a third part of the talk, I will highlight recent research in deep-learning-based foundation models, which promise to learn representations of single-cell data that can predict the effects of genetic or chemical perturbations. I will report on benchmark efforts of recently published, well-advertised foundation models as well as deliberately simple baselines, and discuss the roles of objective statement and critical benchmarking in dissecting overpromises from real progress.