A mutagenetic tree hidden Markov model for longitudinal clonal HIV sequence data

Niko Beerenwinkel, Mathias DrtonBiostatistics 2006 · DOI

TL;DR

This paper introduces a novel "mutagenetic tree hidden Markov model" (mtree-HMM) to analyze how HIV drug resistance mutations emerge over time within individual patients. By combining evolutionary pathways (mutagenetic trees) with temporal dynamics (hidden Markov models), the approach provides a more accurate understanding of resistance development than previous methods, crucial for designing better treatment strategies.

What Problem Does This Paper Address?

HIV treatment relies on combination drug therapy to suppress viral replication. However, HIV's high mutation rate allows it to evolve drug resistance, limiting treatment effectiveness. Understanding the specific order and timing of mutations that confer resistance is vital for predicting treatment outcomes and designing more robust drug cocktails. Previous methods often analyzed cross-sectional data (snapshots in time) or population-level sequencing, which can obscure the precise evolutionary path and the co-occurrence of mutations within individual viral genomes. This paper tackles the challenge of analyzing longitudinal clonal sequence data, which provides detailed information about viral evolution within individual patients over time, to infer these resistance pathways.

Methods

The core innovation is the mutagenetic tree hidden Markov model (mtree-HMM). This model integrates two key concepts:

  • Mutagenetic Trees: These represent the evolutionary pathways of mutations. They are directed acyclic graphs where nodes are mutations, and edges indicate that one mutation typically arises after another. This captures the idea that mutations don't occur randomly but follow preferred "paths" due to genetic interactions and selective pressures. The model assumes mutations are non-reversible (once fixed, they don't disappear) and that they occur independently along these tree edges via Poisson processes.
  • Hidden Markov Models (HMMs): HMMs are used to model systems that transition between unobservable (hidden) states over time. In this context, the "hidden states" are the true mutational patterns of the virus population within a patient at a given time. The model assumes that observed clonal sequences are "noisy" copies of these hidden states, with a certain probability of false positives (detecting a mutation that isn't there) and false negatives (missing a mutation that is present).

The mtree-HMM combines these by treating the mutagenetic tree as defining the possible hidden states and their transitions over time, while the HMM framework accounts for the observed clonal data and the uncertainty in inferring the true population state. The model parameters (mutation rates, false positive/negative rates) are estimated using an Expectation-Maximization (EM) algorithm.

Datasets

The study uses longitudinal clonal HIV sequence data from three phase II clinical studies of the reverse transcriptase (RT) inhibitor efavirenz. This dataset comprises 3350 clones from 163 patients, collected at multiple time points. The data focuses on seven specific amino acid substitutions in the HIV-1 RT associated with efavirenz resistance.

Key Insights and Ideas

  • Ordered Accumulation of Mutations: The study strongly supports the idea that drug resistance mutations don't appear randomly but follow specific, ordered pathways. This is crucial because the combination of mutations, not just individual ones, dictates the level of resistance.
  • Non-Reversibility of Mutations: The data suggests that once a resistance mutation becomes established in the viral population, it generally remains fixed, rather than reverting to the wild-type. This is likely due to the strong selective advantage conferred by these mutations under drug pressure.
  • Clonal Variation as Noise: The model treats the diversity observed among clones from the same patient at the same time point primarily as a result of experimental noise (PCR errors, viral copying errors) rather than as evidence of multiple distinct, co-existing viral strains dominating the population. This assumption is justified by the strong selective pressure that tends to favor a single dominant strain.
  • Predictive Power for Therapy: Understanding these evolutionary pathways and rates is directly applicable to predicting treatment response and designing more effective, durable drug regimens.

Key Results

  • Mutagenetic Tree Structure: The analysis identified a primary pathway for efavirenz resistance, initiated by mutation 103N, followed by other mutations like 100I, 101Q, 108I, or 225H. A secondary, slower pathway involves mutation 190S followed by 101E.
  • Mutation Rates: Mutation 103N was found to be the earliest to occur, with an expected waiting time of about 19 weeks. Mutation 190S was significantly slower, appearing around 478 weeks. The order of subsequent mutations was also estimated.
  • Assumption Validation: Statistical tests supported the assumptions of non-reversibility of substitutions (A2) and a single dominant population state (A3) for most mutations, although some exceptions were noted for 101Q and 190S.
  • Model Performance: The mtree-HMM successfully modeled the longitudinal clonal data, providing estimates for mutation progression rates and clonal variation parameters.

Limitations and Future Work

  • Single Dominant Strain Assumption: The model assumes the population is dominated by a single strain. While justified by strong selective pressure, it doesn't explicitly model multiple co-existing lineages or complex parallel evolutionary pathways, which could contribute to elevated false negative rates for some mutations. The authors suggest future work could involve modeling multiple strains or grouping clones into pathways.
  • Recombination: The study assumes mutations occur independently and doesn't explicitly model recombination, although the mutations considered are relatively close on the genome, making recombination less likely. Incorporating recombination would be a significant extension.
  • Interval Censoring: The data has interval censoring (mutations are observed at specific time points, but their exact fixation time within that interval is unknown). The paper acknowledges this but assumes its impact is minor due to short interval lengths. For longer intervals, modeling interval censoring would be more important.
  • Generalizability: While the model is applied to efavirenz resistance, extending it to other drugs or combinations, and to different viruses, would be valuable.

Summary updated

Abstract

RNA viruses provide prominent examples of measurably evolving populations. In human immunodeficiency virus (HIV) infection, the development of drug resistance is of particular interest because precise predictions of the outcome of this evolutionary process are a prerequisite for the rational design of antiretroviral treatment protocols. We present a mutagenetic tree hidden Markov model for the analysis of longitudinal clonal sequence data. Using HIV mutation data from clinical trials, we estimate the order and rate of occurrence of seven amino acid changes that are associated with resistance to the reverse transcriptase inhibitor efavirenz.
BibTeX
@inproceedings{beerenwinkel2006mutagenetic,
  author = {Niko Beerenwinkel and Mathias Drton},
  title = {{A mutagenetic tree hidden Markov model for longitudinal clonal HIV sequence data}},
  booktitle = {Biostatistics},
  year = {2006},
  doi = {10.1093/biostatistics/kxj033},
}
Figure