Keynote: Thomas Lengauer - Chasing the AIDS Virus - ISMB/ECCB Stockholm 2009

Background in mathematics and Computer Science - Allyson Lister

protein bioinformatics, computational drug screening and design - Ruchira S. Datta

previously full prof at U. Paderborn - Ruchira S. Datta

on steering board of ECCB since its founding - Ruchira S. Datta

an exciting story, downstream of the bulk of computational biology in the medical field - Ruchira S. Datta

process flow usually ends with finding and optimizing potential drug targets - Ruchira S. Datta

Start when the drugs are available on the marketplace and they support personalized medicine, and which drugs to give to AIDS patients. - Allyson Lister

Personalized medicine, they start when the drugs are in the market place - Diego M. Riaño-Pachón

in this case, support difficult decision of a doctor: what drug to give to the AIDS patient - Ruchira S. Datta

33m HIV infected patients in 2007 - Peter Menzel

25 M deaths since 1981 - Gabriele Sales

greatly affected in Africa - Ruchira S. Datta

Aids awareness is waning - Oliver Hofmann

Europe appears little affected, but this may be deceptive - Ruchira S. Datta

...and increasing infection rate nowadays - Peter Menzel

AIDS awareness campaigns have waned in recent years, and as a consequence there is an increase in infection rates again. - Allyson Lister

At least in Germany, AIDS awareness is reducing - Diego M. Riaño-Pachón

AIDS is rampant and almost a lost cause. AIDS is nowhere near under control. - Ruchira S. Datta

'AIDS almost a lost cause' (no way it is under control currently) - Oliver Hofmann

small molecule: 10000 letters - Ruchira S. Datta

HIV virus: small genome (10k bases) - Gabriele Sales

double single-strand RNA genome - Gabriele Sales

virus attaches via surface proteins to the T-cell - Ruchira S. Datta

HIV has a duplicated single-stranded RNA genome - Peter Menzel

transferred into nucleus via viral protein called integrase - Ruchira S. Datta

no way to get the virus out of an infected cell - Ruchira S. Datta

It's not possible to remove the virus from the infected cell. - Gabriele Sales

viral particles assemble in a complex process, not completely understood - Ruchira S. Datta

during maturation process, strings of proteins are cut into their functional parts - Ruchira S. Datta

viral protein protease makes virus effective at infecting a new cell - Ruchira S. Datta

AIDS virus is by far the best understood of all viruses - Ruchira S. Datta

Drug design starts with understanding the life cycle. Best understood virus. - Oliver Hofmann

life cycle of HIV is already understood very well - Peter Menzel

HIV is the best understood virus - Diego M. Riaño-Pachón

Various drugs blocking different phases of the virus infection. - Gabriele Sales

There are a number of drugs that blocks the fusion of the virus with the cell, 17 blocking reverse transcriptase, etc. - Allyson Lister

17 drugs blocking transcription - Diego M. Riaño-Pachón

one new drug blocks attachments of virus to cell; another blocks fusion; 17 drugs block reverse transcription; 1 drug blocks integration; 8 block maturation - Ruchira S. Datta

17 or 70? :) - Allyson Lister

will explain why just one drug doesn't suffice - Ruchira S. Datta

HIV is extremely dynamically evolving, possibly the most dynamically evolving virus known - Ruchira S. Datta

17 :-) - Peter Menzel

moving target: over 10 million virus particles turned over per day per patient - Ruchira S. Datta

A drug may be efficient against the wild type, but not against mutants. - Gabriele Sales

wild type viruses are most fit under natural condition; drug will be very effective on wild type but will very quickly select for resistance - Ruchira S. Datta

Always going to be minority variants resistant to the drug which will escape drug treatment - Oliver Hofmann

reverse transcriptase is a bad copier, enabling variation every time the virus replicates - Ruchira S. Datta

There is no drug targeting all mutants. - Gabriele Sales

Hence drug cocktails that catch all variants. Doesn't work, best we can do is postpone the onset - Oliver Hofmann

therefore need drug cocktail that catches all of them, but this is utopia and doesn't happen; there is no drug therapy that works forever - Ruchira S. Datta

The virus always wins, we can only postpone the loss - Diego M. Riaño-Pachón

HAART: highly active anti-retroviral therapy; administer at least two drugs of different classes (targetting different proteins, working in different ways) - Ruchira S. Datta

number of viral RNA in the blood is a major clinical indicator - Peter Menzel

therapy is effective for some time, until new strain develops that is resistant - Ruchira S. Datta

50 copies is the current limit of detection for blood tests (unclear per what?) - Oliver Hofmann

Detection limit: 50 copies of the virus. - Gabriele Sales

this is the main question in treating patients, and is very difficult - Ruchira S. Datta

so far only viral genome, not host genome, is being considered - Ruchira S. Datta

people have built mutation tables: synopsis of global clinical experience of how virus responds to treatment - Ruchira S. Datta

In the past, they've built mutation tables - global collection of clinical experience - Allyson Lister

Mutation tables: collection of responses of the virus to various treatments. - Gabriele Sales

e.g., protease inhibitors - Ruchira S. Datta

Gotta love the overlap in comments! :) - Allyson Lister

Allyson: heh - Ruchira S. Datta

@Allyson: too many bloggers here ;-) - Gabriele Sales

(I am now waiting before posting.. only to see everyone else waiting, too... gah) - Oliver Hofmann

An expert group will build this table. - Allyson Lister

Mutation table created by a group of experts, after heated discussions - Diego M. Riaño-Pachón

a particular SNP suggests that the virus is resistant to Atanovir - Ruchira S. Datta

I say just go with the flow and just get some overlap - it's good! - Allyson Lister

Allyson: knowledge of the crowd.. :-) - Peter Menzel

can't collect a lot of data: if have seen resistance, don't want to subject future patients to this therapy - Ruchira S. Datta

@Peter - definitely! - Allyson Lister

but if there are too many mutations in the table, won't be able to administer therapy to any patient--every patient will have some of these - Ruchira S. Datta

Tables limited because mutations are not acting independently. - Gabriele Sales

mutation tables carry not enough information.. - Peter Menzel

interdependencies cannot be captured by mutation tables; need rule-based expert systems - Ruchira S. Datta

-> expert systems ? - Peter Menzel

Mutations act in the context of the remaining genome (and the host genome) - Oliver Hofmann

Mutation table ignores the context of mutations and synergies - Diego M. Riaño-Pachón

unfortunately, the medical community calls these "algorithms" - Ruchira S. Datta

virologists ask: "Is this kind of resistance analysis objective?" "Can we not let the clinical data speak for themselves?" i.e., circumvent political process of decisionmaking of what goes in the tables - Ruchira S. Datta

Started by building a clinical database. - Gabriele Sales

Then the comp biol at his group enter, by request from MDs, into the picture - Diego M. Riaño-Pachón

at the start, no clinical database existed - Ruchira S. Datta

Avoiding community decisions by querying clinical databases.. which did not exist at the beginning of the project - Oliver Hofmann

phenotypic data: extract patient's virus and expose it to drugs - Peter Menzel

phenotypic data: expose the virus to different drug concentrations in cell culture - Ruchira S. Datta

observe fitness of resistant vs wild type curve of how much drug needed to suppress replication - Ruchira S. Datta

measure the "resistence factor": how much more drug is necessary to keep a mutant under control as the wild type - Gabriele Sales

One value: drug increase required to overcome mutation effect - Oliver Hofmann

quantify as "resistance factor" - Ruchira S. Datta

map viruses to drug concentrations that are effective - Peter Menzel

only a few labs can do this kind of analysis - Ruchira S. Datta

but this data is too expensive and too slow to make for clinicians - Allyson Lister

people have resorted to viral genome, as sequencing the viral genome is easy and fast - Ruchira S. Datta

bioinformatic resistance analysis can replace phenotypic lab test - Ruchira S. Datta

now: sequencing individual virus' genomes - Peter Menzel

Multivariate statistical learning approaches - Oliver Hofmann

multivariate statistical learning on db of genotype-phenotype pairs - Ruchira S. Datta

Training on a genotype-phenotype pairs database (1000+ HIV variants). - Gabriele Sales

The training data is the genotype-phenotype pairs of 1000+ HIV variants - Allyson Lister

"Every HI virus is ugly" - Peter Menzel

want statistical model that will either regress or classify - Ruchira S. Datta

quality criteria: predictive power, interpretability - Ruchira S. Datta

A statistical model that regresses or classifies. Statistical power is not the only factor, results need to be interpretable - Oliver Hofmann

doctors in the field want interpretable models, and will sacrifice a few % accuracy for this - Ruchira S. Datta

regression: estimate resistance factor - Ruchira S. Datta

Clinicians require interpretable models, why they make the predictions they do - Diego M. Riaño-Pachón

Classification into two classes: susceptible or not. - Gabriele Sales

classification based on cutoff values - Ruchira S. Datta

interpretable model is a decision tree - Ruchira S. Datta

Decision tree classifier: tease out interdependence between different mutations - Oliver Hofmann

along branches of decision tree, query different amino acid positions - Ruchira S. Datta

Decision tree encoding the ammino acids conferring resistance. - Gabriele Sales

Beerenwinkiel et al. PNAS 2002 99 (12) 8271-6 - Allyson Lister

much more informative than mutation tables. - Peter Menzel

virus can be resensitized by multiple mutations - Ruchira S. Datta

Allows to find re-sensitization (sp?) effects - Oliver Hofmann

one decision tree for each drug - Diego M. Riaño-Pachón

Web tool: www.geno2pheno.org - Gabriele Sales

http://www.geno2pheno.org, so far most used clinical tool by them - Ruchira S. Datta

one particular patient was a difficult case, convinced doctors that this tool has some use - Ruchira S. Datta

Genotype is aligned to the wt and mutations are identified - Allyson Lister

Identifies genome variations (alignment to the wild type) - Oliver Hofmann

on server, using regression with linear SVMs, not classification with decision trees - Ruchira S. Datta

now uses SVMs instead of decision tree - Peter Menzel

List of drugs by estimated effect (based on SVM). - Gabriele Sales

gives estimated resistance factor - Diego M. Riaño-Pachón

Using linear SVM for regression: a line for each drug and have est resistance factor, and normalization with Z-score, and the scored mutations. - Allyson Lister

use z-scores, as absolute values of resistance factors are not comparable btw drugs - Ruchira S. Datta

Each drug with an estimates resistance factor, Z-score (for comparative purposes) and a list of scored mutations based on their weight - Oliver Hofmann

this difficult patient is full of mutations, has resistance to every known drug per the mutation table - Ruchira S. Datta

"out-therapy" - doctors say positively they can't help him any more - Ruchira S. Datta

but they saw that some of these mutations actually resensitize! - Ruchira S. Datta

Some mutations, which confer resistance to some things (e.g. 76V in the anecdotal example) actual confers re-sensitisation and therefore would have a positive effect. Couldn't have been done with mutation tables! - Allyson Lister

Give one drug to retain re-sensitation mutation, add second drug to exploit the re-sensitation effect - Oliver Hofmann

one mutation conferring resistance to two drugs, resensitized the virus to other two drugs - Diego M. Riaño-Pachón

this could not have been found via the mutation table; the patient was on the recommended therapy from March 2003 until April 2009 and blood was clear of virus - Ruchira S. Datta

natural next question: predict in what direction the virus will evolve under a given drug therapy - Ruchira S. Datta

Now: model the viral evolution - Peter Menzel

Next question is: how the virus will evolve in reaction to a certain drug? - Gabriele Sales

not possible by mutation table - Ruchira S. Datta

hacking viral evolution neat. :) - Nav

The virus does not change randomly; it follows specific mutational paths. - Gabriele Sales

The virus "chooses" mutation paths, do not know why - Diego M. Riaño-Pachón

simulated by fair mutations, but the virus does not mutate by flipping a fair coin, it chooses useful mutations (!) don't know how it does that - Ruchira S. Datta

virus follows specific mutational paths into resistance - Ruchira S. Datta

('chooses' is probably not the right word for this :) ) - Oliver Hofmann

want to find such paths in the database - Ruchira S. Datta

Oliver: hopefully not, it's very strange - Ruchira S. Datta

What are the preferred directions the virus goes? - Diego M. Riaño-Pachón

Not enough data to make accurate models. - Gabriele Sales

Longitudinal data missing - Oliver Hofmann

would like longitudinal data, but have cross-sectional data: lots of patients, but few data points on each one - Ruchira S. Datta

used to build mutagenetic trees - Ruchira S. Datta

Viral evolution is modeled using tree structures. - Gabriele Sales

the TAM1 path is found by seeing the virus does *not* follow every possible path - Ruchira S. Datta

mutagenetic trees with probabilities on branches - Peter Menzel

method derives mixtures of mutagenetic trees, not single ones - Ruchira S. Datta

22% of data follow one tree, 78% another - Ruchira S. Datta

can make time models, to estimate average time until a mutation is acquired - Ruchira S. Datta

predict length of tunnel created for the virus by drug therapy, on its path to resistance - Ruchira S. Datta

Estimate probability that a virus will acquire a certain resistance within a given time - Oliver Hofmann

applet predicts which therapy appears most promising - Ruchira S. Datta

Can try to maximize that duration - Oliver Hofmann

Now: Algorithm calculates success probabilities for a certain therapy suggested by the software - Peter Menzel

Do they actually follow the other tree? Could they be transitioning through the other mutations faster than you are sampling? - Nav

therapy optimization with THEO - Ruchira S. Datta

this strange sampling of effective paths reminds me of quantum computing - Ruchira S. Datta

Doctors still do not trust computers to the very end. Good! - Peter Menzel

EuResist: Europe-wide collection of resistance data - Ruchira S. Datta

Rosen-Zvi, Altmann et al BIoinformatics, 2008 - Gabriele Sales

server akin to geno2pheno on the internet; ROC curves show it has better performance than expert systems - Ruchira S. Datta

Most 'fun' consortium he's been involved in (would love to know the criteria for that :) ) - Oliver Hofmann

so does THEO - Ruchira S. Datta

http://www.euresist.org/ - Peter Menzel

German database of ineffective therapies (thanks to honesty of doctors) - Ruchira S. Datta

without THEO, chance of failure is >24%, without <10% - Ruchira S. Datta

Error in therapy classification: 24% without THO, below 15% with it. - Gabriele Sales

geno2pheno accessed from 30 countries - Ruchira S. Datta

2/3 AIDS patients in Germany treated with geno2pheno server - Peter Menzel

also need to keep finding new drugs - Ruchira S. Datta

Constant need new drugs as the virus _will_ evolve with time towards resistance - Oliver Hofmann

one target: the viral entry - Ruchira S. Datta

the cell cooperates in drawing the virus in - Ruchira S. Datta

nice movie showing HIV infection of the cell. Wish I could link that - Oliver Hofmann

CD4 receptor and coreceptor on human cell; first GP120 attaches to CD4, then to coreceptor - Ruchira S. Datta

then virus drills down and the particles fuse - Ruchira S. Datta

so coreceptor protein will be the target; there is one drug targeting this, from Pfizer, Maroviroc (sp?) - Ruchira S. Datta

some people cannot be infected. - Allyson Lister

1% of Caucasian population does not have this coreceptor, and cannot be infected with HIV - Ruchira S. Datta

1% of caucasians do not have the coreceptor - Peter Menzel

Close enough: http://www.scienceinreview.com/2007... - Oliver Hofmann

some viral variants can use a different coreceptor (CCR5 vs CXCR4) - Ruchira S. Datta

once you're in therapy the virus can switch - Allyson Lister

people with the CCR5 deletion don't get AIDS, so the virus first goes through here, but it later switches to the other one - Ruchira S. Datta

@Oliver, thanks for the link - Diego M. Riaño-Pachón

have duotropic strain; need genotypic assay - Ruchira S. Datta

have genotypic prediction of viral tropism; this server has the most hits - Ruchira S. Datta

the lab is in San Francisco; German doctors don't want to send samples all the way there and are incented to use the genotypic test - Ruchira S. Datta

35 amino acids in V3 loop; used to look at residues 11 and 25, but multivariate server does much better - Ruchira S. Datta

supply method with structural descriptor of the V3 loop, and use it in the predictor, which increases the sensitivity - Ruchira S. Datta

results shown were toy, based on clonal data - lab sample of *single* virus strain - Ruchira S. Datta

Viral sequences are ambiguous in numerous positions. - Gabriele Sales

in patient have several strains, so ambiguous base calls; in practice, these are "bulk data" and greatly reduce sensitivity - Ruchira S. Datta

therefore add clinical correlates such as virus load, CD4 load, etc. that are easily drawn from patient - Ruchira S. Datta

Add extra information -- clinical parameters -- to increase sensitivity - Oliver Hofmann

still phenotypic test gives much more accurate result, so clinical value is contested - Ruchira S. Datta

Sensitivity goes from 80% to 40% if you move from clonal to bulk data. - Allyson Lister

study shows that in clinical picture, no difference btw genotypic and phenotypic test - Ruchira S. Datta

go away from sanger sequencing to increase accuracy - Allyson Lister

still not satisfying, so want to increase accuracy - Ruchira S. Datta

move away from Sanger sequencing; if virus occurs in only 10%, won't even see it - Ruchira S. Datta

therefore use deep sequencing - Ruchira S. Datta

Sanger sequencing can now be replaced by deep sequencing, solving ambiguities. - Gabriele Sales

need to assemble 454 reads - Ruchira S. Datta

coreceptor of 35 aa can be resolved by single read - Ruchira S. Datta

ultra-deep sequencing for getting sequence information for the whole quasi species - Peter Menzel

Up to 130k sequences (reads?) from a single drop of blood - Oliver Hofmann

elbow shaped curve in one case, less of one coreceptor type => drug against coreceptor likely to be effective - Ruchira S. Datta

need to establish cutoffs; with prediction specificity of 90%, size of X4 minority must exceed 5%; call this R5 and use the drug - Ruchira S. Datta

o/w, call it X4, drug will be ineffective - Ruchira S. Datta

but need to choose parameter values - Ruchira S. Datta

There might be better descriptors, but establishing these in the community is going to be difficult - Oliver Hofmann

Zero delay between results and bedside application (unlike traditional drug development with lead times of 10+ years) - Oliver Hofmann

Niko Beerenwinkel worked on haplotype prediction, now at ETH Zürich - Ruchira S. Datta

this is not just academic software, need this to be highly available, not just dependent on grad students - Ruchira S. Datta

Rolf Kaiser, w/o exposure to computers, conceived of this project; thankful for his vision - Ruchira S. Datta

formed society for furthering the software, stable under various losses of funding - Ruchira S. Datta

Blog post: http://themindwobbles.wordpress.com/2009... - Allyson Lister

(only to get the blog on top of the ISCB portal site) - Reinhard Schneider