Thursday, Aug 7: 10:30 AM - 12:20 PM
0195
Invited Paper Session
Music City Center
Room: CC-104A
Microbiome
AI
Virome
Multi-omics
Applied
Yes
Main Sponsor
ENAR
Co Sponsors
Section on Statistics in Genomics and Genetics
WNAR
Presentations
Phage, their bacterial hosts and phage–bacteria interactions strongly influence global biogeochemical cycles, incidence of human diseases, and patterns of microbial genome diversity. It is important to characterize such interaction patterns and underlying processes using the empirical data. We present several latent variant models to study phage–bacterial interactions as networks rather than as coupled interactions in isolation. We show that such models fit empirical data of phage-bacterial interactions well. Such models provide computational tools to connect genomic, functional, and ecological information in predicting cross-infection at the community scale.
Keywords
Network modeling
Bipartite graph
Markov chain Monte Carlo
Matrix completion
Phage-bacteria coevolution
Controlling the total number of sequence reads for a given sample (read depth or library size) in virome sequencing studies is experimentally challenging. Since most alpha-diversity metrics are strongly dependent on read depth, statistical or computational normalization is necessary to ensure samples are measured on the same scale and to avoid confounding by read depth. Unfortunately, standard tools commonly used for read depth normalization in other omics fail for virome data. To address this problem, we propose the Read Depth Adjustment by Quantile-regression (ReDAQ) approach which regresses out the effect of depth from alpha-diversity metrics. To accommodate irregular distributions, we employ a quantile process regression approach and further allow for nonlinear read depth effects at each quantile. We show via real data analysis and simulations that ReDAQ works well in removing read depth effects, avoiding potential confounding, whereas extant approaches (developed primarily for bacteriome or other omics) fail.
Keywords
virome
diversity
quantile-regression
microbiome
library size
The human microbiome offers valuable insights into health and disease, yet leveraging it for accurate phenotypic prediction remains challenging. In this talk, I will present two AI-driven approaches that integrate biological knowledge to enhance microbiome-based prediction. One harnesses microbial phylogeny to improve clinical outcome prediction, while the other employs large language models guided by domain knowledge to enable more accurate and efficient disease diagnosis. Together, these strategies demonstrate the potential of combining AI with biological insight to improve phenotypic prediction from microbiome profiles.
Keywords
AlphaFold2
Computational Pipeline
Function-Based Annotations
Human Virome Catalog
Protein 3D Structures
Motivated from the polygenic risk score framework, we propose an integrated microbial risk score (MRS) framework to aggregate the complicated microbial profile into a summarized risk score that can be used to measure and predict disease susceptibility.
Keywords
risk score
metagenomic
disease prediction
multi-omics
Emerging evidence underscores the dynamic interplay between the human microbiome and the immune system, with the microbiome contributing to disease pathogenesis via mediation of causal pathways in conditions such as Alzheimer's disease and cancer. Yet, traditional mediation analyses are ill-suited for microbiome data, given its compositionality, sparsity, and high dimensionality. To overcome these limitations, we propose Absolute Microbiome Mediation Analysis (AMMA) — a novel framework leveraging the natural effects model, sure independence screening, and microbiome-specific bias correction to infer mediation effects based on absolute microbial abundances. To our knowledge, AMMA is the first method capable of inferring absolute abundances in a mediation framework, demonstrating superior performance in simulation studies.