Transcript Estimating and forecasting in vivo drug disposition and

Estimating and forecasting in
vivo drug disposition and effects
using distributed computer
systems
Niclas Jonsson
Division of Pharmacokinetics and Drug Therapy
Department of Pharmaceutical Biosciences
Uppsala University
Pharmacometrics
• Describes the dynamic interaction between
drugs and individuals using quantitative
models.
• The models are highly non-linear and are
based on pato-physiological and
pharmacological knowledge.
• The models are used in clinical drug
development with the goal to make more
efficient use of available data and to optimize
future clinical trials.
Cross-disciplinary field
Pharmacokinetics,
Pharmacodynamics,
Pharmacology
Numerical issues
(simulation, optimized
computations, hardware
strategies)
Pharmacometrics
Statistics/Mathematics
(Theory, model formulation)
(Pato-)physiologi
(Disease states and
progression)
Main application - NONMEM
• NON-linear Mixed Effects Modeling
• Old (fashioned) FORTRAN 77 program
– First version appeared 1980
– Still the most widely used
• Single threaded…
Scope of problems to be solved
• Typical data consist of plasma concentrationdrug effect-time data from tens to thousands of
patients
• Run times for a single model fit varies
depending on model complexity and amount of
data
<1 min
Short
>10 min Moderate
> Days
Long
• A typical analysis involves 30-100 runs
Future trends
• More computer intensive methods, e.g.
– Stepwise variable selection (40-300 runs)
– Cross-validation (100-500 runs)
– Bootstrapping (200-2000 runs)
– Monte Carlo simulations
• Combinations of the above
• More mechanistically based models, i.e. more
complex models (>Days)
• Wider use of pharmacometrics in commercial
and academic research.
• Parallelized “NONMEM”
Impact of parallelization/distributed
computing
• Present software (NONMEM) does not allow
for parallel execution of single runs.
• Most computer intensive methods do lend
themselves to parallelization!
• Distributed computing solutions will allow us to
investigate the properties of methods that will
be tractable to the regular users 5+ years from
now.
Our current environment
• 20+ users
– The majority in applied work
– ~5 in theoretical research
• Cluster (since 2001) of 15+ CPUs
– Load balancing
– Parallel execution of multiple runs
Current hardware
200 GB
File
Server
Desktops/
Laptops
Fast Ethernet
5 double CPU
Computational
Servers
5
Workstations
Software
•
•
•
•
Red Hat Linux 7.3
openMOSIX kluster patch for kernel 2.4.19
Perl 5.6.1
Perl-speaks-NONMEM+Parallel::Forkmanager
– In house support library for parallelization of
multiple NONMEM runs.
– Recently completed.
Technical requirements
• Computational resources needed:
– Fall 2003 and onwards, as much as possible…
• Inter-processor relative processing speed:
– Inter-processor>processor
• Primary memory:
– For some applications 1Gb is too little
• Secondary storage:
– Up to 200 Gb for months
• Data access frequency:
– ?
• Status of software to be used:
– Depends on porting issues (help needed?)
– Late beta stage (PsN)