Tyler Grimes

Assistant Professor of Statistics
Department of Mathematics and Statistics
University of North Florida

Google Scholar LinkedIn Github CV

Hi, I’m Tyler, an assistant professor of statistics at the University of North Florida. My work sits between statistics and biology. Before joining UNF, I spent five years analyzing biological data from vaccine clinical trials in industry. That experience shapes how I approach research: methods should work on real data, answer the questions scientists and clinicians actually ask, and come with free software that other researchers can use. I’m always happy to hear from students interested in working at that intersection.

On this site, you can find various notes containing random experiments and code snippets used for testing ideas, along with talks and papers from past research projects.

Grimes Lab

Our group develops statistical methods for studying systemic lupus erythematosus (SLE) and related autoimmune diseases using large-scale molecular data such as gene expression. Lupus is one of the most common autoimmune diseases and can affect many organ systems. It is a heterogeneous disease, meaning it presents differently from one person to the next, which makes it hard to diagnose and means there is no single treatment that works for everyone. That heterogeneity is also a statistical problem. Patients and healthy controls differ in many ways at once, and the standard tools for comparing them can mistake those differences for new biology. We study what common methods actually estimate in these settings, and we develop methods that give each patient their own answer: how their genes are coordinated, how that coordination differs from other patients, and how it changes with disease activity or treatment. The goal is analysis that holds up in real cohorts and can eventually inform how patients are grouped and treated.

Statistical research

We develop statistical methods for valid inference on dependence structure (co-expression and association networks) in heterogeneous patient cohorts: asking what widely used procedures actually estimate when groups differ in means, variances, composition, and size, and building estimators for the quantities they were meant to estimate, with lupus as the motivating application.

Our statistical research centers on:

• Valid inference for association and co-expression networks in heterogeneous cohorts
• Subject-level and conditional estimands: methods that give each patient their own network, and a correct account of what that network measures
• Pseudo-value and influence-function methods, including their semiparametric theory
• Adjustment for cell-type composition and other sources of between-group difference in bulk-tissue data
• Permutation and distance-based tests for unbalanced and repeated-measures designs
• Open-source software implementing these methods in R

Current research questions include: When sample-specific networks are compared between patients and controls, how much of the difference is differential expression or cell composition in disguise, and how do we remove it? Can we estimate how two genes’ coordination changes with disease activity or treatment, continuously and after adjusting for confounders, with valid confidence intervals? When adjusting a comparison for estimated cell composition, how do we tell whether the adjustment has created or erased a difference? How should distance-based tests be run when some patients contribute many more observations than others?

Earlier work focused on RNA-seq gene expression data in cancer, developing methods in graphical modeling, prediction and survival analysis, and asking whether gene regulatory networks differ between high-risk and low-risk patients and how gene expression can improve prediction of survival.

Across all of this, simulation is an indispensable tool. Beyond assessing performance, building a simulation forces us to write down the data-generating process and the context of the problem, and that step is often where insights appears. Much of our current work follows the same pattern: specify the process, derive what a standard method actually estimates under it, and confirm the result by simulation before proposing a fix.