Research

Sequencing gives us relative, not absolute, abundances. Ignoring that fact inflates false positives and can reverse the direction of an estimated effect. Our work on ANCOM-BC and ANCOM-BC2 estimates and corrects the sample-specific sampling fraction directly, delivering valid confidence intervals and controlled false discovery rates for differential abundance, extended to multiple groups, covariate adjustment, and repeated measures.

Related publications ANCOMBC software

Microbiome, metabolome, proteome, and clinical measurements each capture a partial view of the same biology. We develop correlation and latent-structure estimators that link these layers into a single coherent model; one example is SECOM, which recovers both linear and nonlinear taxa–taxa associations without the spurious correlations induced by compositionality.

Most omics studies are observational. We work on mediation analysis, confounding adjustment, and treatment-effect estimation tailored to high-dimensional, sparse, compositional exposures and mediators, asking not just which features differ, but which plausibly lie on a causal pathway.

Predictive accuracy alone rarely changes clinical or regulatory practice. We build machine-learning methods that pair competitive prediction with interpretable structure, uncertainty quantification, and feature attributions that hold up under the correlation patterns typical of omics data.

Methods are developed against real scientific questions. Current and past collaborations span:

• Cancer: biomarker discovery and treatment response.
• Infectious disease: gut microbiome, systemic inflammation, and HIV-1 acquisition.
• Aging: longitudinal shifts in the microbiome and metabolome.
• Maternal & child health: early-life exposures and developmental outcomes.
• Environmental health: exposure–microbiome–outcome pathways.
• Clinical trials: design and assessment of trial endpoints.