Speaker: 熊健康博士，湖南师范大学

Title: Structured Priors for Variational Inference in Single-Cell Lineage and Spatial Omics Data

Language: Chinese

Time & Venue: 2025.12.10 09:30-10:30 ZOOM ID: 867 3407 5806 (Passcode: 541764)

Abstract: Single-cell lineage tracing and spatial omics generate high-dimensional observations together with rich tree- or graph-structured information. I will discuss how to incorporate such structure directly into the prior of variational autoencoder–type models, leading to mathematically transparent generative formulations that still admit tractable posterior inference. For lineage data, we place a Gaussian process prior on latent variables with covariance kernels defined by lineage distances and develop scalable posterior variational inference under this structure-aware prior. For spatial omics data, we combine a Gaussian prior on a base latent with graph-based priors on spatial latents defined over tissue neighborhood graphs, so that spatial organization is incorporated at the prior level of the generative model while the resulting variational posterior inference retains a simple and tractable form. I will emphasize the common probabilistic structure behind these models and its implications for structured variational inference on trees and graphs.