Mining HIV data

1
Mining HIV data

• Koen Deforche
• Department of Evolutionary Virology, Faculty of
  Medicine, KULeuven

2
Topics

• HIV
• The problem anti-retroviral therapy failure
• Looking for
• (1) Better interpretation of resistance
• (2) Other causes for anti-retoriviral therapy
• Illustration of (1)
• Using BN structure learning to learn dependencies
  between amino acids in protease

3
HIV

• Retro virus
• Relatively small genome (10 Kbp)
• High replication rate high mutation rate
• Probably most sequenced organism
• Many open questions
• How does HIV cause AIDS ? And why not always ?
• How does HIV kill our immune system ?

4
The problem

• Seemingly inevitably, anti-retroviral drug
  therapy fails after a while.
• Limited number of drugs (20).
• Limited number of attacks (3, PI, NRTI, NNRTI).
• Simple, accute question what to do when patient
  fails? Change therapy? What drugs?

5
Current SOC

• Do population genotyping of HIV genes that code
  for drug target proteins, on plasma sample.
• Use other expert knowledge past therapy history,
  patient drug adherence,

Algorithm (for drug X)
Drug susceptibility score (for drug X)
Genotype
6
(1) Resistance algorithms

• Built (also at our lab) using expert knowledge
• Contradicting results very subjective
  construction !
• Data mining was used to predict in vitro
  susceptibility based on genotype (Beerenwinkel et
  al. (decision trees), Larder et al (neural
  networks))
• Does not account for evolution
• Different environment in vivo
• Expert rule systems proven more accurate for in
  vivo behaviour (!)

7
Data sets!

• Free 4300 non-annotated sequences in Los Alomos
  database.
• Not free, but available from patient medical
  records
• (genotype, last therapy ,previous therapies,
  viral load, CD4 count)
• Harder to come by
• (genotype before last therapy, last therapy,
  viral load at control points)

8
Example of things missed now

• Interactions between amino acids
• Interesting on their own
• Needed to understand subtype behaviour
• Evolutionary distance playing by the rules of
  Darwin and Kimura
• Evidence for primary versus secondary mutations

9
(No Transcript)
10
Bayesian network around 89
11
HIV-1 protease
12
(2) Other factors

• Observation gt1/3th of therapy failures with PIs
  could not be explained for by my BN.
• Not significantly more mutations than naive
  sequences.
• Hypotheses
• Some naive sequences may well be resistant (not
  likely)
• No drugs in blood
• Drug adherence 1/3th of patients does not take
  their pills!
• Drugs are cleared by metabolism
• Drug-drug interactions

13
Data set

• Close cooperation with UZLeuven all medical,
  epidemeological and virological data for Leuven
  HIV patients (/- 300)
• Social background
• Drug adherence estimate from clinician
• Other drugs
• Drug dosage
• Drug adherence project
• Relational database

14
ViroDB
15
Conclusions ?

• Data mining can add (valuable) new insights on
  HIV and anti retroviral therapy response.
• You have worked too hard.

16
Thank you!