From: Paul Hutson prhutson@pharmacy.wisc.edu
Subject: [NMusers] Filing
Date: 10/21/2003 9:55 PM
All:
Prompted by Nick's suggestion of thousands of fits of iterated data,
with comparison of the objective functions and description of the
parameters, I am very curious to learn from you PowerUsers out there
how you:
1) Manage and keep track of file names (and any variations in initial
estimates associated with each run),
2) Extract the parameter estimates (Thetas, Etas, Etcs) for descriptive
and statistical analysis.
For example, do you have Fortran strip out the parameters of interest
with a WRITE command to a file that is appended with subsequent runs,
or do you create a $TABLE text file, then parse it into a spreadsheet
like EXCEL with one or two macros that strip out the parameters to
another spreadsheet page?. Either way, how do you keep track, in a
given line of parameters gleaned from a given run, of what run and
what initial estimates it was derived?
Thanks in advance for your insight.
Paul
Filing
From: bvatul bvatul@verizon.net
Subject: Re: [NMusers] Filing
Date: 10/22/2003 1:59 AM
Hello
Wings for NONMEM developed by Nick is a very useful tool and one can extract
the parameter estimates etc by using simple command files.
Venkatesh Atul Bhattaram
CDER, FDA.
From: mark.e.sale@gsk.com
Subject: Re: [NMusers] Filing
Date: 10/22/2003 8:54 AM
Paul,
Re: your last comment, below is Fortran code for this. Put a call to
this (CALL WRITERES) right before the STOP in cfiles.for and recompile
including this source code file. By putting just before STOP (after all
other files are closed) you can be sure not to change anything else. (please
forgive my non-standard use of actual comments in fortran code). We have not
tested this with any compiler other than G77.
SUBROUTINE WRITERES
C BY MARK SALE, GLAXOSMITHKLINE, LAST REVISED 8/13/2003
INCLUDE 'NSIZES'
COMMON /ROCM9/ IERE,IERC
C IERE AND RC AS INTERNAL ERROR CODES FOR ESTIMATION AND COVARIANCE
INTEGER IERE,IERC
C MIN IS 0,1 MINIMIZED SUCCESSFULLY
INTEGER MIN
C CORR IS 0,1 PASSED CORRELATION TEST, COV IS 0,1 PASSED COVARIANCE TEST
INTEGER I,N,CORR,P,COV
C ABSOLUTE VALUE OF OFFDIAGONAL CORRELATION MATRIX
REAL ABSVAL
COMMON /CM2/ NETA,NEPS,NRE2
COMMON /CM1/ NTHETA,NTH,DB(LTH),LB(LTH),UB(LTH),ITR(LTH)
C OBJECTIVE FUNCTION VALUE
COMMON /ROCM8/ OBJECT
DOUBLE PRECISION OBJECT
C CORRELATION MATRIX IS CORRM
COMMON /CM12/ COVM(LPAR3),COVINM(LPAR3),STHTA(LTH),SEN(LVR,LVR),
1 CORRM(LPAR3),NVLS,VLS(LPAR)
COMMON /CM4/ MAXFN,NSIG,OPRPT,OPINT,OPFLP,OPETA1,IDUM(8),OPTWO
MIN = 1
C ERROR CODES 0 = OK, 139 = BOUNDARY, OTHER IS ERROR
IF(IERE.EQ.0.OR.IERE.EQ.139) MIN = 0
C IF IT FAILS MINIMZATION, IT FAILS COVARIANCE
IF(MIN.EQ.0.AND.IERC.EQ.0) THEN
COV=0
ELSE
COV = 1
END IF
C open OUTPUT file
OPEN(UNIT= 44,FILE = 'GSKNOUT')
WRITE(44,8) OBJECT
8 FORMAT ('OBJECTIVEFUNCTIONVALUE=',F16.2)
WRITE(44,'(A13,I3)'), 'MINIMIZATION=', MIN
WRITE(44,'(A11,I3)'), 'COVARIANCE=', COV
C AND THE CORRELATION, FIRST HOW MANY EFFECTS (DIMENSION OF MATRIX)
NEFF = NTHETA + (NETA*(NETA+1))/2 + (NEPS*(NEPS+1))/2
CORR = 0
P = 0
DO 66 I = 1,NEFF
DO 65 N = 1,I
P = P + 1
C CORRELATION MATRIX IS IN CORRM, LOWER TRIANGULAR
ABSVAL = ABS(CORRM(P))
IF((ABSVAL.GT.0.95).AND.(I.NE.N).AND.(ABSVAL.LE.1.00)) CORR = 1
65 CONTINUE
66 CONTINUE
C Correlation?
IF(MIN.EQ.0.AND.COV.EQ.0.AND.CORR.EQ.0) THEN
WRITE(44,'(A15)') 'CORRELATION= 0'
ELSE
WRITE(44,'(A15)') 'CORRELATION= 1'
END IF
CLOSE(44)
END
Mark
From: mark.e.sale@gsk.com
Subject: Re: [NMusers] Filing
Date: 10/22/2003 9:43 AM
One other issue of thetas, these are documented in ROCM6 as below
COMMON /ROCM6/ THETAF(40),OMEGAF(30,30),SIGMAF(30,30)
DOUBLE PRECISION THETAF,OMEGAF,SIGMAF
and can be written to a file like this:
WRITE(44,'(A5)') 'THETA'
DO 27 I = 1,NTHETA
WRITE(44,45) THETAF(I)
27 CONTINUE
45 FORMAT (E16.7)
Similarly if you want OMEGA
DO 50 I = 1,NETA
DO 49 N = 1,NETA
WRITE(44,60) OMEGAF(I,N)
49 CONTINUE
WRITE(44,'(A2)') ' '
50 CONTINUE
60 FORMAT (30E16.7,$)
Again, this code is for the G77 compiler, some of the formats are
different for digital/compaq.
As for keeping track of them, obviously this needs to be done by
software, not by hand. And we would of course recommend grid
computing for this.
Mark
From: "Bachman, William"
Subject: RE: [NMusers] grid computing
Date: 10/22/2003 11:04 AM
Hi, Mark,
We, on the other hand, would recommend cluster computing!
(yet another shameless plug).
On a serious note, now that we've figured out clusters, I'm
looking at interfacing PDx-Pop to grids. I've done some reading
up on the Sun Grid Engine. The nice feature of SGE is its
heterogeneity (Windows/Linux/UNIX) and of course the improved use
of existing corporate CPU power. [Internal versions of PDx-Pop have
the same heterogeneous abilility to access either Windows or Linux servers].
Can you share with us some details on the grid you've implemented at GSK?
Bill
ps I tried your code (with Compaq 6.6B and it worked fine, although
the output is a bit sparse and is hard coded with "GSK". I guess
I'm not the only "shameless plugger" out there ;)
The other sources of output summaries available to users are:
1. Alison Boekman's "nmsee"
2. the .sum file produced by PDx-Pop:
PDx-Pop 1.1j Release 4 Run Summary File Run No: 101
Date: Sep 26 2003 10:43
Output Extracted from file: c:\pdxpop1.1j\example1\101.res
DataFile: 002.CSV
MINIMIZATION STATUS
MINIMIZATION SUCCESSFUL
NO. OF FUNCTION EVALUATIONS USED: 88
NO. OF SIG. DIGITS IN FINAL EST.: 3.2
ETABAR IS THE ARITHMETIC MEAN OF THE ETA-ESTIMATES,
AND THE P-VALUE IS GIVEN FOR THE NULL HYPOTHESIS THAT THE TRUE MEAN IS 0.
ETABAR: 0.20E-01 -0.17E-01
P VAL.: 0.64E+00 0.76E+00
MINIMUM VALUE OF OBJECTIVE FUNCTION : 725.602
MODEL DEFINITION
TVCL=THETA(1)
CL=TVCL*EXP(ETA(1))
TVV=THETA(2)
V=TVV*EXP(ETA(2))
K=CL/V
S1=V
TAD=TIME ; for a single dose only
SID=ID
DEL=0
IF(F.EQ.0) DEL=1
W1=1
W2=F
IPRED=F
IRES=DV-IPRED
IWRES=IRES/(W1+W2)
Y=F + W1*EPS(1); + W2*ERR(2)
ERR1=EPS(1)
95% CONFIDENCE INTERVAL DESCRIPTOR/
FINAL ESTIMATE %RSE LBOUND UBOUND VARIABILITY
________________________________________________________________________________
THETA
1 0.00589 7.81% 0.00499 0.00679 CL, L/hr
2 1.44 5.78% 1.28 1.60 V, L
________________________________________________________________________________
INTERINDIVIDUAL
OMEGA VARIABILITY
1,1 0.198 43.1% 0.0306 0.365 CV = 44.5%
2,2 0.201 24.8% 0.103 0.299 CV = 44.8%
________________________________________________________________________________
RESIDUAL
SIGMA VARIABILITY
1,1 7.81 16.6% 5.26 10.4 SD = 2.79
From: mark.e.sale@gsk.com
Subject: RE: [NMusers] grid computing
Date: 10/22/2003 12:05 PM
Bill,
We use United Devices Metaprocessor. It is a web services based system.
Basically we package up everything needed (nonmem .obj and lib .files,
compiler files, control file, data set and a small executable) into a
zip file (about 3.5 Mbytes) and send it off to a computer where it is
unpacked, compiled, run and the required files zipped up and sent back.
All the files on the "agent" computer are then deleted. Nothing is
required on the agent computer except operating system and the metaprocessor
software. It runs at low priority, so the owner of the computer doesn't
notice at all. All the computers currently on the grid are at a single
physical site (this is not a metaprocessor requirement, but a NONMEM
license requirement), and we use the g77 compiler because Compaq wanted
$500,000 for enough licenses for their Fortran.
As for the file name GSKNOUT, feel free to change it to GLOBOMAXNOUT,
note there is no clause about not modifying code without first consulting with us ;).
Mark
From: Terri J. Fisher tjfisher@clinapps.com
Subject: RE: [NMusers] Filing
Date: 10/22/2003 12:40 PM
Hi Paul,
CLINapps Inc. has developed a clustered computing solution called SmartPK
that addresses all of your concerns below. It allows you to harness the
power of multiple CPUs, handles bootstrapping and batch loading, tracks
file names and parameter estimates, and much more.
We will be showcasing SmartPK at next week's AAPS meeting in Salt Lake City.
For a sneak peak at the SmartPK datasheet, please see our Online Expo entry:
http://www.aapspharmaceutica.com/onlineExpo/Exhibitor_Detail.asp?ID=156030
Terri Fisher
CLINapps Inc.
910 Hampshire Road, Suite G
Westlake Village, CA 91361 USA
Phone: (805) 494-1777, Ext. 15
Fax: (805) 494-1702
tjfisher@clinapps.com
www.clinapps.com
From: "Bonate, Peter" pbonate@ilexonc.com
Subject: [NMusers] grid computing
Date: 10/22/2003 1:01 PM
Okay, so for those of us who still work in the dark ages, the question
I have is, are things really faster? I'm not talking about faster because
the processor of the computer is faster, but faster because NONMEM can take
advantage of the parallel processing or grids or parallel algorithms or
whatever you want to call it.
Thanks,
pete bonate
From: Jill Fiedler-Kelly Jill.Fiedler-Kelly@cognigencorp.com
Subject: [NMusers] grid computing
Date: 10/22/2003 1:23 PM
Paul, Pete, and others,
Grid computing is an ideal solution for running NONMEM on lots of
systems. We at Cognigen have been using Grid computing for NONMEM for
almost 2 years now. As Bill mentioned in his response, we also
implemented grid computing to optimize use of our existing computer
infrastructure. When you run NONMEM on your desktop PC, regardless of
how fast it is, not only are you limited by the number of jobs you can
run, but you are also tying up your computer while it is processing
those jobs. When we utilize either of our internal grids, scientists
can submit hundreds of jobs simultaneously (bootstrapping) and still
utilize their workstation for other tasks.
Cognigen is currently utilizing three types of grids for NONMEM. One is
SUN-based using SUN's SPARC processors; the second is Intel-based
running Linux; and the third is for clients which has a web front end
for remote access via the Internet. We initially used SUN Grid Engine
for the grid platform, but have since moved over to GridEngine, which is
the opensource product that Sun Grid Engine is based on. Also, Sun Grid
Engine/GridEngine only runs on a Unix host. It does not run on Windows
PCs, but the opensource product does allow for job submission from a
Windows PC.
Jill
______________________________________
Jill Fiedler-Kelly
Cognigen Corporation
395 Youngs Road
Buffalo, NY 14221
(v) 716.633.3463, ext. 228
(f) 716.633.7404
(e) Jill.Fiedler-Kelly@cognigencorp.com
http://www.cognigencorp.com/
From: Jill Fiedler-Kelly Jill.Fiedler-Kelly@cognigencorp.com
Subject: [NMusers] grid computing
Date: 10/22/2003 1:26 PM
Paul,
Since the flood gates have now been opened ...
In response to your questions about file names, parameter estimates,
etc. you may be interested in looking at Cognigen's PERSPECTIVE
Hypertext Data Analysis Mapping software. The primary function of this
software is the organization of analysis results, which includes
documenting these things you mentioned. There is a NONMEM component in
this application that generates summary tables of results, calculates
model-specific sumary statistics and allows you to assign labels to
parameter estimates. The output from PERSPECTIVE is published to HTML
and the entire output is also convertable to PDF for storage or
submission. If you'd like to take a look at this, go to
http://www.cognigencorp.com/perspective.
Jill
______________________________________
Jill Fiedler-Kelly
Cognigen Corporation
395 Youngs Road
Buffalo, NY 14221
(v) 716.633.3463, ext. 228
(f) 716.633.7404
(e) Jill.Fiedler-Kelly@cognigencorp.com
http://www.cognigencorp.com/
From: "Bachman, William" bachmanw@globomax.com
Subject: RE: [NMusers] grid computing
Date: 10/22/2003 1:31 PM
Pete,
In the case of multiple runs sent to a multiple-processor cluster, the
speed-up factor is roughly inversely related to the number of processors
in the cluster.
From my AAPS poster presentation in Salt Lake City:
Mathematical expressions for total run-time as a function of the
number of runs in a batch and the number of processors in the cluster, where:
P = no. of processors in the cluster (equal speed CPU's assumed)
R = no. of run in the batch (equal individual run-times assumed)
T(P,R) = total run-time as a function of P and R
T1 = run-time for one run on one processor
TO = network and software overhead time
N = 1, 2, 3, ... (integer values)
M = the number of integral multiples of runs compared to the number of processors
= (R - modulus (R/P))/P
For R>P and R is an integral N multiple of P:
T(P,R) = N (T1 - T1/P) + TO
For R< P:
T(P,R) = (T1 - T1/R) + TO
For R>P and R is not an integral multiple of P:
T(P,R) = M (T1 - T1/P) + (T1 - T1/R) + TO
As T1 increases, TO approaches 0 and is negligible for all but very fast runs.
William J. Bachman, Ph.D.
Manager, Pharmacometrics Research and Development
GloboMax
The Strategic Pharmaceutical Development Division of ICON plc
7250 Parkway Drive, Suite 430
Hanover, MD 21076
410-782-2212
bachmanw@globomax.com
From: Nick Holford n.holford@auckland.ac.nz
Subject: Re: [NMusers] Filing
Date: 10/22/2003 2:28 PM
Mark,
You wrote:
>>(please forgive my non-standard use of actual comments in fortran code).
The de facto NONMEM coding standard does not preclude inclusion of 'actual'
comments. For example,
C DO NOT MODIFY CODE
C WITHOUT FIRST CONSULTING WITH THE NONMEM PROJECT GROUP
It is however part of the NONMEM standard not to include 'informative'
comments. You should be seeking forgiveness for including comments
that explain what the code is doing
Nick
>>
>> SUBROUTINE WRITERES
>> C BY MARK SALE, GLAXOSMITHKLINE, LAST REVISED 8/13/2003
>> INCLUDE 'NSIZES'
>> COMMON /ROCM9/ IERE,IERC
>> C IERE AND RC AS INTERNAL ERROR CODES FOR ESTIMATION AND COVARIANCE
>> INTEGER IERE,IERC
>> C MIN IS 0,1 MINIMIZED SUCCESSFULLY
>> INTEGER MIN
>> C CORR IS 0,1 PASSED CORRELATION TEST, COV IS 0,1 PASSED COVARIANCE TEST
>> INTEGER I,N,CORR,P,COV
etc...
--
Nick Holford, Dept Pharmacology & Clinical Pharmacology
University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand
email:n.holford@auckland.ac.nz tel:+64(9)373-7599x86730 fax:373-7556
http://www.health.auckland.ac.nz/pharmacology/staff/nholford/
_______________________________________________________