The optimal antipsychotic drug dose in patients is usually based on PET imaging with an optimal window between 70% and 80% of D2
receptor occupancy 
, with an exception for the partial D2
receptor agonist aripiprazole. However the D2
receptor occupancy only partly accounts for the clinical response and our increased understanding of other neurotransmitter systems and systems interactions has not been effectively integrated into antipsychotic drug discovery. We demonstrate here a quantitative mechanistic computer-based model as a translational tool that combines preclinical physiology data with patient-centered data on neuronal circuits, pathology and pharmacology, eliminating some of the inherent limitations of preclinical animal models 
. Due to its mechanistic nature this model is limited to specific disease areas like schizophrenia, in contrast to the more generic systems biology data-mining approaches often applied to different disease areas.
We showed that retrospective evaluation of drug efficacy with a wide range of pharmacological activities using this computer model is more effective than simple receptor D2 receptor competition or multivariate regression analysis. We have further tested this translational model by predicting, in a blinded manner, the clinical profile of two compounds for which clinical data had been collected but not published or available to the modelers at the time of evaluation. To our knowledge, this is the first time that any simulation model has been tested in such a blinded manner.
The results suggest that the mechanistic disease model correctly predicts the relative performance for JNJ37822681 in PANSS total score improvement and EPS liability with respect to olanzapine, but not for ocaperidone. The low-affinity property of JNJ37822681 differentially modulates only the dopaminergic striatal pathway effects during burst and tonic dopamine activity. The model prediction of the potent clinical improvement with clozapine and of the efficacy of olanzapine as compared to the two highest doses of JNJ37822681, despite the same or lower D2 receptor occupancy also suggests that the computer model adequately captures the beneficial contribution of additional non-D2 receptor actions.
In line with the reported clinical benefit of trazodone in Parkinson disease patients 
, our model suggests that cortical 5-HT2
A activity is a key modulator of EPS liability and that the fast dissociation rate of JNJ37822681 may only compensate partially for the EPS liability induced by significant D2
receptor inhibition during burst firing, This is not unlike remoxipride that has a substantial EPS liability despite a low affinity for the D2
. We believe this translational disconnect is likely due to species difference of the dopaminergic synapse physiology between primates and rodents 
While a simple multivariate regression analysis can already give a good idea on the expected outcomes, our analysis suggests that the mechanism-based computer model is superior in predicting the clinical outcomes of both JNJ37822681 and ocaperidone. This is likely due to the fact that the multivariate analysis assumes independent processes that affect the outcome in a linear way. In contrast actual physiological modeling can account for a non-linear processes such as the threshold for action potential generation or the complex non-linear interaction between different receptor systems (for instance one neurotransmitter regulating the release of another neurotransmitter) that modulate the membrane potential. While the multivariate regression analysis can identify a possible target that drives the clinical outcome (for instance the 5-HT2AR in the case of the unexpected high EPS liability of JNJ), the computer-based mechanistic mode can add the appeal of quantitatively understanding the neurobiology, i.e. clarifying the link from receptor modulation to membrane excitability through modulation of specific ion channels in specific parts of the neuronal network. In addition, in contrast to the mechanism-based computer model, multivariate regression analysis is unable to predict the outcome of a new target that hasn't been tested in the clinic before, or the effect of comedication often used in clinical trials.
The failure of the model to predict ocaperidone clinical outcome may be due to imperfect representation of the off-target physiology in the model. Alternatively, with steady-state dosing, the ocaperidone levels could accumulate leading to increased functional brain concentrations. Indeed hypothetical higher D2
receptor occupancy for ocaperidone substantially reduces the differences between predicted and reported values for EPS liability and PANSS total score. In that regard it is interest to note that ocaperidone is much more potent in vivo than haloperidol or risperidone with ED50
values in the amphetamine test below 1 microg/kg 
. Additionally, for missing data we assumed the drug did not affect those receptors and that affinities to rodent receptors are identical for human receptors, but species differences in affinity are commonly present.
There are several issues, however, for which the model falls short. First, the results represent a relative difference from baseline, rather than an absolute predictor of clinical outcome. However, this approach is the only ‘preclinical’ model that predicts actual PANSS total score or EPS liability outcomes, in contrast to animal models that give more of a binary prediction.
The model fell short on the absolute prediction of the placebo improvement. The increased placebo improvement which has been observed lately in clinical trials cannot be effectively modeled by this approach, because they are presumably associated with issues like expectancy bias on the part of the investigator and the patient 
. Additionally, the model has been calibrated using historical values for the placebo effect collected in 26 different papers since 1988, where the placebo effect was much less prominent.
It is also important to realize that the model predictions are limited by the current state of knowledge. For example, the computer-based model is much less effective in predicting akathisia-related side-effects compared to Parkinsonian type side-effects 
. Although historically it has been classified as an extrapyramidal disorder 
, akathisia might be driven by pathophysiological mechanisms more reflective of anxiety than motor signs 
. The current version of the computer-based EPS model is focused on the cortico-nigrostriatal-thalamic pathway pathology and does not take into account other extrastriatal pathways.
The current EPS computer model is limited to Parkinsonian physiology and is well calibrated with historical data from patients initiated on anticholinergic medication to treat EPS symptoms. This might lead to differences between potential and expressed pathological changes - i.e., the drug may have increased EPS liability, but its expression in humans may not correspond to a given clinical readout unless very large numbers of patients are used. For instance, high EPS liability may not be optimally assessed by the use of anticholinergics, which is an ‘all-or-none’ approach that depends both on the subjects' description of the event and the physician's ability to elicit, characterize and manage that symptom. However the results suggest that the model correctly captures the ranking of the investigative drugs as compared to olanzapine with regard to the EPS liability. The platform also has reasonable correlations with some other measures of EPS liability 
, such as Simpson-Angus Scale (SAS) and the Abnormal Involuntary Movement Scale (AIMS) which capture different clinical aspects of this complex side-effect profile.
It is of interest to compare the predictivity of this computer-based modeling approach with the more traditional animal models currently used in psychiatry Research & Development. Both JNJ37822681 
and ocaperidone 
passed all preclinical animal tests to the point that they were deemed of interest for a (financial and resource-intensive) investment in clinical development. Yet the computer model would have been able to raise a red flag about the EPS liability for JNJ37822681, because it quantitatively showed that the fast dissociating properties at the D2R did not compensate for the lack of effect at the cortical 5-HT2
AR. The computer model further predicted a lack of clinically relevant differentiation between ocaperidone and olanzapine and suggested that higher doses of ocaperidone would reduce the therapeutic ratio between effect on PANSS Total and EPS liability. In addition, we are not aware of preclinical animal models that can quantitatively predict a PANSS total score, especially in comparison to an existing drug on the market.
The computer-based model has been calibrated using average values of treatment groups and do not reflect inter-individual differences caused by differences in individual genotypes and co-medications; however, the model, in principle can accommodate genotypic profiles if such information is obtained from the patient population evaluated, for instance through imaging genetics 
We chose to focus on the PANSS Total as readout because this is usually the primary readout for clinical trials with antipsychotics and there are more historical data available. We have been testing our computer model against other subscales, such as PANSS positive and PANSS negative subscales, for which we have less published data. Not unexpectedly, the calibration with PANSS positive subscale is very similar to the PANSS Total scale and the relative effect of the drugs on the PANSS positive scale is similar to their effect on PANSS Total.
Future developments include the implementation of more detailed subcortical anatomy and physiology 
that will take into account the different properties of the direct versus indirect pathway in combination with detailed modeling of the globus pallidus interna and externa, the subthalamic nucleus and part of the thalamus. Alternatively other receptor types and neurotransmitter systems can be implemented in the appropriate brain region to build a model that is for instance, more suited for cognitive or negative symptoms Such an approach could, in principle, lead to other models for Parkinson's and Huntington's diseases.
Current preclinical animal models generally provide binary information relative to safety and efficacy, but they rarely predict relative performance of a novel investigative drug to a comparator. This computer-based mathematical model, calibrated retrospectively using published clinical data of many antipsychotic drugs can predict relative clinical outcomes, important in prioritizing discovery projects. In addition, when no target engagement data in humans are available, the computer-based model allows for the relative therapeutic window between PANSS effect and EPS liability to be estimated.
The ‘Quantitative Systems Pharmacology’ approach is being increasingly recognized as a possible translational tool for drug discovery and development in the field of oncology and metabolism 
and contributed to a number of newly approved, rationally designed cancer drugs. Although the current understanding of human neurobiology in general and in schizophrenia pathology in particular is currently limited, the combination of the existing large academic expertise in computational neuroscience and the availability of endophenotype studies of the human brain using PET imaging and electroencephalogram (EEG) provides the framework for an increasingly more powerful ‘Quantitative Systems Pharmacology’ approach. In this context, it is of interest to note, that although the current version of the computer model is largely based upon existing dopamine dominated antipsychotic pharmacology; new cholinergic and glutamatergic targets can be readily introduced into the model based upon their preclinical physiology. As they in turn affect more complex neuronal network systems, like the type we model here; we expect that this ‘Quantitative Systems Pharmacology’ approach can yield better insights than pure qualitative reasoning as is done now.
In summary, although the current model did not perfectly predict the clinical outcome for the novel antipsychotic drugs, the comparative results against the active comparator were more reliable than could have been estimated by D2 binding properties or by preclinical animal model outcome. Further refinements using our expanded knowledge about receptor profiles and systems interaction should permit an even better predictive capacity. This approach can provide valuable insight into relative clinical efficacy and EPS liability, as well as into novel drug targets beyond the dopamine system and more efficiently drive drug development by enabling better selection of drugs prior to expensive and time-consuming clinical testing.