Mathematics Colloquium
Speaker: Sagnik Nandy, Ohio State University
Host: Parthanil Roy
Title: Data Integration in Single-Cell Genomics: Message Passing for Multimodal Fusion
Time, day and date: 4:00:00 PM - 5:00:00 PM, Wednesday, July 29
Venue: Ramanujan Hall
Abstract: Modern single-cell technologies increasingly measure multiple modalities — transcriptome, epigenome, proteome — on the same cell, but each view remains noisy and heterogenous. In this talk, we shall explore different ways to combine information in such multimodal datasets. In particular, we shall develop a message passing based framework for principled multimodal fusion, built on empirical Bayes denoising: shrinking noisy estimates toward a prior learned directly from the data.
I will first present OrchAMP, an Approximate Message Passing-based method that jointly denoises multiple modalities by estimating a shared prior on cell signatures, achieving Bayes-optimal signal recovery without collapsing weak-signal modalities the way naive joint SVD does. I will then discuss DAIF, which extends this framework to supervised prediction by adaptively choosing the fusion granularity — bridging early and late fusion in a data-driven way — and show on TEA-seq and TCGA-BRCA data that this adaptivity translates into real predictive gains over existing pipelines (MOFA+, Multigrate, JAFAR, Cooperative Regression).
This talk is based on joint works with Zongming Ma, Samriddha Lahiry, Subhabrata Sen and Pragya Sur.