Astrocytes and cell culture
Fetal human brain tissue was obtained through StemExpress under a protocol approved by the Stanford University Institutional Review Board. Astrocyte cultures were obtained and maintained according to Zhang et al., 2015 (which contains a complete, step-by-step protocol). Briefly, fetal brain tissue was cut into ~1mm3 pieces and enzymatically digested in 7.5 unit/ml papain (Worthington, LS 03126) at 34°C for 30 minutes. Tissue mechanically dissociated to produce a single cell suspension in a protease inhibitor solution (Worthington LS003086). Cells were resuspended in PBS with BSA (Sigma A8806) and DNase (Worthington, LS002007) and passed it through a Nitex ((Tetko Inc, HC3-20)) filter to remove cell clumps. Cells were then incubated for 30min at room temperature on a petri dish coated with anti-CD45 (BD Pharmingen, 550539) antibody to negatively select for microglia/macrophages followed by a petri dish coated with anti-Thy1 (CD90; BD, 550402) to negatively select for neurons. The cell suspension was then incubated for 45 minutes on an anti-HepaCAM (R&D systems, MAB4108) plate to select for astrocytes. The plate was washed with PBS to remove unattached cells and the purified astrocytes removed from the plate by 0.05% trypsin (Sigma, T9935) digestion for 3 minutes at 37°C. Collected cells were stored in Bambanker reagent (BulldogBio BB01) according to manufacturer’s instructions.
Astrocytes were cultured in a defined, serum-free base medium containing 50% neurobasal (Gibco, 21103-049), 50% DMEM (Invitrogen, 11960-044), 100 U/ml penicillin + 100 μg/ml streptomycin (Invitrogen, 15140-122), 1 mM sodium pyruvate (Invitrogen, 11360-070), 292 μg/ml l-glutamine (Invitrogen, 25030-081), 1× SATO (see Zhang et al, 2015), 5 μg/ml of N-acetyl cysteine (NAC, Sigma, A8199), and 5ng/ml HBEGF.
Synaptosome preparation
Synaptosomes were prepared as published previously 18,44. All procedures involving animals were conducted in conformity with Stanford University guidelines that are in compliance with national and state laws and policies. Synaptosomes were prepared from whole Sprague Dawley rat cortex as published previously44 and conjugated to pHrodo Red, succinimidyl ester (Thermo Fisher Scientific, P36600) in 0.1 M sodium carbonate (pH 9.0) at room temperature with gentle agitation45. After a 2-h incubation, unbounded pHrodo was washed out by multiple rounds of centrifugation and pHrodo-conjugated synaptosomes/myelin were re-suspended with isotonic buffer containing 5% DMSO for subsequent freezing.
High-throughput screening
Astrocytes were plated on day 0 into solid-black 384 well Greiner plates PDL coated (EK-30946) plates at 1000 astrocytes/well, in 50uL media. After 3 days in culture, we used the SciClone ALH3000 to add 200 nL of all screen compounds (Supplementary File 1) to the plates, and then added 40uL of additional media with 0.1% serum (all columns) and 0.5uL/well of pHrodo (pHrodo™ Red AM Intracellular pH Indicator) containing labeled synaptosomes (columns 1-22 only). After an additional 24 hours incubation, we added 10uL of 50uM Calcein AM (Fisher C3100MP) to the plates and imaged cells (see below). The screening protocol is also included in our GitHub repository (https://github.com/jenwilson521/phagocytosis_beta2Agonists).
Compound libraries and dosing
We screened the Library of Pharmacologically Active Compounds (LOPAC1280) (Sigma, St Louis, MO, USA) and Microsource Spectrum (MS) compound libraries. The LOPAC and MS libraries included 1,280 and 2,000 unique compounds respectively and all compounds were dosed at 1.39, 2.78, 5.56, 11.11, and 22.22 uM in singlicate. All raw cell imaging data and plate data are included in Supplementary File 1 (sheet names “Well Data” and “Plate Data”). For each plate, we included four rows of cells with media, without any compound (“High Control”), one row of postive_control_compound (“Positive Control”) and media without cells or compound (“Low Control”).
Image analysis and data normalization
After adding calcein AM, the plates were moved into an incubator in a separate IXMicro room where we used a CRS robot to load and image plates using an ImageXpress Micro. We used FITC-FIXED and TRITC-FIXED cubes and used a 600 ms exposure for each. Within each plate, we calculated the median total cells in the high and low controls, the median calcein AM integrated intensity (“green”) and median pHrodo phagocytosis integrated intensity (“red”) of all non-control, data wells, and calculated the median %phagocytosis integrated intensity adjusted for total calcein AM signal (a red/green ratio that represented the relative amount of phagocytosis activity while controlling for live cells). After completing plate-normalization, we calculated the total number of cells with pHrodo red and calcein green signal and calculated the red:green ratio per cell. We then further normalized this ratio to the percent of the median of red:green ratios in the plate. We considered a compound to be toxic if the integrated calcein signal within a well was 30% less than the median plate calcein signal.
Using the normalized red:green ratios, we fit a dose-response curve and estimated the EC50 of pHrodo-red induction or inhibition as a measure of the drug’s effect on phagocytosis. We considered compounds with 100%, <70%, and >130% of the median red:green ratio when the [concentration] > 5uM as compounds as no-effect, decreasing, or increasing phagocytosis respectively. We generated summary images from this analysis using (count_screen_data.py) provided in the GitHub repository.
Incucyte validation of salmeterol
1000 fetal human astrocytes per well were plated on Greiner Bio-One CELLSTAR μClear™ 96-well, Cell Culture-Treated, Flat-Bottom Microplates and cultured for 3 days to establish cell viability. Cells were then treated with various doses of salmeterol as well as 0.1% v/v of fetal bovine serum (Thermo Fisher 10437028) to provide opsonins for phagocytosis. One hour later, cells were treated with 5ul of pHrodo-conjugated synaptosomes and 50uM Calcein AM (live stain) and imaged at 1h intervals. Images were acquired with an Incucyte ZOOM (Essen Bioscience) at 37°C and 10% CO2. For image processing analysis, we took 3 images per well using a 20× objective lens from random areas of the 96-well plates. Custom analysis scripts within the Incucyte ZOOM Software were used to measure phagocytosed particles (pHrodo-positive area) and live cells (Calcein AM area), and intensity and particle size cutoffs were used to discriminate signal from noise.
Running PathFX analysis
We ran PathFX analysis for 239 drugs from the three different experimental groups — increase, decrease, or no effect on phagocytosis — using the PathFX algorithm (version 1.0)15. We also ran PathFX using ADRB2 as a single drug target to simulate effects of ADRB2 agonists, such as mabuterol (analysis contained in the script, run_PathFX_mabuterol.py). We ran PathFX using default parameters and provided DrugBank identifiers as inputs. This analysis was completed in four scripts: run_multiDrug_incease.py, run_multiDrug_decrease.py, run_multiDrug_noeffect.py, and run_PathFX_mabuterol.py. We selected the 239 drugs from the phagocytosis screen (see above) based on whether they had protein-binding targets in DrugBank21. Out of the 239 total drugs selected for PathFX analysis, 52 were determined to increase phagocytosis, 43 decreased phagocytosis, and 144 had no effect on phagocytosis based on the primary in vitro screen.
Uncovering network associations to multiple neuropsychiatric disorders
Using the network and phenotype tables identified from PathFX, we identified drug network genes associating the drug targets to neuropsychiatric disorders. We searched for 77 different psychiatric and immune phenotypes and grouped these 77 neuropsychiatric phenotypes into 10 different clusters based on clinical similarity (Neuropsychiatric_clusters.xlsx). We counted drug networks that were associated with these phenotypes for each of the three groups. This analysis was also completed in three scripts: phenotype_counting.py, phenotype_counting_decrease.py, and phenotype_counting_noeffect.py. These scripts generated an excel matrix (Phenotype_counting.xlsx) that stores whether a phenotype (represented by its CUI identifier) is present in the network of a drug by marking a ‘0’ (for not present) or a ‘1’ (for present) in the cells of the sheet.
Gene counting and phenotypic grouping
We next analyzed network genes associated with psychiatric diseases. We analyzed all schizophrenia-associated genes using the gene-phenotype data from PathFX using the following script: get_schizophrenia_genes.py. This yielded a list of 1,947 genes, aggregated from a variety of different sources, including ClinVar, OMIM, PheGenl, and DisGeNet (as published in15). We identified schizophrenia-associated genes in the 239 drug pathways. This analysis was completed in three scripts: gene_counting.py, gene_counting_decrease.py, and gene_counting_noeffect.py.
GO enrichment
For individual drugs, we conducted GO analysis of PathFX-identified schizophrenia-associated genes (all contained in Supplementary File 2) using the GOrilla tool (http://cbl-gorilla.cs.technion.ac.il/) 46, using a target (mabuterol_schizophrenia_genes.txt) and background (all_schizophrenia_genes.txt) list, selecting for molecular process, function, and component GO ontologies, and using Homo Sapiens as the organism. The full GO enrichment results for mabuterol are contained in Supplementary File 6. We also used GO enrichment to understand proteomic changes measured in the mouse brain. We again used the Gorilla tool but converted significantly changed mouse proteins to their human homologues and used this as the foreground list compared to all schizophrenia proteins and all network proteins. Both comparisons are contained in Supplementary File 8.
Clustering and meta-analysis
To look for network patterns, we clustered drugs based on phenotype-associated genes across experimental groups using the clustering.py script using the fastcluster module in python and one-hot encoding of network proteins. We further modified our initial clustering to create a clarified heatmap image (all_drugs_clustermap.png for schizophrenia). This analysis is contained in clustering_update_fig.py.
Logistic Regression
We used logistic regression to discern network patterns associated with the ‘Increase,’ ‘Decrease,’ or ‘No effect’ phagocytosis experimental groups. For this analysis, we merged the gene counting results from each experimental group into a single matrix (convert_to_single_matrix.py). Then we ran the logistic regression using the LogisticRegression module in python; this analysis is contained in the following two scripts: call_logistic_regression.py and run_logistic_regression.py. These scripts generated an excel sheet for each phenotype, labeled by CUI term, which ranks the genes for each experimental group in order of regression coefficients. For example, the schizophrenia logistic regression results can be found here: Logistic_regression_results/Gene_counting_C0036341_mergedregCoeff.xlsx.
In vivo Model
We investigated the effects of a selective beta-2 adrenergic receptor (ADRB2) agonist, mabuterol in 3-month-old 5XFAD male mice from MMRRC (JAX #034840-JAX). This neuroinflammatory model of amyloidosis was selected because previous studies have evidence of altered synaptic pruning22. Mice were treated daily for 2 months with vehicle (0.9% Saline) or 0.3mg/kg mabuterol. Behavioral testing, including Activity Chamber and Y-maze: Forced Alternation, was performed at set time points as described below. Methods for each behavioral test are described below. Mice were group-housed under a reversed light-dark cycle with lights off at 8:30 AM and on at 8:30 PM. Mice were handled prior to the experiments to habituate mice to interacting with the experimenter. All procedures related to animal maintenance and experimentation were approved by the Stanford University Administrative Panel for Laboratory Animal Care and conformed to the U.S. National Institutes of Health Guide for the Care and Use of Laboratory Animals. Efforts were made to minimize the number of mice used and their suffering.
Activity Chamber
The Activity Chamber (AC) was used to assess general locomotor activity and exploration as described previously 24. Briefly, mice were placed in one corner of a square Open Field Activity Arena (43x43x30 cm; Med Associates Inc., St. Albans, Vermont; Model ENV-515) located inside of a dark sound-attenuated chamber (74×60×60 cm) and allowed to freely explore the arena. Movement was tracked by an automated tracking system with three planes of infrared detectors during a 10-min trial. Parameters measured included distance moved, vertical counts (rearing), and time spent in the periphery and center of the arena. The periphery was defined as the zone within 5 cm of the arena wall. Between each trial, the surface of the arena was cleaned with 1% Virkon disinfectant. AC was conducted at baseline, 4-, and 7-weeks post-dosing.
Y-Maze: Forced Alternation
The Y-maze was used to assess behavior believed to be associated with hippocampal-dependent spatial reference memory. This test is based on the tendency of rodents to preferentially explore a novel environment over a familiar one. In this case, a normal rodent prefers to explore a different arm of the maze than an arm they previously explored. The maze was made of plastic with 3 arms in a “Y” shape (each arm 40x8x15cm). The test consisted of two 8-min trials separated by a 1-hour intertrial interval (ITI) as described in 22. To start each trial, mice were placed at the end of one of the arms (Start Arm). During the first trial (Training), mice were only allowed to explore two of the three arms (Familiar Arms). A plastic insert blocked off the third arm (Novel Arm). The Novel Arm was pseudorandomized to avoid any location bias. During the second trial (Testing), the insert was removed, and the mice were allowed to explore all three arms. The trials were recorded with an overhead camera and tracked with Ethovision XT (Noldus Information Technology, Wageningen, Netherlands). Between each trial, the surface of the maze was cleaned with 1% Virkon disinfectant. Y-maze was conducted 4- and 7-weeks post-dosing.
Tissue Collection
Brain and plasma were collected after 2 months of dosing. Mice were dosed 1 hour (±15 minutes) prior to tissue collection. Mice were deeply anesthetized with isoflurane. Prior to perfusion, whole blood was collected from the left ventricle via cardiac puncture (23 g needle) into K3EDTA-containing vials (Greiner Bio-One, MiniCollect Tube Reference #450475). Blood was spun (10min, 3000G, 4°C) within 1 hour of collection, and plasma was stored at -80°C. For perfusion, the right atrium was opened, and mice were transcardially perfused with ice-cold phosphate-buffered saline (PBS; pH 7.4) through a 23 g needle. The perfused brain was removed. The brain was bisected coronally at the level of the mammillary bodies into the forebrain and hindbrain. The forebrain was hemisected. The left hemisphere was immediately flash-frozen on dry ice and stored at -80°C for later analysis. The right hemisphere and hindbrain were post-fixed with 4% paraformaldehyde in a 15ml conical centrifuge tube (48 hours, 4°C). Following post-fixation, the hemisphere and hindbrain were transferred to 30% sucrose in phosphate buffer (PB) and stored at 4°C for later analysis.
Multiplex mouse cytokine assay
Multiplex tissue cytokines were analyzed in plasma and brain homogenate using a Luminex 48-plex (Affymetrix) mouse cytokine assay as described in 23. Brain homogenate was prepared from a 2 mm coronal section of the left hemisphere just anterior to the hippocampus. The Luminex assay was performed in the Human Immune Monitoring Center at Stanford University, following manufacturer instructions. Briefly, hippocampal tissue was homogenized in a protein extraction buffer (1% triton X100, 0.5% tergitol, 25 mM Tris HCl, 100 mM NaCl containing Halt protease inhibitor cocktail, 1x and 1 uM phenylmethanesulfonyl fluoride, 1x). The tissue was lysed by pulling through a 23 g needle (10×) and then sonicated for 3 × 3 second pulses. Homogenate was spun at 14,000 g for 10 min, and protein concentrations were determined by Pierce BCA assay. Homogenate samples were diluted to a common concentration of 6 μg/uL. Plasma samples were diluted 1:3. Plasma and brain homogenate samples were run in singlet on a 96-well plate alongside standard curve and quality control calibration samples. Significance was assessed using Šidák’s posthoc after 1-way ANOVA.
Proteomics
Proteins were analyzed from a 2 mm coronal section of the left hemisphere posterior to the olfactory bulb as previously described 22. Proteomics analyses were performed at the Vincent Coates Foundation Mass Spectrometry Laboratory, Stanford University Mass Spectrometry (SUMS - RRID:SCR_017801). Lysis buffer (5% SDS, 50 mM TEAB, and 1X Protease and Phosphatase Inhibitors) was added to tissue samples, and they were homogenized by bead mill. The resulting lysate was cleared and transferred for filter supported digestion. Proteins were reduced with 10 mM DTT at 550°C for 30 minutes, followed by alkylation with 30 mM acrylamide for 30 minutes at room temperature. 0.5 μg of Trypsin/LysC protease (Promega) was added to each sample for digestion at 37°C overnight. After digestion, the reaction was quenched using 1% formic acid, and peptides were eluted and dried. Peptide quantification was performed with the Pierce Quantitative Fluorometric Peptide Assay kit (Thermo Fisher Scientific). The peptide mixture was dried by speed vac before dissolution in reconstitution buffer (2% acetonitrile with 0.1% formic acid). 1 μg was used for subsequent LC-MS/MS analysis. The mass spectrometry experiment was performed using an Orbitrap Eclipse Tribrid mass spectrometer RRID:022212 (Thermo Scientific, San Jose, CA, USA) with liquid chromatography using an Acquity M-Class UPLC (Waters Corporation, Milford, MA, USA). A flow rate of 300 nL/min was used, where mobile phase A was 0.2% formic acid in water and mobile phase B was 0.2% formic acid in acetonitrile. Analytical columns were prepared in-house with an I.D. of 100 microns pulled to a nanospray emitter using a P2000 laser puller (Sutter Instrument, Novato, CA, USA). The column was packed using C18 reprosil Pur 1.8 micron stationary phase (Dr. Maisch) to a length of ~25 cm. Peptides were directly injected onto the analytical column using a gradient (2%–45% B, followed by a high-B wash) of 80 min. The mass spectrometer was operated in a data-dependent fashion using CID fragmentation for MS/MS spectra generation. For data analysis, the RAW data files were processed using Byonic v4.1.5 (Protein Metrics, Cupertino, CA, USA) to identify peptides and infer proteins. Proteolysis with Trypsin/LysC was assumed to be semi-specific, allowing for N-ragged cleavage with up to 2 missed cleavage sites. Precursor mass accuracies were held within 12 ppm and 0.4 Da for MS/MS fragments. Cysteine modified with propionamide was set as fixed modifications in the search. Proteins were held to a false discovery rate of 1%, using the standard reverse-decoy technique 47. Significant protein-level changes were assessed using a t-test.
Western Blot
Flash-frozen, 2 mm coronal partial section containing hippocampus and cortex from the left hemisphere was homogenized in 10 ul/mg of T-PER (Tissue Protein Extraction Reagent; Thermo Scientific, Cat: 78510, Waltham, MA, USA) with Halt Protease Inhibitor Cocktail (Thermo Scientific, Cat: 78429, Waltham, MA, USA) and Phosphatase Inhibitor Cocktails (Abcam, Cat: ab201112, ab201113, ab201114, Cambridge, UK) on ice by sonication using Ultrasonic Probe Homogenizer (Omni International, Kennesaw, GA, USA). The homogenate was centrifuged at 12,000 rpm for 10 min at 4°C. The protein concentration was determined using the Pierce BCA (bicinchoninic acid) protein assay kit (Pierce, Cat: 23227 Rockford, IL, USA). Samples were prepared with Novex Bolt lithium dodecyl sulfate sample buffer and Novex Bolt sample reducing agent (Invitrogen, Cat: B0007, B0009). For some antibodies, as noted below, samples were then boiled at 95°C for 5 min. Samples were loaded 20 ug/well in 10% or 4-12%, Bis-Tris, 1.0 mm, Mini Protein Gel (Invitrogen, Cat: NW00107BOX, NW04122BOX, Waltham, MA, USA). The protein was transferred to a polyvinylidene difluoride membrane (Abcam, Cat: ab133411, Cambridge, UK) and incubated in Intercept (TBS) Blocking Buffer (Li-cor, Cat: 927-60001, Lincoln, NE) for 1 hour at room temperature. The membranes were incubated at 4°C overnight on a shaker with anti-Akt (1:2000, Cell Signaling, Cat: 2920, Danvers, MA, USA), anti-Atg5 (1:1000, Cell Signaling, Cat: 12994, Danvers, MA, USA), anti-DAPK1 (1:1000, Cell Signaling, Cat: 3008, Danvers, MA, USA), anti-phospho-DAPK Ser308 (1:1000, Sigma-Aldrich, Cat: D4941, Burlington, MA, USA), anti-GluA1 (1:1000, Cell Signaling, Cat: 13185, Danvers, MA, USA), anti-LC3B (1:500, Novus Biologicals, Cat: NB100-2220, Littleton, CO, USA), anti-NR2B (1:1000, Invitrogen, Cat: MA1-2014, Waltham, MA, USA), anti-PSD95 (1:1000, Cell Signaling, Cat: 2507, Danvers, MA, USA), anti-synapsin-1 (1:1000, Cell Signaling, Cat: 5297, Danvers, MA, USA), anti-phospho-synapsin-1 Ser62, Ser67 (1:1000, Millipore, Cat: AB9848, Burlington, MA, USA), anti-synaptophysin (1:1000, Millipore, Cat: MAB329-C, Burlington, MA, USA), and anti-tubulin (1:1000, Sigma-Aldrich, Cat: T5168, Burlington, MA, USA) primary antibodies. Samples were boiled, as described above, except for those used with anti-Akt, anti-Atg5, anti-DAPK1, and anti-NR2B. The following day, membranes were washed (4×10 min) with 0.01% Tween-20 in 1x TBS and incubated for 1 hour at room temperature on a shaker with IRDye IgG Secondary Antibody (1:10000, goat anti-mouse Cat: 926-68070, goat anti-rabbit Cat: 926-32211, goat anti-mouse Cat: 926-32210, goat anti-rabbit Cat: 926-68071, Li-cor, Lincoln, NE, USA). Following secondary antibody incubation, membranes were washed (4×10 min) with 0.01% Tween-20 in 1x TBS. Membranes were then scanned with the Sapphire Biomolecular Imager (Azure Biosystems, Dublin, CA, USA) in the appropriate wavelengths. Azurespot Version 2.0 (Azure Biosystems, Dublin, CA, USA) was used for densitometry analysis of target protein levels and normalized to internal level of tubulin or the non-phosphorylated protein for each sample as a control. In order to stain membranes with multiple primary antibodies, membranes were stripped by incubating in 1X Restore Fluorescent Western Blot Stripping Buffer (Thermo Scientific, Cat: 62300, Waltham, MA, USA) for 20 minutes at room temperature on a shaker. Membranes were washed (4×5 min) with 0.01% Tween-20 in 1x TBS and incubated in Intercept (TBS) Blocking Buffer for 1 hour at room temperature. Membranes were then incubated in respective primary antibodies overnight as described above.
Analysis of clinical records
We designed an observational study to assess the effects of a pediatric exposure to ADRB2 agonists on schizophrenia-associated in-patient visits following a young-adult diagnosis of schizophrenia. The Optum Clinformatics™ Data Mart Database (OptumInsight, Eden Prairie, MN) is a de-identified database from a large national insurance provider. The dataset contains over 88 million patients largely under the age of 65 and is frequently used for observational studies. We used a version of Optum standardized to OHDSI’s Observational Medical Outcomes Partnership (OMOP) common data model (CDM) version 5 (https://github.com/OHDSI/CommonDataModel). The OMOP CDM used standard vocabulary concepts mapped to international coding systems into a consolidated data resource.
We implemented the analysis using the CohortMethod package 32 in R to fit a Cox Regression model for assessing differences between target and comparator patient populations. We designed the study following OHDSI guidelines and generated parameterized code to assess data from the Optum data set in CDM v 5.0 and make this code available in the EHR_code folder in the GitHub. The full, anonymized analysis pipeline is contained in EHR_code/count_schizophrenia_drugs_patiens_beta2agonists_PrimaryDiagnosis_anonymized.R. This analysis consisted of multiple selection steps and to increase transparency we have highlighted which script completes each step:
- We selected all child drug concepts relative to the “Selective beta-2-adrenoreceptor agonists“ ATC drug concept, and this included 540 individual drug concepts (e.g. “Albuterol 4 MG Oral Tablet [Nu-Salbutamol]” or “Terbutaline 5 MG Oral Tablet [Bricanyl SA] Box of 60 by Astrazeneca”); full concept list in Supplementary File 9, EHR_code/beta2agon_pediatric_exposure_noPriorObsReq.sql.
- We defined all young adult patients (age > 18) with a schizophrenia diagnosis using EHR_code/Schizophrenia_young_adult_diagnosis_noPriorObsReq_v2.sql.
- We defined our target cohort as patients with the pediatric exposure and the young adult diagnosis (EHR_code/count_shared_patients.sql).
- We defined our comparator cohort as young adult patients without the pediatric exposure (EHR_code/remove_doubles_from_single_cohort.sql).
- We required in-patient visits to be associated with a diagnosis of schizophrenia on the date of the visit or up to 3 days post visit to accommodate diagnoses logged for billing purposes (EHR_code/inpatient_pysch_visit_new_inclusion_rule.sql).
- We fit a large-scale propensity model to control for confounding and the model AUC was 0.87 suggesting that covariates were not sufficiently predictive of treatment assignment. We used inverse propensity weighting (IPW) and patient matching to compare target and comparator outcomes. Patient matching generates a subset of target-comparator patient comparison based on covariate similarity, but eliminates patients when a reasonable match is not available. IPW instead uses all patients but weights their contribution to hazard ratio estimation based on the representation of their covariate profile in the study population. Because we had relatively few patients, we considered both methods in this exploratory analysis. (EHR_code/ count_schizophrenia_drugs_patiens_beta2agonists_anonymized.R).