Mystery Rays from Outer Space

Meddling with things mankind is not meant to understand. Also, pictures of my kids

December 19th, 2009

Baseball: Predictive value of UZR

Jacoby Ellsbury diving catch
Typical Ellsbury catch

(I see it’s been a long time since my last baseball-related post. Here’s a long one to make up, so I’m good for another year, baseball-post-wise.)

Jacoby Ellsbury, the centerfielder for the Boston Red Sox, was just given the “Defensive Player of the Year” award, as voted by baseball fans in the “This Year In Baseball Awards.”  This is interesting because, statistically, Ellsbury in 2009 was actually one of the worst defensive players in all of baseball. Does this mean that baseball fans got it wrong, or did the statistics lie to us?

The answer isn’t immediately obvious, because it’s generally agreed that statistical analysis of baseball defense1 still lags well behind most offensive and even pitching measurements. If a defender makes a catch, is that because everyone, including me, would have caught it? Was it a difficult catch, but one that most major-league players should have made? Or was it a really difficult catch, that almost every major-leaguer would have flubbed? Fans tend to judge these by the spectacularity of the catch, and there’s no doubt that Ellsbury made his share of spectacular, diving, all-out catches in 2009. But did he make them harder than they were?

In this case, actually, defensive stats and a lot of astute observers agree. Ellsbury didn’t get “good jumps” on a lot of hits this year — he hesitated before running, and he may not have run to the best spot for a catch. As a result, he had to make spectacular catches, on hits that a better defender would have caught quite routinely.

Easy, hard, or making it harder

J.D. Drew catch
Typical J.D. Drew catch

An interesting contrast to Ellsbury is the guy to his left, J.D. Drew, who plays right field for the Red Sox. Drew makes few spectacular catches, rarely diving or getting mussed up. His catches almost always look routine and easy. But statistically, he was a much better fielder than Ellsbury in 2009. In UZR/150 games, Ellsbury was -18.3, while Drew was at +15.7 — second-highest among right-fielders2 in the majors last year. Again, many careful observers agree with the stats; Drew doesn’t have to make spectacular catches, because he instantly sees where the ball will go, moves to the right place immediately, and makes the catch easily. Casual fans don’t think of Drew as a high-quality defender, because he doesn’t seem to make difficult plays; but in fact, he is making the difficult plays, he’s just making them look easy.

So UZR seems to agree fairly well with careful observers’ analysis of Ellsbury and Drew’s respective abilities. But that’s not really the interesting question. It’s mildly entertaining to say, “Yeah, baseball fans got it wrong”, but it’s much more interesting to ask what this tells us about the future. In other words: UZR seems to be an reasonable descriptive statistic. Is it a useful predictive statistic as well? For example, should Ellsbury play CF for the Sox next year? What are the chances that he’ll be a good centerfielder next year? What are the chances that Drew was just lucky this year, and next year he’ll be a lousy right fielder?

The numbers and the future

And here the waters get much more muddy. One puzzling point is that in 2008, Ellsbury was an excellent fielder, by UZR/150. He was pretty good at CF (+6.9) and superb in right field (+18.6, although in limited time — 36 games). Do players often show this kind of 20-odd point swing in UZR? And what does it tell us about that player’s future? (My answer, for those with tl;dr disease: Based on history, Ellsbury has about a 40% chance of being at an average or better fielder next year, and a 13% chance of being either good, or very good.)

Let’s ask a number of questions about UZR/150 and its ability to predict future defense.

(1) How well does a player’s UZR/150 correlate, year to year? That is, if a player has a particular UZR/150 this year, how similar will his next year’s UZR/150 be?

(2) How many players have the kind of huge drop in UZR/150 that Ellsbury showed? What happened to their defense in the years following that drop?

(3) If a defender is “very bad” this year,3 how likely is it that he’ll be a decent or good defender next year?

I’ve scraped FanGraph’s fielding ratings and dumped them into an SQLite database on my own computer, so that I can look at these questions. 4 Here are my attempts at answers.

Correlations between years for UZR/150
Correlations between years for UZR/150 (click for larger version)

(1) UZR/150 from one year does correlate with the next, but not very well. If we limit our analysis to outfielders who spent more than 65 games at a particular position in a year,5 and plot out each year vs. the following year in a scatter graph, the R2 is just 0.1823, and it only gets a little better (0.2505) if we limit it to players with at least 150 games at the position (see the figure at left).  In fact I tried all kinds of variations, and the only R2 that was over 0.5 was if I limited the analysis to the very best and the very worst outfielders6  who played over 130 games at a position; there the correlation with their next year was 0.5494.

So yes, there’s some correlation, on the bulk level.  But not much.  On an individual basis — which is, of course, what we’re interested in — you couldn’t be at all confident that next year’s UZR/150 will be very similar to this year’s.

(2) 20-point drops in UZR/150 aren’t unheard of, and players can bounce back from them. This kind of swing isn’t all that common, but it’s happened.  I turned up 24 outfielders since 20027 who had at least a 20-point change in UZR/150 from one year to the next. Of the 21 players with at least a 20-point drop, 7 of them stopped playing that position.  Six of the rest had a drop in 2009, and some of those won’t be back.  There were 12 who had a 20-point drop and continued at the position; 8 and of these, at least half bounced back, at least temporarily:

  • Andruw Jones (from 34.7 in 2005 to 13.1 in 2006; then 22.2, then 0.2, and then out of the position)
  • Kenny Lofton (19.9 in 2005; -17 in 2006; 8.3 in 2007)
  • Corey Patterson (-11 in 2002; then 14.8, 33.8, 11.3, 14.2, 1, 0.7)
  • Willy Taveras (22.6 in 2006; then -7.1, -3, and back up to 14.1)
  • Jeff Francoeur (30.1 in 2005, 7.4 in 2006, 16.9 in 2007; but then -4.9, -19 in 2008 and 2009)
  • Juan Encarnacion (all over the place. From 2003: -11.4, 13.5, -11.1, 7.1, -26.7)
  • Reggie Sanders (12.9 in 2002; then -7.2, 4.3, and -9.2)

The complete table9 is here.

Very good 1st year (n=105)
Percent at least OK Percent at least good
85.72 65.72
Good 1st year (n=271)
Percent at least OK Percent at least good
81.56 44.29
OK 1st year (n=396)
Percent at least OK Percent at least good
68.69 29.8
Bad 1st year (n=229)
Percent at least OK Percent at least good
54.14 17.9
Very bad 1st year (n=70)
Percent at least OK Percent at least good
38.57 12.86

(3) Very good and very bad defenders tend to be consistently good or bad. Although the fine correlation just isn’t there, can we draw a more general conclusion?  If we have a player who (based on UZR/150) is very good, good, just about average, bad, or very bad this year,10 what are his chances of being at least average, or of being at least good, next year? A summary of those chances, based on historical analysis of the 1071 players who qualified, is shown at the right; a more detailed breakdown is here.

What we see is that a player who was very good one year, has an 85% chance of being at least decent the following year, and a 66% chance of being good or very good. But a player who was very bad one year (for example, Jacoby Ellsbury this year) has a 40% chance of being at least decent the following year, and a 13% chance of being good or very good. So, again, there is reasonable predictive value when we look in this rather coarse-grained way, but there’s a ton of year-to-year variability.

I won’t show the data but here increasing the number of games played to 130 or 150 per year doesn’t help very much, the percents remain surprisingly similar although the numbers drop.

The bottom line

So what can we expect from Ellsbury next year (assuming he plays centerfield in 2010)?  Well, looking at the history of outfielders with that kind of drop in UZR/150, maybe there’s around a 20-50% chance that he’ll bounce back to be a decent CF (from question 2, above).  Looking at all players (question 3 above) there’s about a 40% chance that he’ll be decent, and a 12% chance that he’ll be good or very good next year.

Not great numbers, but that’s what we see.  My own suspicion is that Ellsbury will be a pretty good defender next year, but I wouldn’t put a lot of money on it.


  1. The most widely used is probably “UZR”, the “ultimate zone rating”; see here, here, and here for explanations. The other contender is the plus/minus rating system. UZR is freely available from the FanGraphs web site, while plus/minus requires a subscription to Bill James Online, so I’m only using UZR — actually, UZR/150, which is UZR normalized to 150 games.[]
  2. Those who played more than 100 games in RF[]
  3. I.e. has a low UZR/150[]
  4. Not that other, better-qualified, people haven’t already asked the questions. But poking at data is how I try to understand it, so here it is.[]
  5. Which, not entirely coincidentally, includes Ellsbury’s 2008, when he played 66 games at CF[]
  6. UZR/150 of < 15 or > 15[]
  7. When UZR was introduced[]
  8. Several players had more than one swing year, so these numbers don’t add up all that nicely[]
  9. Reminder: this is for outfielders only, not all players[]
  10. I used UZR/150 cutoffs of > +15, +5  to + 15, -5 to + 5,  -15 to -5, and less than < -15 for the different grades[]
August 28th, 2009

Influenza – more diverse than you thought

Virons le virus (Institut Merieux Benelux, 1991)
“Virons le virus” (Institut Merieux Benelux, 1991)

One of the important drivers of influenza virus evolution is mixed infection: Infection of the same individual with two different strains of virus, which can then reassort to generate brand-new viral genomes. This presumably what happened, for example, with the recent swine-origin influenza virus (SOIV): some pig was simultaneously infected with North American swine flu and a Eurasian swine flu, the two reassorted so that two of the Eurasian virus’s segments joined with 6 of the North American segments, and the new virus thus produced turned out, just by chance, to be good at infecting humans.

Reassortment, notoriously, can generate rapid large changes in the personality of the virus. Pandemic influenzas have been reassortants, unrecognized by the population’s immune systems. But that’s not the only possible outcome; reassortants between closely-related viruses can lead to small changes, reassortants between two circulating strains would still be recognized by the immune response, and so on. Reassortment per se isn’t inevitably devastating, the big concern is reassortment between widely-differing viruses — human and avians strains being the major issue today.

I’ve tended to think of multiple infection and reassortment as quite a rare phenomenon. Reassorted influenza viruses appear and circulate relatively often, but not, you know, daily;1 and those are the product of millions upon millions of infected individuals. On the other hand, most reassortments are probably either dead on arrival (their different segments are simply not compatible) or at best very unfit (their different segments make them easily outcompeted by the wild flu that’s already well adapted to the individuals in question). That means we don’t know the frequency of reassortants, because most of them would be invisible to us.

I’ve also tended to think, perhaps naively, that multiple infections would be a little unusual, because the timing would have to be fairly precise. Viruses generally rely on a couple of days of relative peace (immunologically speaking) to quickly replicate and bank a virus load that then keeps pace with the increasing immune response. If Virus B tries to infect you a couple of days after Virus A is already present, Virus B is going to run right into the thick of the immune response to Virus A, never have that chance to bank its progeny virus, and you’d expect it to be quickly overwhelmed. So you probably need nearly simultaneous infections to get a real multiple infection.

Chicago influenza poster 1918
“Influenza is prevalent” (Chicago, 1918)

But this is all speculation. A recent paper2 went out and actually looked for evidence of mixed infection in humans.3 They used previously-collected samples, and this is going to greatly underestimate the extent of mixed infection,4 but they did detect evidence of several mixed infections in their collection of over 1000 influenza samples. A plausible number they offer is about 0.5% of their samples — half a dozen individuals — were potentially mixed infections.5

(Later they suggest, as unpublished data, that the number may be as high as 3%. An important caution, that they don’t mention here, is that the 3% number is from influenza database analysis, and we know that these databases are not high quality — see On the accuracy of the influenza databases and the paper referenced therein6 — in fact, about 3% of the samples in the database are contaminated, so I don’t know if the present authors took this into account when interpreting evidence for mixed infections.)

However, sticking with the 0.5% figure — which is still remarkably high, and would represent tens of thousands of cases per year — they were able to look more closely at several of these samples and confirmed that they did, in fact, represent true mixed infections. This is another spinoff of the rapid, high-throughput sequencing that’s now becoming widely available. One patient, from New Zealand, was simultaneously infected with two viruses:

…one closely related to viruses cocirculating in New Zealand during 2004 and a second lineage that clustered with A/H3N2 viruses that became dominant in the following (2005) influenza season in the southern hemisphere 2

Another, in New York, was infected with two different influenza strains that are antigenically distinct — that is, viruses that would require different vaccines for protection. Remember that influenza vaccines are customized, year by year, to match up against the dominant circulating virus of that particular year. This patient would have needed two distinct vaccines to get adequate protection from his two infections.

A third, “even more dramatic” example was another New Yorker who was infected with two viruses that were not merely antigenically different, but that came from two distinct, broad groups — influenza A and influenza B viruses. I don’t think A and B can reassort, or at least the progeny would be very unlikely to be fit, but it illustrates that very mixed infection is quite possible.

It’s important to note that they were looking for mixed infection, not reassortment. Reassortment woud be much less common than mixed infection — you need mixed infectio nfor reassortment, but it’s not inevitable following mixed infection. Still, the background of mixed infection seems to be rather higher than I thought it would be.

In sum, we propose that mixed infection of diverse influenza viruses, a necessary precursor to reassortment, is a common occurrence during seasonal influenza in humans and will in turn accelerate the rate of adaptive evolution in this virus. In addition, intrahost populations of influenza virus will harbor genetic diversity generated by de novo mutation, which we have not assessed in the current study. As a consequence, we urge that intrahost sequencing be more routinely employed to assess the degree of genotypic and phenotypic diversity in populations of acute RNA viruses. With the advent of high-throughput next-generation sequencing platforms, viral variants are being much more explicitly revealed within specimens, and this type of data can be made available on a routine basis.2


  1. Offhand, actually, I don’t know how often reassortants have been identified. I’ll try to find that[]
  2. Ghedin, E., Fitch, A., Boyne, A., Griesemer, S., DePasse, J., Bera, J., Zhang, X., Halpin, R., Smit, M., Jennings, L., St. George, K., Holmes, E., & Spiro, D. (2009). Mixed Infection and the Genesis of Influenza Virus Diversity Journal of Virology, 83 (17), 8832-8841 DOI: 10.1128/JVI.00773-09[][][]
  3. It would probably be more interesting to look for mixed infection in swine, or wild ducks, but it’s only humans that have enough close attention to detect these relatively rare events.[]
  4. Most samples of a mixed infection are simply going to pick up the more abundant of the viruses present[]
  5. This comes with a large helping of caveats; it could over- or under-estimate the frequency. But it’s a reasonable starting point and they did confirm some of them.[]
  6. Krasnitz, M., Levine, A., & Rabadan, R. (2008). Anomalies in the Influenza Virus Genome Database: New Biology or Laboratory Errors? Journal of Virology, 82 (17), 8947-8950 DOI: 10.1128/JVI.00101-08[]
July 2nd, 2009

Simple, obvious, and wrong answers

Macrophage and mycobacterium
Macrophage phagocytosing mycobacteria

Sometimes the simple, obvious answer is right, and sometimes it’s completely backwards.

Tuberculosis was a terrifying, ubiquitous killer in the 19th century, but is relatively rare today (at least, in developed countries). The reason for the drop in Tb deaths isn’t entirely clear; it started with social factors probably including accidental or deliberate isolation of Tb patients, antibiotic treatment also knocked the disease back, and in some areas the vaccine (known as BCG) made a difference as well.

BCG is one of the oldest vaccines still in wide use; it was developed in the 1920s when a strain of Mycobacterium bovis (tuberculosis of cattle, contagious to humans) spontaneously lost virulence in culture. This avirulent strain of the bacterium was sent around the world and cultured independently, resulting in many distinct vaccine strains in different places and times. These strains are not only distinct genetically, but also phenotypically — they look different in culture, or grow differently, or whatever.

Over time, the vaccine has changed functionally, as well. Very early on the vaccine abruptly became even less virulent. More gradually, it seems that BCG has also become less effective; it’s no longer is able to protect against pulmonary Tb (although it’s still protective against other forms of the disease). Why is this?

At first glance this seems unsurprising. The bacterium has been grown in culture — outside of any animal host — for nearly 100 years. It’s had no selection to maintain its ability to grow in animals, or to avoid their immune responses, so of course it’s going to lose its ability to grow in animals.

But a recent paper1 suggests that exactly the opposite happened. Whether randomly, or because of some unexpected type of selection, the BCG strain has actually amplified an immune evasion function. This modern variant of the vaccine strain isn’t simply passively failing to induce an immune response; it’s actively suppressing the immune response.

Specifically, the authors argue that normal (wild, virulent) Mycobacterium secretes antioxidants as an immune evasion mechanism; that modern BCG also secretes lots of antioxidants; and that this is related to genomic duplications in some BCG strains:

Some BCG daughter strains exhibit genomic duplication of sigH, trxC (thioredoxin), trxB2 (thioredoxin reductase), whiB1, whiB7, and lpdA (Rv3303c) as well as increased expression of genes encoding other antioxidants including SodA, thiol peroxidase, alkylhydroperoxidases C and D, and other members of the whiB family of thioredoxin-like protein disulfide reductases.1

Further reading
Tb family trees
Conspicuous consumption
Life & Death, pre-vaccination

MycobacteriaIn other words, the long-term culture of BCG has yielded variants that are less immunogenic, because they are more actively suppressing the immune response. If their reasoning is correct, then reducing the antioxidant secretion from BCG should increase its immunogenicity. They took a BCG strain and deleted the duplicated antioxidant gene sigH (as well as the overexpressed SodA), and sure enough, the deleted version was more immunogenic and more protective in mice. “By reducing antioxidant activity and secretion in BCG to yield 3dBCG, we unmasked immune responses during vaccination with 3dBCG that were suppressed by the parent BCG vaccine.1

As a possible explanation, they note that their deletion variant also grows more slowly in culture than the “wild-type” BCG, and especially under certain culture conditions, and that this has led, coincidentally, to the reduced immunogenicity:

The practice of growing BCG aerobically with detergents to prevent clumping may have increased oxidant stress to cell wall structures and selected for increased antioxidant production. Then with each transfer the bacilli making more antioxidants represented a slightly greater proportion of the culture until they became dominant. In vivo, these mutations caused the vaccine to become less potent in activating host immunity. In effect, we believe that as BCG evolved it yielded daughter strains with an increased capacity for suppressing host immune responses. 1

If this turns out to be generally true, then there’s a relatively straightforward handle for converting BCG back into a more effective, and safer, vaccine; whereas if the reduced immunogenicity was because of over-attenuation, it’s not so simple — you’d be trying to make a vaccine more virulent, which is a tricky tightrope to walk.

Incidentally, I frequently complain about the terrible, terrible quality of press releases about scientific advances  (and therefore the terrible quality of much “science reporting”, which is basically regurgitating the terrible press releases) so I want to give props to the person at Vanderbilt University Medical Center who put together the release for this paper — it’s a clear, simple, interesting, and as far as I can tell accurate account of the finding, background, and observation.  It can be done well — I wish it was done this well more often.


  1. Sadagopal, S., Braunstein, M., Hager, C., Wei, J., Daniel, A., Bochan, M., Crozier, I., Smith, N., Gates, H., Barnett, L., Van Kaer, L., Price, J., Blackwell, T., Kalams, S., & Kernodle, D. (2009). Reducing the Activity and Secretion of Microbial Antioxidants Enhances the Immunogenicity of BCG PLoS ONE, 4 (5) DOI: 10.1371/journal.pone.0005531[][][][]
May 27th, 2009

How the aphid got its wings

Rosy Apple Aphid (Whalon lab)
Rosy Apple Aphid (Whalon lab, MSU)

While nothing can match the pure undiluted awesomeness that is the parasitoid wasp/bracovirus symbiosis,1 there are other symbioses that are at least in the same ballpark.  The latest one I’ve learned about is the relationship between a densovirus and the rosy apple aphid. 2  I can’t do better than to quote the abstract:

Winged morphs of aphids are essential for their dispersal and survival. We discovered that the production of the winged morph in asexual clones of the rosy apple aphid, Dysaphis plantaginea, is dependent on their infection with a DNA virus, Dysaphis plantaginea densovirus (DplDNV). Virus-free clones of the rosy apple aphid, or clones infected singly with an RNA virus, rosy apple aphid virus (RAAV), did not produce the winged morph in response to crowding and poor plant quality. DplDNV infection results in a significant reduction in aphid reproduction rate, but such aphids can produce the winged morph, even at low insect density, which can fly and colonize neighboring plants. Aphids infected with DplDNV produce a proportion of virus-free aphids, which enables production of virus-free clonal lines after colonization of a new plant.2

So without the virus, the aphids don’t grow wings, and they’re not able to disperse to new sites. When infected, they can sprout wings, and spread to a new site. Presumably without a flying aphid to carry them the virus can’t disperse, either.

Apart from anything else, my kids, having learned about this at dinner,3 are now hoping to have their wings turned on the next time they’re infected with a virus.


  1. Bioweaponized wasps shooting mutualistic immune suppressive viruses at their prey! Pew! Pew! Pew! []
  2. Ryabov, E., Keane, G., Naish, N., Evered, C., & Winstanley, D. (2009). Densovirus induces winged morphs in asexual clones of the rosy apple aphid, Dysaphis plantaginea Proceedings of the National Academy of Sciences, 106 (21), 8465-8470 DOI: 10.1073/pnas.0901389106[][]
  3. We have interesting dinner conversations at my house[]
April 14th, 2009

Tumor immunity: The Goldilocks approach

GoldilocksWe know that the immune system can destroy tumors. We also know, unfortunately, that by the time we see a tumor, immunity probably won’t destroy the tumor. There are lots of reasons for that. One is that tumors are essentially part of the normal body, so it’s normal for the immune system to ignore them. It looks as if you need to have immunity that’s just right to get rid of a tumor.

Tumors arise from normal self cells,1 that the immune response has been programmed to ignore. Now, the process of becoming a tumor is not normal, and so tumors are not entirely normal self any more — meaning that there are likely to be some targets in most if not all tumors. But in all but the most reckless tumors the differences between abnormal and normal are relatively small, compared to, say, a virus-infected cell that contains many potential targets.

There’s actually a long list of known tumor antigens; the T-cell tumor peptide database lists many hundreds of them. But most are not truly specific for the tumor.  The’re actually normal self antigens; they’re derived from proteins that are overexpressed in tumors, or that are differentiation antigens or “cancer-germline” antigens that are normally also found in self tissues. What’s more, these normal self antigens are the most interesting tumor antigens, as far as clinical utility is concerned. Mutations can make brand-new, non-self targets for the immune system, but they’re going to be sporadic targets, often unique to individual tumors — not something you can prepare for. The normal antigens, though, are likely to be predictable, common targets; it’s conceivable that tumor vaccines can be prepared in advance.

Melanoma cell (Eva-Maria Schnäker, University of Münster)
Human melanoma cell

If these antigens were common (which they are, in some tumor types — like melanoma), and they were good targets for the immune system, then we wouldn’t see much cancer. We do see melanomas quite often, and part of the reason may be that the immune system generally responds quite weakly to these antigens.  Why is that? And, more to the point, how can we make the immune system respond more strongly? A recent paper in the Journal of Experimental Medicine2 offers answers for both of these questions.

From work in the past couple of years, we now have decent estimates of how many T cells there are that can react with any particular target. (See here and here for my discussion of the earlier papers.) A reasonably strong immune response to a non-self epitope might originate from maybe 100 or so precursor T cells. There’s a rather wide range of frequency for these precursor cells, say from 20 to 1000; and to some extent, the fewer T cells there are the weaker (the less immunodominant) the immune response.

We expect T cells against normal self targets to be less common, because they should be eliminated as they mature in the thymus. Some may survive, though, and we would count on these survivors to attack the normal (albeit overexpressed, or abnormally present) target in the cancer cells. But just how rare are they?

Rizzuto et al say they’re really rare (this was in mice, by the way); at least ten times less abundant than T cells against non-self antigens.  If you look at the range I gave for “normal” precursors, that could mean there are fewer than 5 or 10 precursors.  If the average is “fewer than five”, then quite possibly some mice have only two, or one, or no precursors.  You can’t have much of a response with no precursors.

So there’s a weak anti-tumor response because there aren’t many T cells in the body that can respond to the normal self targets in the tumor. That’s not really a surprise, but it does raise the question, What if there were more of the T cells? To ask that question, Rizzuto et al. tried transferring more of these precursor T cells into tumor-bearing mice — starting at around the normal level for a precursor to non-self antigen, and going up from there — and then vaccinating with the appropriate target.

The effects were pretty dramatic. With no supplemental T cells (that is, with the natural, very low, level of T cell precursors) the mice all died of the tumor quickly. At the middle of the range, almost all of the mice rejected the tumor. And at the highest levels of transfers? The mice all died again. Having enough T cells to respond was protective, but putting in too many made them useless.

These results identify vaccine-specific CD8+ precursor frequency as a remarkably significant predictor of treatment and side-effect outcome. Paradoxically, above a certain threshold there is an inverse relationship between pmel-1 clonal frequency and vaccine-induced tumor rejection.2

Melanoma cell
Mouse melanoma cell

(My emphasis) This paradoxical effect is probably because the numerous T cells started to compete with each other so that none of them were properly activated; they only saw effective-looking polyfunctional T cells at the lower transfer levels.

In other words, if you’re going to transfer T cells to try to eliminate a tumor, more is not necessarily better. Quality and quantity are both important factors, and quantity helps determine quality.

One question I have is how this relates to tumor immune evasion. Many tumor types  acquire mutations, as they develop, that block presentation of antigen to T cells. Are these mutations perhaps only partially effective — giving the tumors sufficient protection against the tiny handful of natural precursors they “expect” to deal with, but not against a larger attack after, say, vaccination — or are they more complete, and protective even if the optimal number of T cells are transfered? I’d guess that it would depend on the tumor, but it looks as if it might be a relevant question and it would be nice to have more than a guess.

Our results show that combining lymphodepletion with physiologically relevant numbers of naive tumor-specific CD8+ cells and in vivo administration of an effective vaccine generates a high-quality, antitumor response in mice. This approach requires strikingly low numbers of naive tumor-specific cells, making it a new and truly potent treatment strategy.   2


  1. I’m ignoring here crazy things like the contagious tumors of Tasmanian Devils and dogs[]
  2. Rizzuto, G., Merghoub, T., Hirschhorn-Cymerman, D., Liu, C., Lesokhin, A., Sahawneh, D., Zhong, H., Panageas, K., Perales, M., Altan-Bonnet, G., Wolchok, J., & Houghton, A. (2009). Self-antigen-specific CD8+ T cell precursor frequency determines the quality of the antitumor immune response Journal of Experimental Medicine, 206 (4), 849-866 DOI: 10.1084/jem.20081382[][][]
March 12th, 2009

A successful trial of a malaria vaccine

Plasmodium and RBCThe point of a vaccine trial is to test whether the vaccine works.  If you get an answer to that question, the trial is a success.  The answer may be “No”, in which case the vaccine is a failure, but the trial would still be a success.  (The STEP HIV vaccine trial was therefore a success, though the vaccine was a failure.)

Malaria vaccines have been desperately needed forever, and in the past year there have been a few clinical trials. 1  An encouraging, though unspectacular, trial was reported last year, where the vaccine offered modest protection in children. 2

The most successful vaccines seem to be T-cell based, rather than antibody-based, and the latest report, of a Phase II trial in Kenya,3 drives another nail in the antibody/malaria coffin:

The FMP1/AS02 vaccine did not protect children living in Kombewa against first episodes of P. falciparum malaria; it did not reduce the overall incidences of clinical malaria episodes or of malaria infections, and did not reduce parasite densities … Because of the clearly demonstrated overall lack of efficacy in this trial, FMP1/AS02 is no longer a promising candidate for further development as a monovalent malaria vaccine. … We therefore propose that future MSP-142 vaccine development efforts should focus on other antigen constructs and formulations. 3

For more reading about immunity to malaria:


  1. I don’t actually know much about the history of malaria vaccines, as far as trials go, so I don’t know how unusual it is to have clinical trials.  People have been working on malaria vaccines for decades, but none have worked very well.[]
  2. Abdulla, S., Oberholzer, R., Juma, O., Kubhoja, S., Machera, F., Membi, C., Omari, S., Urassa, A., Mshinda, H., Jumanne, A., Salim, N., Shomari, M., Aebi, T., Schellenberg, D. M., Carter, T., Villafana, T., Demoitie, M. A., Dubois, M. C., Leach, A., Lievens, M., Vekemans, J., Cohen, J., Ballou, W. R., and Tanner, M. (2008). Safety and immunogenicity of RTS,S/AS02D malaria vaccine in infants. N. Engl. J. Med. 359, 2533-2544. doi:10.1056/NEJMoa0807773

    Bejon, P., Lusingu, J., Olotu, A., Leach, A., Lievens, M., Vekemans, J., Mshamu, S., Lang, T., Gould, J., Dubois, M. C., Demoitie, M. A., Stallaert, J. F., Vansadia, P., Carter, T., Njuguna, P., Awuondo, K. O., Malabeja, A., Abdul, O., Gesase, S., Mturi, N., Drakeley, C. J., Savarese, B., Villafana, T., Ballou, W. R., Cohen, J., Riley, E. M., Lemnge, M. M., Marsh, K., and von Seidlein, L. (2008). Efficacy of RTS,S/AS01E vaccine against malaria in children 5 to 17 months of age. N. Engl. J. Med. 359, 2521-2532. doi:10.1056/NEJMoa0807381[]

  3. Ogutu, B., Apollo, O., McKinney, D., Okoth, W., Siangla, J., Dubovsky, F., Tucker, K., Waitumbi, J., Diggs, C., Wittes, J., Malkin, E., Leach, A., Soisson, L., Milman, J., Otieno, L., Holland, C., Polhemus, M., Remich, S., Ockenhouse, C., Cohen, J., Ballou, W., Martin, S., Angov, E., Stewart, V., Lyon, J., Heppner, D., Withers, M., & , . (2009). Blood Stage Malaria Vaccine Eliciting High Antigen-Specific Antibody Concentrations Confers No Protection to Young Children in Western Kenya PLoS ONE, 4 (3) DOI: 10.1371/journal.pone.0004708[][]
January 14th, 2009

Why a vaccine failed, and maybe a fix

Jenner vaccinating a child
Jenner vaccinating a child

As I said last week, one of the biggest vaccine fiascos was the vaccine against respiratory syncytial virus (RSV) that was introduced in the 1960s. RSV is essentially a universal infection of children; it usually causes fairly mild respiratory disease, but because it’s so common the small fraction of cases that are more severe, end up being a leading cause of hospitalization for children. The vaccine was supposed to prevent that. As it happened, the vaccine itself didn’t cause any problems on its own; but children vaccinated with this RSV vaccine, who then later on were infected with RSV, actually had worse disease than those children who were uninfected. (Two children died.)

This enhanced respiratory disease (ERD) was really puzzling at the time, because the vaccine actually did induce a good, strong antibody response. But the antibody turned out to be non-protective. Just having an antibody response is not enough; the overall immune response needs to be involved and protective.

(I think we’re seeing some parallels to this concept now with T cell responses, where we are discovering that just having CD8 T cells doesn’t necessarily offer protection against things like HIV and hepatitis C virus, whereas the quality of the CD8 cells — now being measured as the range of cytokines they can produce — seems to be correlated with protection.)

The RSV vaccine turned out to trigger a TH2 type immune response. TH1/TH2 type responses are now a fundamental concept in immunology, but that hypothesis is a relatively new. Tim Mossman proposed it in 19861 and there was a significant lag before it was widely accepted. I think one of the findings that helped make TH1/TH2 accepted was the finding that the RSV vaccine triggered a strong TH2 immune response,2 compared to the actual virus infection which mainly causes TH1-type immunity. This — to me, anyway — abruptly made the paradigm look less like a laboratory curiosity only seen in mice, and more like a real, clinically important phenomenon.

ABCs of RSVSo the TH2 immune response seemed to more or less explain why the RSV vaccine caused disease. TH1 immune responses are generally protective against viruses, while TH2 immune responses are apparently more geared toward parasitic worms; TH2 responses tend to induce eosinophils and allergic-type responses, and that’s consistent with the clinical disease seen in the vaccinated children who got ERD.

But why did the vaccine induce a TH2 response? This is, of course, a huge question, especially if you’re trying to develop a new antiviral vaccine. One suggestion was the the vaccine screwed up the viral antigens too much. The vaccine used a formalin-inactivated virus, and the proposal was that the formalin alters the virus antigens and that directly caused the abnormal response3 If so, then this is a potential problem for any formalin-inactivated vaccine.

A new paper4 reaches a different conclusion. They say that formalin isn’t the main problem; rather, it’s the lack of adjuvant stimulation. Specifically, they say, you need to stimulate innate immunity via toll-like receptors (TLRs). Unless you do this, B cells don’t become completely activated, and though B cells produce antibodies the B cells don’t progress toward affinity maturation. That is, the normal process where antibodies are selected and shuffled to produce ultra-strong binders to their target antigens never gets underway. As a result, the vaccine induces low-affinity antibodies, and these low affinity antibodies are not protective.

It’s not clear — according to this model — whether the TH2 bias is actually the problem. Immune responses become biased to TH2 when there’s little innate immune stimulation, so the low affinity antibody and the TH2 response go hand in hand. Steve Varga (who has a nice commentary5 on this paper) has shown that some of the TH2 effects that were believed to be important in the pathogenesis of the ERD are not necessarily critical after all. Still, Varga and Delgado et al do seem to still feel that the TH2 shift is part of the disease.

The really exciting part of this finding is that it might actually be easy to fix. We now know a lot about TLR stimulation, and it should be possible to include TLR ligands along with the RSV vaccine:

These findings … open the possibility that inactivated RSV vaccines may be rendered safe and effective by inclusion of TLR agonists in their formulation. 4

Will this induce strong, protective immunity? Hopefully we’ll find out soon.


  1. Mosmann TR, Cherwinski H, Bond MW, Giedlin MA, Coffman RL. Two types of murine helper T cell clone. I. Definition according to profiles of lymphokine activities and secreted proteins. J Immunol 1986; 136: 2348-2357[]
  2. Priming immunization determines T helper cytokine mRNA expression patterns in lungs of mice challenged with respiratory syncytial virus. Graham BS, Henderson GS, Tang YW, Lu X, Neuzil KM, Colley DG. J Immunol. 1993 Aug 15;151(4):2032-40.[]
  3. A potential molecular mechanism for hypersensitivity caused by formalin-inactivated vaccines. Moghaddam A, Olszewska W, Wang B, Tregoning JS, Helson R, Sattentau QJ, Openshaw PJ. Nat Med. 2006 Aug;12(8):905-7.[]
  4. Maria Florencia Delgado, Silvina Coviello, A Clara Monsalvo, Guillermina A Melendi, Johanna Zea Hernandez, Juan P Batalle, Leandro Diaz, Alfonsina Trento, Herng-Yu Chang, Wayne Mitzner, Jeffrey Ravetch, José A Melero, Pablo M Irusta, Fernando P Polack (2008). Lack of antibody affinity maturation due to poor Toll-like receptor stimulation leads to enhanced respiratory syncytial virus disease Nature Medicine, 15 (1), 34-41 DOI: 10.1038/nm.1894[][]
  5. Steven M Varga (2009). Fixing a failed vaccine Nature Medicine, 15 (1), 21-22 DOI: 10.1038/nm0109-21[]
November 19th, 2008

Electronic notebooks

Cavemen (Life archives)A couple of years ago I published a paper characterizing a mutant cell line. 1  I had been working, on and off, on the cells for around ten years, and they were already present in the lab when I joined it.  To write the paper I needed to know the details of their generation.  I clambered the ladder to the box marked “1992 LAB BOOKS”, pulled out Ethan’s notes for the year, flipped through them for a few minutes, and copied down the procedure — concentration of EMS, duration of treatment, and so on.  

Since 1992 I’ve used electronic data stored on 5¼-inch floppies, 3½-inch floppies (single and double-sided), Bernoulli drives, zip drives, Jazz drives, CDs, DVDs, and USB flash sticks; as well as on computer hard drives from at least four different OSes, and in God knows how many formats.  

The data on at least five of those media are now almost entirely inaccessible to me (if we were desperate, I’m fairly sure we could retrieve them, but it would be a huge chore).  Probably more than half of the different formats are almost unreadable today.  

Meanwhile, the data in those old-fashioned paper notebook are just as usable today as they were in 1992; and they will be equally usable in another sixteen years.   

I’m seeing a lot of discussion online about electronic lab notebooks, but this is an aspect that I don’t think has been emphasized nearly enough.  I know when you plan an experiment, you expect to publish it (in Nature) next week; but that’s not what always happens, is it.  And even if you do publish in a timely manner, who know what’s going to happen in fifteen years?  (I just thawed out some cells, frozen by a colleague in 1985, to analyze their antigen presentation pathways; something he had no interest in at the time.  He still has his lab notebooks describing his characterization, though, including stuff he didn’t publish at the time.)

Searchable experiments
A crude searchable experiments interface

How many of the protocols out there today are going to be functional in 15 years?  How many web sites from 1992 are still readable today?  (Since HTML wasn’t specified until 1993, the answer is “Not many”.)  History suggests that those electronic notebooks of today will be the impenetrable floppy disks of tomorrow. 2

Electronic notebooks do have one gigantic advantage over paper: Search.  I do use electronic notebooks of one kind or another, and the main reason is so I can search for the half-remembered experiment that used brefeldin A, and find out what concentration.  For years I’ve just used a cobbled-together thing I wrote myself, a HTML interface to an SQLite database linked with a Python cgi script (e.g. the screenshot to the right; click for a larger version).  It works nicely for searching, but it’s not as future-proof as I’d like (it depends on Python, which is being updated to a partially incompatible version soon; SQLite, which is likely to be stable for a few years, but I’m not counting on fifteen; and html, which is evolving as well.)  As well, it’s a little irritating to not have real data in there; so in the past year or so I’ve started using a wiki to keep lab notes in as well.  

I’ve actually made multiple false starts at the wiki/notebook thing, and there’s no guarantee that this latest version will stick, but it’s looking more promising than previous runs.  I’m using DokuWiki, which uses flat text (marked up) files for each page. I trust txt to be readable in 10 or 15 years, so even if (when) the rest of the interface is incompatible there should be usable information there.  It’s also easy to back up, and the wiki in general seems friskier and more responsive than some of the other wikis I’ve looked at.  I’m reasonably sure this will work.

But I’m still backing up to a paper lab notebook, because I know that works.


  1. York IA, Grant EP, Dahl AM, Rock KL (2005). A mutant cell with a novel defect in MHC class I quality control. J Immunol 174:6839–6846. []
  2. Note that I haven’t looked in any detail at the electronic notebooks of today, and really have no idea how future-proof they are.  This is just my prejudice.[]
November 16th, 2008

Slow death, fast death

 

Death and the Doctor
“Death and the Doctor”
Published by William Humphrey, 1777 

Last April I commented on a series of experiments  that used intravital microscopy to visualize cytotoxic T lymphocytes (CTL) attacking a tumor. 1 Immensely cool though the movie is, I noted that I was surprised by their estimate of the rate of cell killing:

Another surprising finding — which is so different from previous work in different systems that I’m hesitant to believe it — is the timing of cell killing. Previous studies (such as the von Andrian paper2 that produced this video) have suggested that CTL kill their targets in something under an hour; maybe 30 minutes or even less. Here. Bousso’s group find that the tumor cells take something like 6 hours to be killed. That’s such a large difference — and has such important implications for effectiveness of CTL killing — that, as I say, I’d like to see it confirmed before I take it to the bank.3

A new paper4 has run another estimate of the time it takes for a CTL to kill its target, and like most of the previous work, they conclude that it takes about a half-hour, give or take, to kill a target. They do come up with a fairly wide range of killing times, that depend on the target and the timing of the immune response — at the peak of the immune response when there are many cells the targets are killed faster (between 2 and 14 minutes), while at later stages, when there aren’t so many CTL, targets have half-lives of 48 min and 2.8 hr.

CTL killing a target
CTL killing a target cell
(From a video by von Andrian)
 

This is not quite looking at the same thing as the video showed, though. In this paper, they were looking at the bulk effects, and that’s what almost all the previous studies have also looked at. The video was looking at a one-on-one interaction. What if targets are killed faster when several CTL gang up on them? Here, having different numbers of CTL caused the half-life of the targets to increase between about 10 and 20-fold. But this is probably simply because, with fewer CTL present, it took longer for them to find the target: Once a CTL found the target, the rate of killing was if anything faster than effectors at killing (“we find that LCMV-specific memory CD8 T cells kill more target cells per day than effectors”). 5

This is actually a disagreement with a previous paper 6 that also looked at killing rates, and offered evidence that different types of CTL can have different killing rates:

We reanalyse data previously used to estimate killing rates of CTL specific for two epitopes of lymphocytic choriomeningitis virus (LCMV) in mice and show that, contrary to previous estimates the “killing rate” of effector CTL is approximately twice that of memory CTL. 6

However, whichever of those studies is correct , both suggest that different types of CTL can have different killing efficiencies. This goes back to a point I’ve made several times, as have others (see e.g. Michael Palm’s TAG post here and references therein, including the comments by me and by Otto Yang) — CTL aren’t a uniform batch, and different kinds of CTL may have different types as well as rates of activities.

Returning to the intravital microscopy killing rate of 6 hours:7 I wonder if that reflects the nature of the CTL there, perhaps influenced by the tumor environment. Tumors are notoriously resistant to killing (probably because those tumors that are not resistant to killing were, um, killed, before they ever become clinically detectable) and it seems quite likely that an immunosuppressive tumor environment may change CTL types, or activities. I wonder if that would offer some way of intervention. Speeding up the rate of CTL killing from 6 hours to 30 minutes seems like it would be a huge influence of clearance of tumors. On the other hand, of course, it may be that the targets themselves are much more resistant to killing (again because tumor cells have been through selection to be resistant to the immune system) and cranking up CTL won’t make much difference.


  1. Breart, B., Lemaître, F., Celli, S., Bousso, P. (2008). Two-photon imaging of intratumoral CD8+ T cell cytotoxic activity during adoptive T cell therapy in mice. Journal of Clinical Investigation, 118(4), 1390-1397. DOI: 10.1172/JCI34388 []
  2. Mempel, T. R., Pittet, M. J., Khazaie, K., Weninger, W., Weissleder, R., von Boehmer, H., and von Andrian, U. H. (2006). Regulatory T cells reversibly suppress cytotoxic T cell function independent of effector differentiation. Immunity 25, 129-141.[]
  3. From this post[]
  4. V. V. Ganusov, R. J. De Boer (2008). Estimating In Vivo Death Rates of Targets due to CD8 T-Cell-Mediated Killing Journal of Virology, 82 (23), 11749-11757 DOI: 10.1128/JVI.01128-08[]
  5. There are also other videos of one-to-one killing, at least in vitro, that are more consistent with the 30-minute ballpark; see the image to the right for one example.[]
  6. Yates A, Graw F, Barber DL, Ahmed R, Regoes RR, et al. (2007) Revisiting Estimates of CTL Killing Rates In Vivo. PLoS ONE 2(12): e1301. doi:10.1371/journal.pone.0001301[][]
  7. Which I have become more relaxed about since my earlier skeptical comment[]
October 6th, 2008

Sex, stats, and sweat

Sweaty t shirtIt’s been suggested for a long time that mice select mates by smelling MHC types, perhaps in the urine. MHC is by far the most variable region in vertebrate genomes, so this would offer a way for mice to avoid inbreeding: The more related the mice, the more likely they are to be similar at the MHC, so selecting a different MHC will help avoid inbreeding.

Partly as an argument by analogy, and partly through some rather poor-quality experiments, it’s also been argued that humans select mates the same way — that differences in MHC type make a partner more desirable. These are the notorious sweaty T shirt experiments that most people seem to have at least vaguely heard of.

I started off very skeptical about the human claims, because the quality of the experiments has, as I say, tended to be poor. There have been small numbers of people, indifference to alternative explanations, and a lot of post hoc hand-waving. (If the preferences turned out to be reversed, why, it was because the female was near her period, or something like that.) I think that most people who have actually looked at the data have had similar reservations, but that hasn’t stopped the concept from becoming pretty well known.

MHC & mate choice I became even more skeptical about the human experiments as I learned more about the mouse data. The evidence for MHC as a mechanism for avoiding inbreeding turned out to be relatively weak, or at least inconsistent (see here for my first discussion); and recently a paper that I found fairly convincing (discussed here) suggested that MHC is not in fact used by mice in this way at all — rather, a much more plausible, highly variable family of molecules called “major urinary proteins” (MUPs) are the source of the anti-inbreeding odor in mouse urine.

Much of the interest in human MHC and sex has been driven by the mouse observation, so I think that if mice don’t use MHC to select mates, then likely humans don’t, either. Still, it remains possible, even probable, that difference species use different methods to select mates. And since humans don’t even have variable MUPs (as far as I know) MHC remains in the chase.

A recent paper1 tries to look at this in a more objective manner, using genome-wide data on couples. Unfortunately the numbers are still quite small (just 30 couples each from a European-American subset, and an African subset) and the results remain slightly ambiguous. Their conclusion was that

African spouses show no significant pattern of similarity/dissimilarity across the MHC region … We discuss several explanations for these observations, including demographic effects. On the other hand, the sampled European American couples are significantly more MHC-dissimilar than random pairs of individuals … This study thus supports the hypothesis that the MHC influences mate choice in some human populations.

So, heads we win, tails you lose, because even though their hypothesis was invalidated overall, some post-hoc wiggling (“demographic effects”) lets them dismiss the data they don’t like.

I’m still pretty skeptical about any real effect from MHC on mate choice. I’m willing to be convinced otherwise, but it’s going to take a larger and more rigorous study than this one to make me interested.


  1. Raphaëlle Chaix, Chen Cao, Peter Donnelly, Molly Przeworski (2008). Is Mate Choice in Humans MHC-Dependent? PLoS Genetics, 4 (9) DOI: 10.1371/journal.pgen.1000184[]