Palmer et al. find that HIV evolution is not so fast

Usually, I use this blog to write about how I do my work, but today I write about science! The blog is cross-posted on Haldane’s Sieve, a website that provides a “feed of preprints in the fields of evolutionary and population genetics.”

Last week, a group of people from Oxford University published an interesting paper on the ArXiv. The paper is about using genealogical data (from HIV sequences), in combination with cross-sectional data (on patient and HIV phenotypes) to infer rates of evolution in HIV.

My conclusion: the approach is very interesting, and it makes total sense to use genealogical data to improve the inference from cross-sectional data. In fact, it is quite surprising to me that inferring rates from cross-sectional data works at all. However, in a previous paper by (partly) the same people, they show that it is possible to infer rates from using cross-sectional data only, and the estimates they get are very similar to the estimates from longitudinal data. The current paper provides a new and improved method, whose results are consistent with the previous papers.

The biological conclusion of the paper is that HIV adaptation is slower than many previous studies suggested. Case studies of fast evolution of the virus suffer from extreme publication bias and give the impression that evolution in HIV is always fast, whereas cross-sectional and longitudinal data show that evolution is often slow. Waiting times for CTL-escape and reversion are on the order of years.

The paper: 

Integrating genealogical and dynamical modelling to infer escape and reversion rates in HIV epitopes, Duncan Palmer, John Frater, Rodney Philips, Angela McLean, Gil McVean, http://arxiv.org/abs/1302.1098

The previous paper: 

Modelling the evolution and spread of HIV immune escape mutants.
Fryer HR, Frater J, Duda A, Roberts MG; SPARTAC Trial Investigators, Phillips RE, McLean AR.
http://www.plospathogens.org/article/info%3Adoi%2F10.1371%2Fjournal.ppat.1001196

1. What rates are they interested in?

The rates of interest here are the rate of escape from CTL pressure and the rate of reversion if there is no CTL pressure.

When someone is infected with HIV, the CTL response by the immune system of the patient can reduce the amount of virus in the patient. CTL stands for cytotoxic lymphocytes. Which amino-acid sequences (epitopes) can be recognized by the host’s CTL response depends on the HLA genotype of the host.
Suppose I have a certain HLA genotype X, such that my CTLs can recognize virus with a specific sequence of about 9 amino acids, let’s call this sequence Y. To escape from the pressure of these CTLs, the virus can mutate sequence Y to sequence Y’. A virus with sequence Y’ is called an escape mutant. The host (patient) with HLA X is referred to as a “matched host” and hosts without HLA X are referred to as “unmatched.” The escape mutations are thought to be costly for the virus.
So, for each CTL epitope there are 4 possible combinations of host and virus:
1. matched host and wildtype virus (there is selection pressure on the virus to “escape”)
2. matched host and escape mutant virus
3. unmatched host and wildtype virus
4. unmatched host and escape mutant virus (there is selection pressure on the virus to revert)

The question is “how fast does the virus escape if it is in a matched host and how fast does it revert if it is in an unmatched host?”

2. Why do we want to know these rates?

First of all, just out of curiosity, it is interesting to study how fast things evolve – it is surprising how little we know about rates of adaptive evolution. Secondly, because escape rates are relevant for the success of a potential HIV vaccine, if escape rates are high, then vaccines will probably not be very successful.

3. What are cross-sectional data and how can we infer rates from them?

Cross-sectional data are snap-shots of the population, with information on hosts and their virus. Here, it is the number of matched and unmatched hosts with wildtype and escape virus at a given point in time.

So how do these data tell us what escape rates and reversion rates are? Intuitively, it is easy to see how very high or very low rates would shape the data. For example, if escape and reversion would happen very fast, then the virus would always be perfectly adapted: we’d only find wildtype virus in unmatched hosts and only escape mutant virus in matched hosts. Conversely, if escape and reversion would be extremely slow, than the fraction of escape mutant virus would not differ between matched and unmatched hosts. Everyone would be infected with a random virus and this would never change.
The real situation is somewhere in between: the fraction of escape mutant virus is higher in matched hosts than in unmatched hosts. With the help of an standard epidemiological SI-model (ODE-model) and an estimate of the age of the epidemic, the fraction of escape mutant virus in the two types of hosts translates into estimates of the rates of escape and reversion. In the earlier paper, this is exactly what the authors did, and the results make a lot of sense. Rates range from months to years, reversion is always slower than escape, and there are large differences between CTLs. The results also matched well with data from longitudinal studies. In a longitudinal study, the patients are followed over time and evolution of the virus can be more directly observed. This is much more costly, but a much better way to estimate rates.

4. Why are the estimates from cross-sectional data not good enough?

Unfortunately, the estimates from cross-sectional data are only point estimates, and maybe not very good ones. The problem is that the method (implicitly) assumes that each virus is independently derived from an ancestor at the beginning of the epidemic. For example, if there are a lot of escape mutant viruses in the dataset, then the estimated rate of escape will be high. However, the high number of escape mutant virus may be due to one or a few escape events early on in the epidemic that got transmitted to a lot of other patients. It is a classical case of non-independence of data. It could lead us to believe that we can have more confidence in the estimates than we should have.

5. Genealogical data to the rescue!

Fortunately, the authors have viral sequences that provide much more information than just whether or not the virus is an escape mutant. The sequences of the virus can inform us about the underlying genealogical tree and can tell us how non-independent the data really are (two escape mutants that are very close to each other in the tree are not very independent). The goal of the current paper is to use the genealogical data to get better estimates of the escape and reversion rates.

A large part of the paper deals with the nuts and bolts of how to combine all the data, but in essence, this is what they do: They first estimate the genealogical tree for the viruses of the patients for which they have data (while allowing for uncertainty in the estimated tree). Then they add information on the states of the tips (wildtype vs escape for the virus and matched vs unmatched for the patient), and use the tree with the tip-labels to estimate the rates. This seems to be a very useful new method, that may give better estimates and a natural way to get credible intervals for the estimates.

The results they obtain with the new method are similar to the previous results for three CTL epitopes and slower rates for one CTL epitope. The credible intervals are quite wide, which shows that the data (from 84 patients) really don’t contain a whole lot of information about the rates, possibly because the trees are rather star-shaped, due to the exponential growth of the epidemic. Interestingly, the fact that the tree is rather star-shaped could explain why the older approach (based only on cross-sectional data) worked quite well. However, this will not necessarily be the case for other datasets.

Question for the authors

Do you use the information about the specific escape mutations in the data? Certainly not all sequences that are considered “escape mutants” carry exactly the same nucleotide changes? Whenever they carry different mutations, you know they must be independent.

Help!

I guess you know the situation. Sometimes there is just too much work for not enough time. A grant proposal is due. A paper needs to be revised. Collaborators want to discuss new ideas. And there is a talk to prepare. You think you’ll drown.

So this actually happened to me last week. As I realized how bad the situation was (some time last weekend) and realized that the stress was probably going to last for at least 10 days (until the grant proposal deadline), I tried to remember everything I had read about stress-management in the last years and I decided to
1. try and keep healthy habits, 2. actively postpone things and 3. get help.

I’ll tell you a bit more.

1. To me, keeping healthy habits means, for example, eating proper food, getting enough sleep and having lunch with my colleagues. These habits are important because they help me to work most efficiently in my working hours and stay sane during a stressful week.

2. Next, I identified a few things that could be postponed, and I postponed them. Most importantly, I sent emails to my collaborators or others involved to let them know that I was postponing things. I find that telling people that I will not be working on our shared project this week, is key to actually freeing my mind from thinking about that project.

Ask for help

3. Maybe most importantly, I decided to ask for help. I found two to-do items that I could ask others to do for me.

One of the to-do items was to run simulations. I needed to run simulations with code that I hadn’t used for a long time. It would certainly take me a few hours to remember how the code worked and to get the results I needed. However, I knew that a graduate student in the office next to mine was working on a related project and he had recently been running simulations that were very similar to what I needed. So I asked him if he were willing to run the simulations for me and he was 🙂
In fact, my request led to a few interesting discussions and who knows, maybe it will lead to a fun collaboration.

The other thing I asked someone else to do, was to edit my paper. I had a paper that was almost accepted, except that the editors wanted me to make a few changes to the text. Because of all the other things on my plate, editing the paper seemed like a huge task. I decided to ask a friend who sometimes works as an editor to help out. I knew that she could do a better job than me in less time. And she’d be happy to do it, because I’d be paying for it. It cost me 100 dollars, but it saved me at least two hours of work and a lot of stress.

Plan B

Now that I have written this post, I realize that there is one more thing that helps me deal with stress: making a plan B.

Last week, the stress was mainly due to a grant proposal I had to write. So what I needed to do is determine what would happen if I didn’t manage to write a good proposal. I needed to answer the question: “What’s the worst that can happen if I don’t make the deadline and how will I deal with that?” By knowing how I will deal with failure, I feel much more at ease, which helps me to sleep well, which helps me to work efficiently, which helps me not to fail.

In the end, it turned out that my grant proposal was not due until May 7th (now I know that the NIH has separate deadlines for AIDS-related proposals!), so my week ended with less stress than expected.

I’d love to hear how you deal with stress!