Deep bayesian biostatistics: AI-accelerated inference for big "n" and big "p"

The National Institute of General Medical Sciences awarded $433,125 to the University of California, Los Angeles on August 10, 2026, under the Biomedical Research and Research Training program (CFDA 93.859, Assistance Listing).

Andrew Holbrook, a biostatistician at UCLA, will develop AI-accelerated inference strategies and open-source software for Bayesian hierarchical models applied to high-impact biostatistical problems. The work addresses a computational bottleneck in biomedical research: Bayesian inference, while flexible for analyzing complex biological data structures, becomes prohibitively expensive as datasets grow in both sample size and dimensionality. Holbrook will focus on two primary applications: phylogenetic continuous-time Markov chain modeling for evolutionary biology, which is computationally intensive for high-dimensional traits, and high-dimensional regression. The resulting tools will accelerate Bayesian analysis across biomedical discovery workflows, particularly those leveraging genomic sequencing data and electronic health record datasets.

UCLA, operating through its Office of Research Administration, will serve as recipient. The award period runs from August 10, 2026, through April 30, 2031, with performance taking place in Los Angeles, California.

Abstract

Biotechnological advances such as fast genome sequencing accelerate data generation within biological and biomedical laboratories. Bioinformatical advances such as common data models for electronic health records (EHR) provide health systems the unprecedented ability to efficiently store and access massive quantities of OB- servational health data. Revolutions in data production, storage and access result in a deluge of biomedical data in the 21ST century, and the biostatistical analysis of observational data becomes a bottleneck in the larger work- flow contributing to biomedical progress. On the one hand, bayesian hierarchical models flexibly adapt to data's diverse, intrinsic structures and provide an important set of tools within biological and biomedical exploratory analysis and knowledge discovery. On the other hand, bayesian analysis often requires significant computational resources, and the time required to perform these analyses grows with the number (n) and size (p ) of individual data points. Thus, bayesian inference constitutes a rate-limiting step within many biostatistical analyses, in PAR- ticular, and larger biomedical discovery workflows, in general. Andrew holbrook (ucla biostatistics) will leverage his expertise in bayesian biostatistics, statistical computing and machine learning to develop AI-accelerated infer- ence strategies and concomitant open-source software for a collection of high-impact applications of structured hierarchical models. The phylogenetic continuous-time markov chain (ctmc) is the mathematical workhorse of evolutionary biology, and mixed-effects modeling of evolutionary substitution processes is an assumption-robust, gold-standard approach to phylogenetic ctmc modeling across a broad swath of applications that becomes pro- hibitively computationally expensive for high-dimensional traits. High-dimensional regression stands as a canoni- CAL statistical modeling task within EHR-based observational data analysis, and bayesian hierarchical models with sparsity-inducing shrinkage priors help discover an outcome's relatively small number of important predictors even when n ⇡ 106 and p ⇡ 105. Unfortunately, cutting-edge markov chain monte carlo (mcmc)-based inference for these models is slow on account of challenging model geometry and o(p 3) rate-limiting steps. When observa- tional data includes (j) longitudinally repeated measures, the gold-standard p -variate mixed-effects model with hierarchical subject- and outcome-specific random effects defies computational memory constraints. Analytically integrating over random effects drastically reduces model size, but inferring the remaining o(J2P 2) random-effect covariance matrix remains challenging. For each of these high-impact bayesian biostatistical tasks, holbrook will: build bespoke deep surrogate models to accelerate key computational bottlenecks in advanced yet computation- ally expensive mcmc algorithms; fine tune existing large-language models (llms) that completely replace these mcmc algorithms; use his priority-access cluster of 40 gpus to build and train completely novel llms that COM- pletely replace mcmc; and use diverse real-world big-n, big-p data from evolutionary biology (traits + genome) and EHR to inform methods that will effectively eviscerate the bayesian bottleneck of biomedical discovery.