NONMEM Tips-n-Tricks

log-transform both sides: cognigen.com/nonmem/nm/99apr232002.html

Model Evaluation

  • In Pfizers tutorial (Byon et al. 2013), there is an argument as to not associate OFVs with p-values, and to regard a drop in OFV with ~10
  • Driving individuals? (.phi file)
    • Sharkplot (ΔOFV = OFV_full - OFV_reduced vs #subjects_removed): compare between runs. Should be able to remove 5 ID without loosing significance.
  • |CWRES|<5 (estimate change ->remove, doesn’t change -> keep)
  • ETA ShrinkageSD < 20% ? (then we can use EBE based diagnostics)
  • EPS ShrinkageSD < 20% ? (then we can look at IPRED vs DV)
  • boxcox ETA: dOFV > -10.827 (p<0.001)
  • VPC
  • QA
  • RSE% considered precise: <30% for theta, <50% for eta
  • Use $COV MATRIX=R for consistency between statistical software.
  • Condition number (CN) – $COV PRINT=E
    • Calculated differently in different software
      • In PsN it is calculated as dividing the largest eigenvalue with the smallest
      • Gabrielsson and Wiener calculates it as log(largest/smallest)
    • Different guidelines for ill-conditioning
      • if CN < 10p, where p=number of estimable parameters, then it’s good
      • There are also references that point to CN<106
      • <1000; but this seems to relate more to linear models, or PK-models with 3 parameters (CL, V, ka)
  • SIR
  • SADDLE_RESET
  • MCETA
  • NRD = Number of Required Digits

Notes

  • IMPMAP is equivalent to IMP MAPITER=1 MAPINTER=1
  • NONMEM places the low bound equal to 1/100 (or 1/1000) of the initial value even if it is set to zero in the code. You may rerun with lower initial value or put NOTHETABOUNDTEST to the estimation record. (Credit: Leonid Gibiansky)
  • Use ”Disease modifying” instead of ”protective” (Credit: Mats Karlsson)
  • Mixture prior slide 31. Prior+flat prior ”what if we are wrong”. When conflict: up-weights data, downplays history
  • .ets file gets “random sampled ETAs” (less prone to shrinkage)
  • Tab-separated (less likely to be present in the actual data)
  • EXIT(1) instead of p=0 /+1E-15 (Credit: Gunnar Yngman)
  • Two-stage approach: Apply a model to each individuals data
    • Require much data per individual.
    • OK mean parameters (i.e. typical individual)
    • But IIV id inflated (also diff ID -> diff model)
  • NONMEM: Apply a model to all individuals data
  • .phi: Individual parameters: φi=μi+ηi, φi=(φ1,…,φn)
    • phc=Var(phi)
    • non_mu_ref parameters: φi=ηi
  • PK Studies:
    • “Intensive studies” (Phase I/II)
      • Few subjects (homogeneous)
      • Single dose
      • Frequent (rich) sampling
    • “Population studies” (Phase III/IV)
      • Many subjects (heterogeneous)
      • Multiple dose
      • Sparse sampling
  • Time-to-event
    • Non-parametric: Log-rank test
    • semi-parametric: Cox
    • parametric: Weibull, Gompertz, etc
    • Weibull*COV forces proportional hazard (Credit: Andrew Hooker)
  • Time varying covariate: Specify as “regressor”.
  • When you use F, CMT matters (Credit: Maria Kjellsson)
  • TTE simulations can be done either with a dataset containing all possible event times, or using MTIME (Credit: Joakim Nyberg)
  • Faster parallel NONMEM:
    • execute -nmfe_options=“[sbatchargs=–threads-per-core=1] -maxlim=2 -parafprint=100”
    • FILE* should be size 0
    • FILE07–39 should be size 0 (Credit: Leonid Gibiansky)
  • ETADER=3: Get out of a local minima (Credit: Rikard Nordgren)
  • logistic=expit=inverse logit
    • logit≈probit(scaled)
  • $MIX NSPOP=3 (from lecture_simultaneous UPSS)
    • P(1) = THETA(1)
    • P(2) = (1-P(1))*THETA(2)
    • P(3) = 1-P(1)-P(2)
    • NSPOP = 3
    • $THETA1 (0, THETAI(1), 1)
    • $THETA2 (0, THETAI(2), 1)
  • Start every dataset with a ”Comment” column, the IGNORE=@ will then ignore a row if there is a text comment!
    • The second column should be a REF column with each line number

[NMusers] Interactive Control of NONMEM runs

Thank you both for your help, I feared that Perl was capturing the keyboard commands, and the sig.exe would sadly not help in my case, because I wanted to use the function in a Linux environment.

But after playing with sig.exe a bit, I realized that all it does is creating a file that gives the signal. So by creating the file manually, you can override the whole process.

All you need to do is finding the directory where NONMEM is running (NM_run1 in PSN) and manually create an empty file with a different name according to the desired action. Here’s a list:

Print toggle (monitor estimation progress): print.sig

Paraprint toggle (monitor parallel processing traffic): paraprint.sig

Next (move on to next estimation mode or next estimation): next.sig

Stop (end the present run cleanly): stop.sig

Subject print toggle: subject.sig

The function I wanted is “Next”. NONMEM grinds a couple if iterations more and then it terminates nicely as if the maximum number of function evaluations had been reached.

I hope this info can help others too.

Thank you, Paolo

Random ETA samples:

When used in nonmem 7.5, a recommended example/setup would be (as shown on = page 152 of nm750.pdf)

$SIZES ISAMPLEMAX=250

... 

$ESTIMATION METHOD=SAEM NBURN=500 NITER=1000 MASSRESET=1 ISAMPLE=2 PRINT=20 NOPRIOR=1 RANMETHOD=P

$ESTIMATION METHOD=SAEM NBURN=0 NITER=0 MASSRESET=0 ETASAMPLES=1 ISAMPLE=200 EONLY=1 

$ESTIMATION METHOD=IMP NITER=5 MASSRESET=1 ETASAMPLES=0 ISAMPLE=1000 EONLY=1 PRINT=1 MAPITER=0

(Credit: Robert Bauer)

IMP

I am building a relatively complex PKPD model (with 47 parameters and 11 differential equations).

I had problems using FOCE so I am trying this estimation method :

$ESTIMATION METHOD=ITS INTERACTION LAPLACE NITER=200 SIG=3 PRINT=1 SIGL=6 NOHABORT 
CTYPE=3 NUMERICAL SLOW 

$ESTIMATION METHOD=IMPMAP ISAMPLE=1000 INTERACTION LAPLACE NITER=1000 SIG=3 PRINT=1 
SIGL=6 NOHABORT CTYPE=3 IACCEPT=0.4 MAPITER=0 RANMETHOD=3S2 

$COVARIANCE UNCONDITIONAL MATRIX=S TOL=12 SIGL=12 SLOW

The iteration for the ITS step seems to be quite stable with some artefacts:

iteration 175 OBJ= 4693.4674554341409
iteration 176 OBJ= 4694.2296104065535
iteration 177 OBJ= 4693.7753507970829
iteration 178 OBJ= 4693.9600270372885
iteration 179 OBJ= 4693.5732455834705
iteration 180 OBJ= 4693.6386423202493
iteration 181 OBJ= 4693.6215390721527
iteration 182 OBJ= 4693.6006496138452
iteration 183 OBJ= 4693.7877620448235
iteration 184 OBJ= 4694.1591757809929
iteration 185 OBJ= 4693.2614956897451
iteration 186 OBJ= 4693.5641640401127
iteration 187 OBJ= 4693.5575289919379
iteration 188 OBJ= 4495.6489907149398
iteration 189 OBJ= 4693.7711764252363
iteration 190 OBJ= 4693.6281175153035
iteration 191 OBJ= 4694.1171774559862
iteration 192 OBJ= 4693.7908707845536
iteration 193 OBJ= 4693.7709264605819
iteration 194 OBJ= 4495.9262902940209
iteration 195 OBJ= 4693.3321354894242
iteration 196 OBJ= 4694.3177205227348
iteration 197 OBJ= 4694.1301486616576
iteration 198 OBJ= 4694.2898587322170
iteration 199 OBJ= 4693.8304358341920
iteration 200 OBJ= 4691.6818293505230

#TERM: 
OPTIMIZATION WAS NOT COMPLETED

The IMP step seems less stable :

iteration 120 OBJ= 4314.8310660241377 eff.= 446. Smpl.= 1000. Fit.= 0.96389
iteration 121 OBJ= 4326.9079856676717 eff.= 448. Smpl.= 1000. Fit.= 0.96409
iteration 122 OBJ= 4164.6649529423103 eff.= 479. Smpl.= 1000. Fit.= 0.96392
iteration 123 OBJ= 4299.9887619753636 eff.= 432. Smpl.= 1000. Fit.= 0.96395
iteration 124 OBJ= 4303.9571213327054 eff.= 399. Smpl.= 1000. Fit.= 0.96349
iteration 125 OBJ= 4328.9835950930074 eff.= 417. Smpl.= 1000. Fit.= 0.96423
iteration 126 OBJ= 4304.3861595488252 eff.= 550. Smpl.= 1000. Fit.= 0.96392
iteration 127 OBJ= 4291.0862736663648 eff.= 422. Smpl.= 1000. Fit.= 0.96430
iteration 128 OBJ= 4326.2378678645500 eff.= 407. Smpl.= 1000. Fit.= 0.96409
iteration 129 OBJ= 4157.5352046539456 eff.= 406. Smpl.= 1000. Fit.= 0.96404
iteration 130 OBJ= 4332.6894073732456 eff.= 399. Smpl.= 1000. Fit.= 0.96399
iteration 131 OBJ= 4357.5343346793761 eff.= 493. Smpl.= 1000. Fit.= 0.96414

Convergence achieved 

iteration 131 OBJ= 4336.1893012015007 eff.= 417. Smpl.= 1000. Fit.= 0.96369 

#TERM: 
OPTIMIZATION WAS COMPLETED

The “unstability” of the IMP step is it usual? NONMEM is completed at the end..

In answer to your question; yes, it is usual to see this “unstability” in the final few iteration OFVs.

When using the IMP method, I often include two sequential $EST commands. The first command will perform optimisation of parameter estimates until a global minimum is found. The second command will then take those parameter estimates and calculate more precise estimates of the objective function value. The second $EST command will have a higher ISAMPLE to reduce the Monte Carlo noise, and have ETYPE=1 (no optimisation of parameter values).

I suspect that the number of samples that you are using may not be enough, giving large Monte Carlo noise in the OFV estimate. I suggest that you perform another run with the parameter values set to their final estimates, and with:

$ESTIMATION METHOD=IMP ISAMPLE=10000 INTERACTION LAPLACE NITER=5 SIG=3 PRINT=1 
SIGL=6 EONLY=1 NOHABORT RANMETHOD=3S2

The higher number of samples should give a more stable result (although the run time of each iteration will increase significantly). Taking the average OFV of these 5 iterations will give a more accurate estimation of the final OFV. Jon Moss, PhD

SAEM

http://monolix.lixoft.com/tasks/population-parameter-estimation-using-saem/

SIGNALS for eg SAEM/IMP

https://www.mail-archive.com/nmusers@globomaxnm.com/msg03990.html Find the directory where NONMEM is running (NM_run1 in PsN), and manually create an empty file with a different name according to the desired action. Here’s a list: * Print toggle (monitor estimation progress): print.sig * Paraprint toggle (monitor parallel processing traffic): paraprint.sig * Next (move on to next estimation mode or next estimation): next.sig * NONMEM grinds a couple if iterations more and then it terminates nicely as if the maximum number of function evaluations had been reached. * Stop (end the present run cleanly): stop.sig * Subject print toggle: subject.sig

VPCs

The VPCs central metric are the prediction of data percentiles. If you focus on the difference between e.g. the 5th and 95th percentile based on the simulated data you will have a prediction interval, like Bill states. If you focus on an individual percentile, but consider the imprecision with which it is derived, often given as a shaded area, then it is like other metrics of imprecision a confidence interval. Confidence interval (ci) and prediction interval (pi) for VPC

In general, if the serum creatinine rises at 2-3 mg/dl per day then the GFR is near zero.

University of Maryland

https://ctm.umaryland.edu/#/ms-pharma