Numerical instabilities in NM 7.5 ADVAN7, maybe related to gfortran 9.4.0?

Dear fellow NONMEM users, We recently observed that a standard PK model using ADVAN7 with FOCEI became numerically unstable after we switched from NM 7.3 to NM 7.5. In particular, we saw that the run showed a drastically increased runtime, did not converge anymore and ultimately aborted due to rounding errors. When setting the initial parameter estimates directly equal to the optimum identified previously in NM 7.3, the run would, instead of quickly converging, slightly move away from its starting point and then show the same problems. Adding $ABBRV PROTECT did not change this. Interestingly, when setting MAXEVAL=0, we observed significantly different initial OFVs between NM 7.3 and NM 7.5, differing by approximately 100 units. For comparison, we repeated the runs using ADVAN5 instead of ADVAN7. Then, both NONMEM versions quickly converged to the exact same optimum previously identified using ADVAN7 in NM 7.3, showing the exact same final OFV. The same success occurred when we rewrote the model using ADVAN6 instead, however NM 7.5 ran significantly longer than NM 7.3 (~45'000s vs. ~2'000s). Taken together, ADVAN7 seems to show different results in NM 7.5 (even with MAXEVAL=0 before the first iteration the OFV is different), and ADVAN6 is much slower compared to NM 7.3 We would greatly appreciate any insights or ideas for further exploration you may have on this matter. In a recent message[1], Robert Bauer mentioned that the gfortran version used to install NONMEM can have an influence on the performance of classical estimation methods. In particular, gfortran 14 seems to use faulty optimisation techniques which apparently can cause rounding errors in NM 7.5.1. We were, thus, wondering if anybody knows whether gfortran 9.4.0, which we use, could cause similar problems? Thank you very much for your help and best regards, Matthias M. Fischer [1] https://www.mail-archive.com/[email protected]/msg08937.html Freundliche Grüße / Best regards Dr. Matthias M. Fischer Pharmacometrics Expert ////////////////////// Bayer Aktiengesellschaft Pharmaceuticals Quantitative Pharmacology M004, 05 / 532 Müllerstraße 178 13353 Berlin, Germany E-mail: [email protected]<mailto:[email protected]> Web: https://www.bayer.com/ https://www.bayer.com/de/corporate-compliance/datenschutzhinweise-fuer-ausgewaehlte-verarbeitungstaetigkeiten // Data privacy https://www.bayer.com/en/corporate-compliance/data-privacy-information-for-specific-processing-activities /// Board of Management: Bill Anderson, Chairman | Wolfgang Nickl, Stefan Oelrich, Heike Prinz, Rodrigo Santos, Heiko Schipper /// Chairman of the Supervisory Board: Norbert Winkeljohann /// Registered office: Leverkusen | Amtsgericht Köln, HRB 48248 /// Follow Bayer on social https://go.bayer.com/social-media-channels

Please read the nm751_bug_list.pdf document that is located in: https://nonmem.iconplc.com/nonmem751 In particular items 1 and 6 describe problems pertaining to ADVAN5 and ADVAN7. The work-around is to set the option -prdefault at the command line, or insert $SIZES PC=$SIZES PC=xxx where xxx>=2*NCOMPARTMENTS+1 In addition, this may resolve the speed difference you observed with ADVAN6. It is unlikely that gfortran 9.4 has the optimization switch problem (the -O3 switch in particular) encountered in gfortran 14. Robert J. Bauer, Ph.D. Senior Director Pharmacometrics R&D ICON Early Phase 731 Arbor way, suite 100 Blue Bell, PA 19422 Office: (215) 616-6428 Mobile: (925) 286-0769 [email protected]<mailto:[email protected]> http://www.iconplc.com/