Speaker
Description
Spatial transcriptomics is an emerging field enabling the study of gene expression within its spatial tissue context, offering critical insights into cellular function and organization. Incorporating spatial information is essential for accurate biological hypothesis generation, as it provides context often lost in traditional single-cell analyses. A key challenge in this area is the inference of underlying biological processes, which is central to many research domains. However, current deep learning-based methods often sacrifice interpretability, making it difficult to extract meaningful biological conclusions.
To address this, we propose a nested two-layer Deep Gaussian Process [1] model for the estimation of a latent time variable that captures the progression of a biological process. The first layer learns a latent representation that is explicitly informed by spatial location, embedding spatial correlation directly into the model. The second layer models gene expression as a function of this inferred latent time, which guides the latent time to produce temporally smooth gene expression changes, as observed in many biological processes.
This model estimates a latent time that characterizes an underlying biological process, directly informed by spatial context. By capturing smooth gene expression changes along this inferred trajectory, the model provides an interpretable representation of spatial biological dynamics. This interpretability supports the generation of meaningful hypotheses across a range of biological and medical research fields.
[1] Damianou, A., & Lawrence, N. D. (2013, April). Deep gaussian processes. In Artificial intelligence and statistics (pp. 207-215). PMLR.