Research

High-throughput multi-omics datasets exhibit several properties that complicate statistical analysis, including high dimensionality, discrete and compositional structure, over-dispersion, and complex hierarchical, spatial, and temporal dependence. Our lab develops specialized methods and software to characterize multimodal structure in large-scale population studies, with reproducible workflows that support both experimental and computational research (Mallick et al., 2017).

This work led to widely used statistical frameworks for multivariable association analysis in high-dimensional microbiome data (MaAsLin 2 and MaAsLin 3; >100K downloads), which support flexible confounder adjustment and scalable modeling across studies (Mallick et al., 2021; Nickols et al., 2026). We also led the development of MMUPHin, a general-purpose tool for microbiome meta-analysis that enables principled aggregation of evidence across studies (Ma et al., 2022), and of MelonnPan, an integrative framework that links microbial community features to community-level metabolite profiles (Mallick et al., 2019).

Through the NIH Integrative Human Microbiome Project, where I was the primary statistician, we led efforts in statistical quality control, multivariable modeling, meta-analysis, and integrative data analysis (Lloyd-Price et al., 2019). Building on this work, our lab co-developed HuMMANet, a harmonized resource of 14,405 paired microbiome-metabolome samples from 46 studies, which now supports cross-study pretraining and benchmarking of new multimodal methods (Verma et al., 2026).

Single-cell and other high-dimensional omics data are characterized by extreme sparsity, over-dispersed counts, and complex dependence patterns across features and samples, posing challenges for both modeling and inference. To address these challenges, we have developed Tweedieverse, a tool for single-cell and spatial differential expression, which has potential applications in other data types such as digital pathology, imaging, and single-cell and spatial multimodal data (Mallick et al., 2022). We have extended this framework to spatial omics through TweedSpot, a scalable method for detecting spatially variable genes using self-adaptive spatial Tweedie models. We have also developed DAssemble, an ensemble framework for differential analysis that stabilizes inference by aggregating evidence across multiple statistical models with complementary assumptions (Liu et al., 2026). Further, we have developed IntegratedLearner, an integrated Bayesian machine learner for multi-omics prediction and classification as well as for end-to-end biomarker discovery, which can be used to stratify patients for therapeutic interventions, providing a promising route to stratified medicine (Mallick et al., 2024).

A central theme of our research is the development of Bayesian methods for high-dimensional inference with principled uncertainty quantification (UQ). Building on our work in Bayesian regularization, we have developed cooperative learning frameworks (BayesCOOP; Roy et al., 2025) that integrate multiple data views while preserving principled UQ. Related work extends these frameworks using conformal inference (Coracle; Roy et al., 2026) to provide finite-sample uncertainty guarantees for multiview prediction under minimal distributional assumptions. We have further developed JAFAR, which draws on multiview factor models with cumulative shrinkage priors, separating shared and view-specific signals while enabling joint learning across data views (Anceschi et al., 2026).

Since 2023, our lab has built an independent research program centered on Bayesian multimodal inference, transfer learning, and uncertainty quantification, with a focus on causal multimodal AI for systems biology. Ongoing work includes (i) causal mediation analysis of multi-omics data, (ii) foundation models for microbial community metabolomics, and (iii) scalable, uncertainty-aware, interpretable machine learning for multimodal AI agents.

Our work on Bayesian regularization began with several variants of the Bayesian LASSO, including the Bayesian bridge, the group bridge, and the reciprocal LASSO, with improved posterior computation (Mallick and Yi, 2014; Mallick and Yi, 2017; Mallick et al., 2021). This line of work originated in my doctoral research, where I addressed challenges in posterior sampling through novel data augmentation strategies and efficient Markov chain Monte Carlo algorithms. These ideas informed subsequent contributions to Bayesian graphical modeling for compositional data (BAnOCC; Schwager et al., 2017). The methods are available in the BayesRecipe package.

Our early research focused on population structure discovery in epidemiologic studies, addressing methodological challenges in variant selection, haplotype-based association testing, and population structure estimation. In this work, we developed regularization-based approaches for multilocus association analysis and zero-inflated count outcomes (Mallick and Tiwari, 2016; Chatterjee et al., 2018), along with haplotype-based modeling frameworks (Li et al., 2014) and population structure discovery methods that emphasize principled handling of study heterogeneity (Ma et al., 2022). We also contributed to synthetic benchmarking frameworks that enable controlled evaluation of methodological performance under realistic data complexity (Ma et al., 2021).

We provide statistical leadership in the design, harmonization, and analysis of large-scale datasets that integrate molecular, genetic, observational, and longitudinal data at the population scale. In these projects, we have served as lead biostatisticians, with responsibilities spanning study design, data quality control, cross-platform integration, and the development of scalable and reproducible analysis strategies. This includes large national studies such as the National Birth Defects Prevention Study, the Nurses' Health Study, and the Health Professionals Follow-Up Study (Abu-Ali et al., 2018; Nguyen et al., 2020), and randomized trials such as the WIPE trial (Kelleher et al., 2013). Through a secondary appointment in the Division of Gastroenterology and Hepatology, where I serve as statistical lead, we extend this work to clinical and population-based studies of liver disease (Buckholz et al., 2026a; Buckholz et al., 2026b; Kushner et al., 2026).