Exploring Glutamatergic and Cholinergic Dose-Response in Schizophrenia Using a Novel Quantitative Systems Pharmacology Approach Hugo Geerts 1,2 , Athan Spiros 1 , Patrick Roberts 1,3 1 In Silico Biosciences, Berwyn, PA; 2 University of Pennsylvania, Philadelphia, PA, 3 Oregon Health & Sciences University, Portland, OR, USA Background  Clinical effect in CNS therapeutics unlike target engagement does not always follow a monotonic dose-response.  Inverse U-shape clinical dose-responses are often caused by off-target effects at higher doses.  In some cases, inverse U-shape dose-response is a consequence of the underlying neurobiology at the target or at the circuit level.  Alpha-7 nicotinic AChR modulation is an attractive target for cognitive impairment in schizophrenia and in Alzheimer’s disease. (1)  mGluR2 partial agonist is currently being tested as a stand-alone medication in schizophrenia. (2)  Computer-based mechanistic disease modeling is a quantitative systems pharmacology approach (http://isp.hms.harvard.edu/wordpress/wp- content/uploads/2011/10/NIH-Systems-Pharma-Whitepaper-Sorger- et-al-2011.pdf) that uses mathematical algorithms to simulate the firing dynamics of neuronal cells in biophysically realistic networks under pathological conditions. (Fig 1) Methods  The mathematical simulations combine preclinical neurophysiology of relevant brain circuits and targets with human imaging studies, pathology and neuropharmacology and are therefore a real translational tool.  The PANSS model simulates the neurophysiology of the N accumbens Medium Spiny Neuron (MSN), driven by cortical input, gated by hippocampal and amygdala input and modulated by dopaminergic projections from the VTA, 5-HT input from the Dorsal Raphe, cholinergic input from tonically active interneurons and NE input from the Locus Coeruleus. (Fig 2A) Computational Neuropharmacology Platform Calibration Platform Validation Platform Application Receptor Physiology CNS Drug Pharmacology Clinical Trials Independent Data Sets Pharmacology Drug (PET Target Engagement) Human Pathology Imaging Studies Predictive Validation PANSS, EPS Cognitive Effect in AD, SZ C m g n (t, V) E n g L E L I P Intracellular Medium Extracellular Medium Preclinical Data Academia Figure 1: General overview of the development of computer based mechanistic-disease platform. Starting from the Hodgkin-Huxley approach we add receptor physiology, data on CNS active drugs and human pathology data from imaging and postmortem studies, calibrate the model with retrospective data on clinical outcomes; test the model with independent data sets and blinded predictions of clinical outcome. Finally, the model uses the drug pharmacology against human receptors to determine an estimated clinical outcome. Mechanistic Brain Modeling AMPA NMDA AMPA AMPA/ NMDA AMPA NMDA NMDA D1 D1 D1 D1 D1 D1 D1 GABA GABA GABA D4 A4b2 nAChR D4 D4 A4b2 nAChR A4b2 nAChR a7 nAChR a7 nAChR a7 nAChR a7 nAChR a7 nAChR a7 nAChR a7 nAChR distal proximal soma basal Figure 2B: Diagram for the physiological processes implemented in the 120 neuron cortical network. All interactions are derived from preclinical data whereas the basal values for ion channel conductances are calibrated using primate electrophysiology data. The model simulates the electrical activity of the network that can be measured by EEG or BOLDfMRI readouts.  Further calibration is done by optimizing the correlation between retrospective clinical trial results (71 drug-dose combinations with 24 different antipsychotics) and the outcome of the same drug-dose combinations in the computer model.  The model captures threefold more variance (0.69) than the simple D2R occupancy measure (0.18) (Spiros, submitted). (Fig 3A) Correlation Between Cortical Model Output and Clinical Dataset 4.00 5.00 6.00 7.00 8.00 9.00 10.00 40.00 50.00 60.00 70.00 80.00 90.00 N-Back Outcome Working Memory y = 0.0786x + 3.001 R 2 = 0.765 Correlation with working memory Figure 3B: Correlation between working memory span outcome for the computer-based cortical network model and clinical results on working memory performance in 17 clinical situations: schizophrenics with 5 antipsychotics, COMT genotype; normals with scopolamine, mecamylamine; Tolcapone and COMT genotype.  Schizophrenia pathology is introduced in cortical network by lower glutamate tone (3), GABA dysfunction at basal dendrites (4), dopamine hypo-activity (5) and increased noise (6).  The mGluR2 affects synaptic glutamate levels as a presynaptic autoreceptor (7) and modulates cortical afferent drive in the N. accumbens. (Fig 4A) MSN Tonically Active Cholinergic neuron Amygdala Cortical Glu afferents VTA neurons Dorsal Raphe 5HT neurons Pre-synaptic mGluR2/R3 D1 D2 mGluR1 M1 GABA interneuron (j+1) k - drug (5-(i+j)) k + drug [drug] R i(j+1) = R ij R (i+1)j = R ij (i+1) k - ACh (5-(i+j)) k + ACh [ACh] R 00 R 01 R 10 R 02 R 11 R 03 R 21 R 30 D 00 D 01 D 10 D 02 D 11 D 20 D 03 D 12 D 21 D 30 R 20 O 11 O 20 O 01 O 10 O 12 O 02 O 03 R 12 O 21 O 30 mGluR2 Striatal Wiring Diagram a7 nAChR Binding Markov Chain Model Figure 4A: mGluR2 wiring diagram showing its impact as a presynaptic glutamate autoreceptor at different positions in the network. Although the cortical contribution is the most important, other effects also play a role. Figure 4B: Computer model for the alpha-7 nicotinic AChR with 29 states reflecting the effect of competition between the endogenous agonist and the partial agonist and calibrated using in vitro experiments. This Markov chain model allows for a detailed study of the balance between activation and desensitization which is particularly important considering the PK profile of an a7 nAChR partial agonist. Subcortical PANSS Model Figure 2A: General overview of PANSS total computer-based model, based upon the neuroarchitecture and neurophysiology of the ventral striatum for a direct pathway medium spiny neuron (MSN). The activity of these neurons is driven by afferent projections from the cortex (modulated by D2R), and background gating signals from hippocampus and amygdala and directly modulated by dopamine afferents from the ventral tegmental area through the D1 receptor. The serotonergic and noradrenergic afferents are driven by dorsal raphe and locus coeruleus activity, respectively, while the cholinergic activity is derived from tonically active interneurons. Pharmacological agents can affect the model in a large number of ways. The model outcome of clinical interventions is then compared to the actual outcome on diverse clinical scales.  The cortical Input into the model is simulated in a 120-neuron model with appropriate neurophysiology dynamics on various targets and is calibrated using primate electrophysiology data. (Fig 2B) (Roberts, submitted) GABA Spiny Neuron D1-R Rec Kir2 L-Ca L-Cl L Ksi D2-R Stimulating Pyramidal Signal Gating Signal Background Pyramidal Signal Rec Spike Train D2-R Striatum PFC Drugs may affect any receptors shown D2-R D1-R D2-R DA Rec Rec DA competition model Info Process Function Measure D3-R Rec Rec RecB RecB’ ACh D2-R Rec Glutamatergic Spike Trains Glutamatergic Spike Trains Hippocampus Amygdala Correlation Between Model Outcome and Retrospective Clinical Results Figure 3A: Correlation between clinically reported PANSS total scores and (top figure) D2R occupancy and (bottom figure) computer-based model output for the same 43 drug-dose combinations for 14 different antipsychotic drugs. The y-axis is the change in PANSS total between baseline and after treatment for particular drug-dose combination. The x-axis (top graph) is the D2R occupancy derived from raclopride experiments and the x-axis (bottom graph) is the systems pharmacology model output. The analysis suggests that the model can explain threefold more variance than the D2R occupancy calculation alone, likely because the PANSS model takes into account the physiology of many other receptors affected by the antipsychotic drugs.  The cortical model is then calibrated using the correlation between model outcome and clinical cognitive readouts on the N-back working memory test. (Fig 3B) Systemic model explains more of variance -35 -30 -25 -20 -15 -10 -5 0 300 400 500 600 700 800 900 Circuit Model Output Improvement in PANSS total R 2 = 0.696 p = 1.6 E-06 43 drug-dose combinations 16 different drugs 12,695 patients Currently used rule-of-thumb -35 -30 -25 -20 -15 -10 -5 0 0 20 40 60 80 100 D2R Occupancy Improvement in PANSS total R 2 = 0.176 p = 0.004  A detailed receptor state mathematical model of the a7 nAchR, based upon preclinical data captures the dynamical balance between activation and desensitization. (Fig 4B)  BOLDfMRI readout is implemented using detailed biophysical description of neuronal activity, glucose metabolism, vascular flow coupling, oxygen changes. (8) Results mGluR2 Activation  Stimulating the presynaptic mGluR2 autoreceptor leads to an inverse U-shape dose-response, because of the balance between excitation and inhibition in cortical network activity projecting to the N accumbens. (Fig. 5A-5B) Discussion  mGluR2 activation likely leads to inverse U-shape dose-response, because of excitation-inhibition balance in the cortical network. (9)  This balance is difficult to monitor in rodent preclinical models as neurophysiology is fundamentally different in primates. (10, 11)  Great care need to be taken in designing clinical trials with alpha-7 nAChR modulators for comedication and disease state.  Computer-based mechanistic disease-modeling in schizophrenia is a relatively inexpensive way to explore the issues associated with inverse U-shape dose-responses in CNS indications.  Identification of the processes that affect the underlying neurobiology can lead to better clinical trial design and probably a higher success rate in clinical trials. mGluR2 Activation Dose-Response Different Dose-Responses for BOLDfMRI, PANSS Total and Target Engagement 2 4 6 8 10 12 14 16 18 Change in PANSS Total 0 5 8 10 12 14 15 16 18 20 % mGluR2 Activation 0 COMT MM COMT MV COMT VV Figure 5A: Effect of increasing level of mGlUR2 activation of model performance on PANSS total. The dose-response shows an inverse U-shape that is dependent upon the COMT genotype. Maximal effect is somewhat limited and in the range of the weaker antipsychotics such as quetiapine. Figure 5B: Effect of mGluR2 activation in the computer model on target engagement (blue), BOLDfMRI readout (green) and PANSS Total (orange) showing that these modalities display a different dose- dependent shape.  Sensitivity analysis reveals that the optimal activation level and the shape of the dose-response curve depend upon a number of processes that can be modulated by the COMT genotype. (Fig 5A)  For instance, optimal treatment window for mGLuR2 activation is between 8-14% for COMT MM and 4-11% for COMT VV.  Target engagement does not reflect the functional response, while BOLDfMRI readout broadly but not exactly reflects the functional dose- response. Alpha-7 nAChR Activation  Alpha-7 nAChR partial agonists, but not Positive Allosteric Modulators (PAM) decrease peak current but increase total area for a regular cholinergic firing frequency. (Fig 6A) 0 5 10 mGluR2 Activation(%) 2 0 15 20 25 30 35 4 6 8 10 12 14 16 18 20 Readout PANSS BOLDfMRI Target engagement Cortical Model  Positive modulation of the a7 nAchR also leads to an inverse U-shape dose- response in the neuronal network because of the balance between activation and desensitization at the level of the receptor. (Fig 6B)  The dose-response shape is dependent upon the ambient basal level of free Ach, which might be modulated by the comedication (some antipyschotics, like haldol, risperidone, olanzapine, but not aripiprazole increase ambient free Ach level) or the disease state (schizophrenia or Alzheimer’s disease).  Optimal doses are different for Alzheimer’s disease and for schizophrenia. Alpha-7 nAChR Partial Agonist in Vivo -0.5 0 0.5 1 1.5 2 2.5 3 3.5 4 0 50 100 150 200 250 300 350 400 450 500 Time (msec) Current 0 nM 15 nM 40 nM 125 nM Cpd A on total current 0 1000 2000 3000 4000 5000 6000 0 100 200 300 400 500 Conc Cpd A (nM) Total current Total 2000 nM Ach Total 1500 nM Ach Total 1000 nM Ach Total 500 nM Ach Optimal AD window Optimal SZ window Figure 6A: Time-dependent current through a postsynaptic alpha-7 nAChR receptor when applying a partial a7 nAchR agonist at different concentrations for a cholinergic stimulation frequency of 6Hz. The results show that a partial agonist reduces peak current, but increases total current over a certain time period. Figure 6B: Dose-response of a partial a7 nAChR agonist on the a7 nAchR postsynaptic receptor current, which is proportional to network performance for different ambient Ach levels. The optimal dose for a low ACh ambient condition (Alzheimer’s disease) is different from the optimal dose for a more normal (i.e. schizophrenia) condition. In addition, certain comedications can affect the level of free Ach in the human brain. Different Dose-Response for a7 nAChR Partial Agonist in Alzheimer’s Disease vs Schizophrenia References 1. Thomsen MS, et al. Curr Pharm Des. 2010;16(3):323-43. 2. Patil ST, et al. Nat Med. 2007;13(9):1102-7. 3. Coyle JT. Cell Mol Neurobiol. 2006;26(4-6):365-84. 4. Lewis DA. Dev Neurobiol. 2011;71(1):118-27. 5. Winterer G, Weinberger DR. Trends Neurosci. 2004;27(11):683-90. 6. Winterer G, et al. Am J Psychiatry. 2004;161(3):490-500. 7. Cosgrove KE, et al.. Mol Neurobiol. 2011;44(1):93-101. 8. Sotero RC, et al. NeuroImage. 2008;39(1):290-309. 9. Isaacson JS, Neuron. 2011;72(2):231-43. 10. Povysheva NV, et al. Neurophysiol. 2007;97(2):1030-9. 11. Povysheva NV, et al. Cereb Cortex. 2006;16(4):541-52.