Mixture modelling (responders vs non-responders)

Hi NMUsers, I am trying to model some data that contain a biomarker time course over 4 weeks (with twice weekly measurements) and steady state of the change in biomarker is attained in this duration in the presence of various doses of a drug. This is a dose ranging experiment and I have value of exposure (AUC(0-tau)) at steady state that drives the biomarker change. I noticed that there are some obvious non-responders to the treatment (no mechanistic/physiological explanation). When using a direct effect Emax model the model fits are good and parameter estimates are well estimated. But the estimate of omega for EAUC50 is extremely high (about 13) due to the very high values estimated for the non-responders, thus simulations are extremely biased. I tried a mixture model on EAUC50 and NONMEM estimated 15% non-responders with high EAUC50 values (while examining the dataset non-responders are about 25%). With the mixture model implementation on EACU50 the omega on EAUC50 dropped to about 4 which is again extremely high for the responder population. I tried mixture model on Emax alone and EAUC50+Emax together also but again got high omegas and lower than expected estimate of non-responder population. I would appreciate if I can get some ideas to take care of this issue. THANKS!! Z Chen