Using epitomes to model genetic diversity: Rational design of HIV vaccines

Nebojsa Jojic, Vladimir Jojic, Brendan J. Frey, Christopher Meek, David HeckermanNeurIPS 2005 · OpenReview

The rapid mutation rate of the Human Immunodeficiency Virus (HIV) presents a formidable "moving target" for vaccine design. A vaccine based on a single viral strain is quickly rendered obsolete by the virus's genetic evolution. To counter this, researchers design vaccine "cocktails" containing multiple variants to cover as much genetic diversity as possible. However, physical and financial constraints limit the size of these cocktails. The core challenge is a classic optimization problem: selecting the smallest possible set of protein fragments that maximizes coverage of potential immune targets (epitopes) across the global viral population. Traditional methods, such as using consensus strains or phylogenetic tree centers, often struggle with the overlapping nature of these fragments and the sheer scale of viral diversity.

To address this, researchers have adapted "epitomes"—a probabilistic model originally used for image and audio compression—to map the vast genetic diversity of HIV.

The Sequence Epitome Model

The authors introduce the Sequence Epitome, a compact summary of a large dataset. Unlike a standard sequence consisting of fixed letters, an epitome is a sequence of probability distributions. At each position, it provides a "profile" of which amino acids are likely to appear based on the input data.

Visualizing the epitome mapping

The model assumes that any fragment of a real virus can be generated by picking a starting point in the epitome and reading off a short window. To train this model, the researchers employ an Expectation-Maximization (EM) algorithm, iteratively refining the epitome to maximize the likelihood of it containing all the overlapping fragments (e.g., 9-mers or 10-mers) found in the raw viral data. To avoid mathematical local minima during learning, the epitome is treated as a circular topology, allowing fragments to wrap around from the end back to the beginning.

For vaccine design specifically, a "greedy" myopic optimization algorithm is used, iteratively appending the most useful fragments to the vaccine sequence until the desired length is reached.

Datasets and Key Insights

The model was evaluated using synthetic data (to test pattern reconstruction amidst noise) and a real-world dataset of 176 HIV-1 gag protein sequences from the WA cohort. The SYFPEITHI database of experimental epitopes was used to validate whether the epitome captured immunologically relevant targets.

A major insight from this approach is that the overlapping nature of immune targets is a feature, not a bug. The human immune system recognizes viruses by breaking them into overlapping 8-11 amino acid fragments. Because the epitome model is built entirely on overlapping "patches," it naturally mimics this biological process. Furthermore, the researchers discovered that optimizing the epitome to cover all possible 10-mers in the sequence data provides nearly identical immune coverage to models relying on expensive, hand-curated epitope databases. The optimization was also surprisingly robust to uncertainties regarding T-cell cross-reactivity.

Superior Efficiency and Coverage

The primary metric for success was Expected Coverage: the percentage of epitopes in the circulating viral population that the vaccine would recognize.

Expected coverage vs. vaccine length

The epitome-based design demonstrated superior efficiency and compression power. It successfully summarized the diversity of 176 full-length proteins into a sequence only slightly longer than a single protein, while maintaining high coverage of the population's variants. For any given vaccine length, the epitome provided significantly higher coverage than "consensus cocktails" (averages of common strains), "tree centers" (phylogenetic midpoints), or randomly selected observed strains. As shown in the results, an epitome-designed vaccine of approximately 1,000 amino acids achieved coverage levels that required nearly double the length using traditional methods.

Limitations and Future Work

While computationally powerful, the approach has limitations. Like many EM-based models, the algorithm can still get stuck in sub-optimal solutions despite the circular topology. Furthermore, this study serves as a computational proof-of-concept. Delivering these artificially designed "epitome" sequences into a human cell as a functional, safe vaccine involves significant downstream bioengineering challenges. Finally, the current model focuses primarily on the cellular (T-cell) immune response via MHC-I presentation; future iterations will need to integrate B-cell responses (antibodies) to facilitate a truly comprehensive vaccine design.

Summary updated

Abstract

We introduce a new model of genetic diversity which summarizes a large input dataset into an epitome, a short sequence or a small set of short sequences of probability distributions capturing many overlapping subsequences from the dataset. The epitome as a representation has already been used in modeling real-valued signals, such as images and audio. The discrete sequence model we introduce in this paper targets applications in genetics, from multiple alignment to recombination and mutation inference. In our experiments, we concentrate on modeling the diversity of HIV where the epitome emerges as a natural model for producing relatively small vaccines covering a large number of immune system targets known as epitopes. Our experiments show that the epitome includes more epitopes than other vaccine designs of similar length, including cocktails of consensus strains, phylogenetic tree centers, and observed strains. We also discuss epitome designs that take into account uncertainty about Tcell cross reactivity and epitope presentation. In our experiments, we find that vaccine optimization is fairly robust to these uncertainties.
BibTeX
@inproceedings{jojic2005using,
  author = {Nebojsa Jojic and Vladimir Jojic and Brendan J. Frey and Christopher Meek and David Heckerman},
  title = {{Using epitomes to model genetic diversity: Rational design of HIV vaccines}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2005},
  url = {https://openreview.net/forum?id=Sk--0Obd-r},
}
Figure