Statistical Issues in AIDS Research

University Of Washington, Seattle WA

Investigators

Linked publications & trials

Abstract

During the MERIT period we will continue to build on the progress made (see progress report) in the current grant cycle in these three areas: 1) Methods for HIV prevention and treatment trials We will continue to investigate clinical trial designs that enable efficient and effective evaluation of interventions for treatment and prevention of HIV, including i) further development of platform trial designs that are particularly efficient when multiple interventions can be assessed, ii) master protocol designs that can be used in diverse and international clinical settings; iii) designs for active control PrEP trials that leverage additional information to infer HIV incidence in the absence of intervention (“counterfactual placebo incidence”), including noninferiority trials augmented with HIV exposure biomarkers and recency assay data. We will develop methods for the design and analysis of stepped wedge designs, including i) using the theory of Structural Nested Mean Models (SNMM) (Robins, 1994) to develop robust and efficient intervention effect estimates for stepped wedge designs, including methods that provide consistent estimates when the treatment effect is not constant and incorporating both design-based and asymptotic inference methods; ii) formulating the log hazard ratio for the treatment effect on the clinical outcome as a continuous, piecewise linear function of time elapsed since the initiation of the treatment. Statistical inference will be based on partial likelihood, with a sandwich variance estimator to account for intra-cluster correlation. Interval-censored outcomes will be accomodated; iii) updating our R software tools to incorporate the methodology developed. We will develop new statistical methods for HIV/AIDS studies in which HIV infection is an intermediate event whose effect on another outcome (e.g., stroke) is of interest. Since HIV infection is only known to occur in a time interval induced by periodic blood tests, we will formulate the effects of covariates on time to HIV infection through the familiar Cox proportional hazards model and adopt nonparametric maximum likelihood estimation with interval-censored observations. We will develop new, highly efficient estimators of an optimal joint dynamic treatment and screening strategy that exploit the no direct effect (NDE) assumption that screening has no effect on a patient’s clinical outcome of interest except through the effect of the screening results on the choice of treatment. For the management of HIV+ individuals, our methods will provide practical guidance on cost-effective strategies that determine at each clinic visit (1) whether to order viral load and/or CD4 count tests (at some cost and burden to the patient) and (2) whether to start or switch anti-retroviral treatment. 2) Charactizing the HIV epidemic To estimate yearly, subnational variation in HIV indicators using household survey data one must account for the complex design. Inference is required at the geographic admin2 level (which is two below the national) but design-based estimators are unstable at this level due to data sparsity. However, at the admin-1 level such estimators can be used and we desire estimates at admin-2 that are consistent with these. We will develop model-based methods for admin-2 level data, but benchmarking to the design-based estimates at admin-1. 3) Analysis of HIV cure and vaccine studies We will continue developing statistical models and computational tools to analyze complex high-dimensional data produced by bioassays used in HIV vaccine studies. In particular, we will focus on methodologies to support the design and evaluation of novel germline-targeting vaccines that seek to elicit broadly neutralizing antibodies through a series of vaccine immunogens. First, we will use our modeling framework that describes evolution of the B cell receptor (BCR) repertoire (using a multitype age-dependent branching stochastic processes) to improve our understanding of affinity maturation. Second, we will develop machine learning approaches using ensembles of deep neural networks to predict the impact of the accumulation of somatic hypermutations on the binding affinity of BCR. Third, we will develop a novel class of computational tools to more robustly preprocess and annotate BCR sequencing data and genotype the immunoglobulin genes of study volunteers by integrating biological knowledge on somatic recombination into the design of the algorithms. We will develop an extension of the serial limiting dilution assay methods developed in Trumble et al. (2017) to accommodate information about the number of distinct viral lineages obtained from viral deep sequencing as described by Lee et al. (2017). In the context of HIV cure research, these methods will provide a point estimate and confidence interval for the size of the latent HIV reservoir. Deep sequencing (as in Lee et al. (2017)) of viruses such as HIV permits examination of genetically diverse populations of viruses, and in particular can lead to the identification of variants which may be of interest. Unfortunately deep sequencing utilizes next gen sequencing platforms which are error prone, which requires particular attention when trying to identify minority variants. We will develop methods to would identify sequence base pairs which exhibit variation across reads beyond what would be expected solely due to sequencing error, while controlling the false discovery rate.