Latent Structure Laboratory
Bacteria may not build cities or have interesting social lives, but they will be here when the Sun explodes.
-- Bill Bryson in A Short History of Nearly Everything
UW Madison Statistics Lightning Talks
Kris Sankaran | krisrs1128.github.io/LSLab | 16 September 2021
Problem Solving in Multi-Omics
Our lab focuses on statistical methods for analyzing multi-omics, and especially microbiome, data:
- We are motivated by interdisciplinary projects with microbiologists, ecologists, and psychologists
- We synthesize and benchmark statistical approaches to solve data analysis problems within these collaborations
Project 1: Multiscale Topic Alignment
- We were motivated by a project about bacterial vaginosis
- The literature collapses microbiomes into 5 “community state types,” but we noticed subtle transitions in participants with the disease
- To addressed this, we developed meta-algorithms for visualizing ensembles of topic models trained across several scales
Project 1: Multiscale Topic Alignment
- We found that properties of the alignment graph are useful for count data analysis more generally
- Can help quantify the degree of mixing across clusters
- Can help highlight mutually exclusive species
- We wrapped our algorithms in the R package
alto
Project 2: Meditation and the Microbiome
- The Center for Healthy Minds at UW - Madison has had a long history of studying the effects of meditation on well-being
- Motivated by increasing evidence of a gut-brain axis, they now also collect microbiome samples
Project 2: Meditation and the Microbiome
- We are supporting a longitudinal RCT that is gathering data from 3000 participants (for reference, my PhD thesis analyzed three subjects...)
- We have adapted concepts from causal mediation analysis to relate sequencing with psychological survey data
An approach to microbiome mediation analysis designed by Xinran Miao and Hanying Jiang.
Techniques
A few ideas recur throughout our research,
- Generative models: These allow us to encode latent structure and integrate across data sources and scales
- Data visualization: Good representations enable hypothesis generation, model criticism, and communication
- Reusable software: We release our methods as accessible R packages
Visualizations and code by Zhuoyan Xu and Tinghui Xu in [2 - 4].