Perturbed iron biology in the prefrontal cortex of people with schizophrenia

Despite loss of grey matter volume and emergence of distinct cognitive deficits in young adults diagnosed with schizophrenia, current treatments for schizophrenia do not target disruptions in late maturational reshaping of the prefrontal cortex. Iron, the most abundant transition metal in the brain, is essential to brain development and function, but in excess, it can impair major neurotransmission systems and lead to lipid peroxidation, neuroinflammation and accelerated aging. However, analysis of cortical iron biology in schizophrenia has not been reported in modern literature. Using a combination of inductively coupled plasma-mass spectrometry and western blots, we quantified iron and its major-storage protein, ferritin, in post-mortem prefrontal cortex specimens obtained from three independent, well-characterised brain tissue resources. Compared to matched controls (n = 85), among schizophrenia cases (n = 86) we found elevated tissue iron, unlikely to be confounded by demographic and lifestyle variables, by duration, dose and type of antipsychotic medications used or by copper and zinc levels. We further observed a loss of physiologic age-dependent iron accumulation among people with schizophrenia, in that the iron level among cases was already high in young adulthood. Ferritin, which stores iron in a redox-inactive form, was paradoxically decreased in individuals with the disorder. Such iron-ferritin uncoupling could alter free, chemically reactive, tissue iron in key reasoning and planning areas of the young-adult schizophrenia cortex. Using a prediction model based on iron and ferritin, our data provide a pathophysiologic link between perturbed cortical iron biology and schizophrenia and indicate that achievement of optimal cortical iron homeostasis could offer a new therapeutic target.

. Clinical and postmortem characteristics of samples obtained from three brain banks. Table S2. Disease-related characteristics of patients in our samples. Table S3. Antibodies used for detection and quantification of ferritin in brain tissue. Table S4. Robust normality statistics for iron distribution in the Combined Cohort, grouped by diagnosis. Table S5. Robust between-group comparison of PFC iron levels in the Combined Cohort, visualized in Figure 1a. Table S6. Robust measures of iron distribution in the Combined Cohort. Table S7. Predicting iron content by diagnosis, controlling for demographic and tissue-quality variables one at a time and simultaneously. Table S8. PFC iron across diagnosis based on regression-adjusted propensity-score matching. Table S9. PFC iron across diagnosis based on regression-adjusted propensity-score matching, including mode of death as a covariate.
Table S10. Predicting iron content by diagnosis, controlling for smoking.   Table S14. Predicting iron content by antipsychotic treatment, controlling for onset age and duration of illness. Table S15. Predicting iron content by type of antipsychotics used. Table S16. Predicting iron content by results of a postmortem toxicology essay for antipsychotics. Table S17. Robust normality statistics for copper distribution in the Combined Cohort, grouped by diagnosis. Table S18. The effect of diagnosis on PFC copper levels in the Combined Cohort, related to Figure S7a. Table S19. Robust measures of copper distribution in the Combined Cohort. Table S20. Predicting iron content by copper, controlling for diagnosis and copper × diagnosis interaction term, related to Figure S7b. Table S21. Robust normality statistics for zinc distribution in the Combined Cohort, grouped by diagnosis. Table S22. The effect of diagnosis on PFC zinc levels in the Combined Cohort, related to Figure   S8a. Table S23. Robust measures of zinc distribution in the Combined Cohort. Table S24. Predicting zinc content by iron, controlling for diagnosis, related to Figure S8b. Table S25. Exploring the effect of iron on diagnosis through zinc, related to Figure S9. Table S26. Robust normality statistics for ferritin distribution in the Combined Cohort, grouped by diagnosis. Table S27. Robust between-group comparison of PFC ferritin levels in the Combined Cohort, visualized in Figure 1b. Table S28. Robust between-group comparison of PFC ferritin levels in the Combined Cohort, controlling for covariates. Table S29. Robust between-group comparison of PFC iron-to-ferritin ratio in the Combined Cohort, visualized in Figure 1c. Table S30. Robust analysis of the relationship between PFC iron and age among control individuals, visualized in Figure 2a. Table S31. Robust analysis of the relationship between PFC iron and age among schizophrenia cases, visualized in Figure 2a. Table S32. Logistic regression analyses predicting disease status based on PFC iron for sequential age cutoffs, visualized in Figure 2b. Table S33. Logistic regression analyses predicting disease status based on PFC iron in young and old subcohorts, visualized in Figure 2c. Table S34. Robust between-group comparison of PFC iron levels in the young and old subcohorts, visualized in Figure 2d. Table S35. Robust analysis of the relationship between PFC ferritin and iron among control individuals, visualized in Figure 3a. Table S36. Robust analysis of the relationship between PFC ferritin and iron among schizophrenia cases, visualized in Figure 3a. Table S37. Predicting ferritin content by iron, controlling for diagnosis and iron × diagnosis interaction term, visualized in Figure 3a. Table S38. Logistic regression analyses predicting disease status based on PFC ferritin for sequential iron cutoffs, visualized in Figure 2b. Table S39. Logistic regression analyses predicting disease status based on PFC ferritin in lowiron and high-iron subcohorts, visualized in Figure 3c. Table S40. Robust between-group comparison of PFC ferritin in low-iron and high-iron subcohorts, visualized in Figure 3d. Table S41. Deriving optimal cutpoints, visualized in Figure 4a-b. Table S42. Logistic regression model predicting disease based on a combination of three predictors following definition of optimal cutpoints. Table S43. Likelihood-ratio tests comparing nested and full regression models. Table S44. Deriving an optimal cutoff for a logistic regression model combining all three predictors. Table S45. Genome-wide association of iron regulatory proteins and schizophrenia based on GWAS results of CLOZUK. Table S46. IREB2 -Association with schizophrenia and integrative (Mendelian randomization) analysis. Table S47. Differential methylation of TAX1BP1 in schizophrenia prefrontal cortex.

Tissue collection
New South Wales Brain Tissue Resource Center (NSW-BTRC) As previously described 1 , in RNAse-free conditions, blocks for tissue sections (14 µm) and for pulverisation/homogenisation were dissected from fresh frozen coronal slabs of prefrontal cortex which were identified as being on the ventral surface below the inferior frontal gyrus, and anterior to the appearance of the lateral ventricle. All blocks for sectioning contained gyrus rectus (Brodmann area [BA] 11) on the medial side. For homogenisation, approximately 0.5 g of predominantly grey matter tissue was excised on a dry ice platform using a dental drill (Cat# UP500-UG33, Brasseler, USA) at moderate speed (up to 40,000 rpm) to minimize heat generation during the dissection. Tissue from each case was then pulverised over a frozen tray placed in dry ice, weighed while frozen, and stored at -80ºC pending further analysis.

Victoria Brain Bank Network (VBBN)
The left hemisphere was removed at autopsy, rapidly processed and frozen to −80 °C using a standardised procedure 6 by the same individual in a way designed to minimise autolytic effects 7 .
The pH of the brain tissue (BA 11) was measured as described previously 8 as this is the best indicator of tissue preservation 9 .
NIMH Human Brain Collection Core (NIMH-HBCC) Each brain was sliced into ~1 cm slabs, which were barcoded and stored in individual plastic bags in -80°C freezers until dissections of circumscribed brain regions (BA 10) were performed using a dental drill or a scalpel. Whole blood was sent to the National Medical Services Lab for toxicological analysis.

Preliminary analysis and data transformation
Demographic and tissue-quality variables were compared across diagnostic groups using independent-samples t-tests for continuous variables or Pearson's chi 2 test (Fisher's exact when expected cell counts<5) for categorical variables. Preliminary analyses were performed separately for the NSW-BTRC samples and jointly for VBBN+NIMH-HBCC samples, as the latter were analyzed later and the VBBN group was relatively small to be analyzed on its own. For assessing differences in iron levels across diagnostic groups we used an independent-samples t-test with Welch correction (NSW-BTRC) and an ANCOVA (joint samples from VBBN+NIMH-HBCC, controlling for tissue origin) with robust standard errors. Following replication of our major iron finding in both preliminary analyses, we decided to generate a Combined Cohort based on controlderived z-scores. Specifically, using in-house Stata code, for each cohort separately we performed a z-score transformation on the data of the control group, and then adjusted the values of the schizophrenia cases group based on the control group's mean and standard deviation (SD). All subsequent analyses were carried out on z-scores across the entire Combined Cohort, except for specific endpoints (e.g. copper levels, iron levels among smokers) that were measured only in subsets of the combined cohort. In the case of heavily skewed distributions (as opposed to relatively normal distributions containing outliers), such as lifetime cumulative antipsychotic dose, a logtransformation was applied.

Linear regression models
Assessing covariates. Demographic data (i.e. sex, age, race, death circumstances [non-suicide vs. suicide]) and tissue-quality parameters (i.e. pH and postmortem interval) were available for all specimens. Additional demographic data (i.e. smoking, weight and BMI), disease-related data (ageof-onset, duration of illness, mean-daily and lifetime-cumulative doses of antipsychotics) and postmortem toxicology data were only available for subsets of individuals. Given its central role in the manuscript, for iron we performed robust linear regressions assessing the importance and effect of each covariate, by running a series of regressions where diagnosis and other covariates serve as predictors and assessing both inference statistics regarding the covariate and change in point statistics of diagnosis coefficient attributed to the addition of covariates. In the case of iron, we also assessed the effect of diagnosis following propensity score matching (also considering the death circumstances) using Stata's 'kmatch' command.
Rationale for robustness. In regression analysis, the presence of outliers in the dataset can strongly distort the classical least-squares estimator and lead to unreliable results. To deal with this, several robust-to-outliers methods have been proposed in the statistical literature. When initially fitting a least squares (OLS) regression to our data, we found some outliers and high leverage data points.
We have decided that these data points were not data entry errors, neither were they from a different population than most of our data, so we had no compelling reason to exclude them from the analysis. In this scenario, robust regression seemed to be a good strategy since it is a compromise between excluding these points entirely or alternatively including all the data points while treating them equally using OLS regression. The idea of robust regression is to weigh the observations differently based on how "well behaved" these observations are. In this sense, it is a form of weighted and reweighted least squares regression. In recent years, it seems that a consensus has emerged to recommend the MM-estimators as a well-suited estimation method, because these estimators combine a high resistance to outliers and high efficiency 10 .
ROBREG. Extending the conceptual framework laid by Verardi et al 10 , throughout our analysis we have used "ROBREG: Stata module providing robust regression estimators," 11 . Specifically, we utilized the "robreg mm" command, which fits an efficient high breakdown MM-estimator. On the first stage, a high breakdown S-estimator (using default breakdown point of 0.5) is applied to estimate the residual scale and derive starting values for the coefficients vector. On the second stage, using iteratively reweighted least squares (IRWLS), an efficient bisquare MM-estimator is applied to obtain the final coefficient estimates. Retaining efficiency (set at 95%), this regression is considered robust to both vertical and high-leverage (i.e., influential) outliers. Stata's robust standard errors can also effectively deal with minor failures in meeting other assumptions such as normality of residuals. The robreg command has been used throughout most linear model analyses, including between-group comparisons (iron, ferritin and iron-to-ferritin ratio) and relationships (e.g., antipsychotic dose-iron, age-iron, iron-ferritin, iron-copper, iron-zinc). ROBREG computes, by default, robust standard errors, which account for heteroscedasticity in regression, and thus also addresses cases of unequal variances between two-groups (e.g., iron in cases vs. controls).
ROBSTAT. As for the assumption of normality of residuals, classical statistical tests proposed to find out whether a sample is drawn from a normal distribution or not, such as the Jarque-Bera test, are based on moments of the data, yielding a zero-breakdown value (i.e., a single outlier can make the test worthless). Thus, it has been proposed that following a robust regression procedure, normality of the residuals will subsequently be tested using a robust test for normality. These robust normality tests were executed in Stata using "ROBSTAT: Stata module to compute robust univariate statistics" 12 .
FRACPOLY. Age-iron and iron-ferritin relationships were examined using Stata's "fracpoly" command, with models based on logarithmic, square-root, quadratic and cubic transformations of the independent variable compared to non-transformed models using Akaike's information criterion and Bayesian information criterion. Models with the highest information criteria were selected.

9
Logistic regression and ensuing prediction models As in linear regression, influential data points may yield biased regression coefficient estimates in logistic regression as well. However, given limited outlier-robust alternatives to standard logistic regression, we decided to identify (and exclude) influential observations and outliers using Pregibon's dbeta ("predict dbeta") which provides summary information of influence on parameter estimates of each individual observation (more precisely each covariate pattern) and is very similar to Cook's D in ordinary linear regression. For assessing if the effect of iron on diagnosis was mediated via zinc, a generalized structural equation model (GSEM) with a logit link function for the dependent variable was built, with significance of indirect effect quantified via bootstrappedderived confidence intervals. For determining discriminatory performance, receiver operating characteristic (ROC) curve analyses were performed, and optimal cutpoints were empirically determined using Stata's 'cutpt' command (Youden method). Following categorical transformation of continuous predictors using these cutpoints, the simultaneous contribution i) of iron among the subcohort of individuals who had died younger than 35 years; ii) of ferritin among the subcohort of individuals with below-mean iron and iii) of iron-to-ferritin ratio among the entire cohort was assessed using a multiple logistic regression model with categorical predictors denoting high-risk, low-risk and unavailable-data categories. The discriminatory performance of this model was compared to nested models through likelihood-ratio-test chi-square statistics.

Data mining
Mining data relating to genetic association and differential expression and methylation was performed using SZDB2.0: an updated comprehensive resource for schizophrenia research 13 . For assessing whether genes coding for of iron regulatory proteins were associated with schizophrenia, genes from a list of iron regulatory genes based on the study by McAllum et al 14 were mined in the CLOZUK GWAS 15 , where a genome-wide significant association with schizophrenia was defined as a P < 5 × 10 -8 . For IREB2, significance of a Summary data-based Mendelian (SMR) randomization analysis (PSMR) 16 was mined. For TAX1BP1, differential methylation data 17 were mined. For FTH1, FTL, IREB2 and TAX1BP1, differential expression in the schizophrenia PFC 18 was mined.

Table S1
Clinical and postmortem characteristics of samples obtained from three brain banks. 1 Significance of between-group (control vs. Scz) comparison based on independent-samples t-test (continuous variables) and Pearson -chi 2 (categorical variables). 2 Pearson -chi 2 was replaced by Fisher's exact test due to expected cell count<5. Smokers were defined as current smokers, heavy ex-smokers, or individuals with a positive postmortem toxicology essay for nicotine. 4 N, number of individuals for whom data on smoking habits or alcohol use were available. 5 Alcohol users were defined as those with a history of drinking an average of ≥20 g ethanol/day. 6 Copper was not measured in specimens from the Victoria Brain Bank Network (VBBN) as the glycerol-containing extraction buffer used during homogenization of these samples interfere with copper and thus bias its quantification. Copper value obtained from one of the NSW-BTRC control specimens was unreliable and thus excluded.

Table S2
Disease-related characteristics of patients in our samples. 1 Dichotomous variable indicating whether antipsychotic treatment had always been limited to typical agents or whether it had also included use of atypical agents. 2 Current/last CPZ equivalent dosage was calculated in all patients in all cohorts by following standard guidelines 19,20 , included were cases with mean CPZ-equivalents ≥ 50 mg/day. Due to the large range over which data were dispersed and right-skewness, log-transformation was applied.

Table S3
Antibodies used for detection and quantification of ferritin in brain tissue.

Table S4
Robust normality statistics for iron distribution in the Combined Cohort grouped by diagnosis, visualized in Figure S2a.
Generalized Jarque-Bera tests for normality (right), as suggested by Brys et al. 21

Table S6
Robust measures of iron distribution in the Combined Cohort, visualized in Figure S3.
(a) Standard deviation and variance of PFC iron among diagnostic groups, as visualized in Figure   S3a. (b) Robust tests of the hypothesis that the variance of ziron is the same across groups, including Levene's statistic (W0) and two statistics proposed by Brown and Forsythe that replace the mean in Levene's formula with alternative location estimators. The first alternative (W50) replaces the mean with the median. The second alternative replaces the mean with the 10 percent trimmed mean (W10). Note that the difference between groups remained prominent across both robust analyses, indicating that the increased iron heterogeneity in schizophrenia patients was not (primarily) driven by outliers, and thus likely represents a true biological difference among diagnostic groups.

Table S7
Predicting iron content by diagnosis, controlling for demographic and tissue-quality variables one at a time and simultaneously.
(a) Robust regression estimates predicting PFC iron content (control-derived z-scores) using either diagnosis alone (first row) or combinations of diagnosis and one additional predictor (Predictor 2, rows 2-7). For categorical variables, reference values of "0" were assigned to control individuals, female sex, Caucasian race, and natural death. t-statistics along with their corresponding significance are presented for both diagnosis and each of the other co-variates tested alongside. (b) Same as above with all covariates entered simultaneously. Based on linear regression models using iteratively reweighted least squares MM-estimators, derived using Stata's "robreg mm" command (95% efficiency, default parameters) 24 . PMI, post-mortem interval; dm01, mode of death.

Table S8
PFC iron across diagnosis based on regression-adjusted propensity-score matching, visualized in Figure S4.
Regression-adjusted propensity-score kernel matching with age, sex, race, pH and PMI as covariates, based on a logit model where diagnosis of schizophrenia was defined as Treatment.
Seven patients could not be matched, and four controls were unused. Estimated effect of diagnosis in patients following PS-matching (ATT) was similar to the non-matched effect (NATE). Based on KMATCH, Stata's module for multivariate-distance and propensity-score matching 25 . PMI, postmortem interval. ncontrols=85, nScz=86.

Table S9
PFC iron across diagnosis based on regression-adjusted propensity-score matching including mode of death as a covariate, visualized in Figure S5.
Regression-adjusted propensity-score kernel matching with age, sex, race, pH, PMI and mode of death as covariates, based on a logit model where diagnosis of schizophrenia was defined as Treatment. Two patients could not be matched, and two controls were unused. Given the gross inequality in mode of death across groups, bandwidth was considerably larger in this analysis compared to that of Table S8. Estimated effect of diagnosis in patients following PS-matching (ATT) was only slightly weaker compared to the non-matched effect (NATE). Based on KMATCH, Stata's module for multivariate-distance and propensity-score matching 25 . PMI, postmortem interval; dm01, mode of death. ncontrols=85, nScz=86.

Table S11
Predicting iron content by diagnosis, controlling for alcohol use.
Robust regression estimates predicting FPC iron levels (ziron, control-derived z-scores) by diagnosis controlling for alcohol use (categorical predictor with categories No, Yes and Unknown).
Alcohol users were defined as defined as those with a history of drinking an average of ≥20 g ethanol/day. Regression coefficients, robust standard errors, t-statistics along with their corresponding significance and 95%CI are presented, based on a linear regression model using iteratively reweighted least squares MM-estimators, derived using Stata's "robreg mm" command (95% efficiency, default parameters) 24 . ncontrols=85, nScz=86.

Table S12
Predicting disease status based on adjusted PFC iron, controlling for BMI.
Robust logistic regression analysis predicting disease status based on PFC iron (adj_iron, µmol/g), after iron has been adjusted for age, sex, ethnicity, sample pH and log-post-mortem interval, among 56 individuals from the New South Wales Brain Tissue Resource Centre for whom BMI data were available (Table 1). Odds ratios (for having been diagnosed with schizophrenia) per 1 µmol/g increase in adjusted iron, standard errors, z-statistics along with their corresponding significance and 95%CIs are presented. ncontrols=28, nScz=28.

a b
Table S13 Predicting iron content by antipsychotic treatment, visualized in Figure S7.

Table S16
Predicting iron content by results of a postmortem toxicology essay for antipsychotics.
Robust regression estimates predicting PFC iron content (ziron, control-derived z-scores) based on a postmortem toxicology essay for antipsychotics (toxap). Regression coefficients, robust standard errors, t-statistics along with their corresponding significance and 95%CI are presented, based on a linear regression model using iteratively reweighted least squares MM-estimators, derived using Stata's "robreg mm" command (95% efficiency, default parameters) 24 . nScz=63.

Table S17
Robust normality statistics for copper distribution in the Combined Cohort, grouped by diagnosis.
Generalized Jarque-Bera tests for normality, as suggested by Brys et al. 21  Table S18 The effect of diagnosis on PFC copper levels in the Combined Cohort, related to Figure S7. Table S19 Robust measures of copper distribution in the Combined Cohort.

Table S20
Predicting iron content by copper, controlling for diagnosis and copper × diagnosis interaction term, related to Figure S7b.
Regression coefficients, robust standard errors, t-statistics along with their corresponding significance and 95%CI are presented, , based on a linear regression model using iteratively reweighted least squares MM-estimators, derived using Stata's "robreg mm" command (95% efficiency, default parameters) 24 . The effect of schizophrenia on iron at mean copper is highlighted.
(b) Based on the regression model above, average marginal effects of copper on iron are depicted according to diagnosis. ncontrols=66, nScz=67.

Table S22
The effect of diagnosis on PFC zinc levels in the Combined Cohort, related to Figure S8.

Table S23
Robust measures of zinc distribution in the Combined Cohort.
Robust tests of the hypothesis that the variance of PFC zinc (zzinc) is the same across groups, including Levene's statistic (W0) and two statistics proposed by Brown and Forsythe that replace the mean in Levene's formula with alternative location estimators. The first alternative (W50) replaces the mean with the median. The second alternative replaces the mean with the 10 percent trimmed mean (W10). No difference between groups was evident. ncontrols=85, nScz=86.

Table S24
Predicting zinc content by iron, controlling for diagnosis, related to Figure S8b.

Table S25
Exploring the effect of iron on diagnosis through zinc, related to Figure S9.

Table S32
Logistic regression analyses predicting disease status based on PFC iron for sequential age cutoffs, visualized in Figure 2b.
Each row represents one of a series of logistic regression analyses predicting disease status based on PFC iron, after iron had been adjusted for relevant covariates (sex, ethnicity, sample pH and post-mortem interval). Beginning at age 31, for each analysis, age cutoff (years), number of 44 individuals (n) who had died younger than the designated age cutoff and were thus entered into the analysis, odds ratio (OR) of having been diagnosed with of schizophrenia (as compared to being a control) attributed to a 1 SD increase in covariate-adjusted iron, significance of the beta-coefficient for iron as a predictor in the model (p-beta) and its negative log10-transformed value (-log10p-beta) are presented. As the odds ratio was maximal when the analysis included only individuals who had died younger than 35, this age (bolded) was selected as a cutoff for generating subcohorts to be contrasted in subsequent panels. Nwhole sample=170 (see also Figure S11).

Table S34
Robust between-group comparison of PFC iron levels in the young and old subcohorts, visualized in Figure 2d.  24 .

Table S38
Logistic regression analyses predicting disease status based on PFC ferritin for sequential iron cutoffs, visualized in Figure 2b.
Each row represents one of a series of logistic regression analyses predicting disease status based on PFC ferritin, after ferritin had been adjusted for relevant covariates (age, sex, ethnicity, sample pH and post-mortem interval). Beginning at -1 SDs, for each analysis, iron cutoff (z-score based on distribution of control individuals), number of individuals (n) who had iron levels below the designated iron cutoff and were thus entered into the analysis, odds ratio (OR) of having been diagnosed with of schizophrenia (as compared to being a control) attributed to a 1 SD increase in covariate-adjusted ferritin, significance of the beta-coefficient for ferritin as a predictor in the model (p-beta) and its negative log10-transformed value (-log10p-beta) are presented. As the odds ratio had a (regional) minimum when the analysis included only individuals who had iron levels below the control group's mean, the value 0 (bolded) was selected as a cutoff for generating subcohorts to be contrasted in subsequent panels. Nwhole sample=169 (see also Figure S10).

Table S40
Robust between-group comparison of PFC ferritin in low-iron and high-iron subcohorts, visualized in Figure 3d.

Table S42
Logistic regression model predicting disease based on a combination of three predictors following definition of optimal cutpoints, visualized in Figure 4.
A logistic regression model that predicts disease based on a combination of three optimal cutpoints (iron in individuals younger than 35, ferritin in individuals with below-mean iron and iron-toferritin ratio). Based on these cutpoints, for each continuous predictor a categorical predictor denoting the high-risk category (Scz, Figure 4b) and a category for no data have been generated. N=167.

Table S43
Likelihood-ratio tests comparing nested and full regression models, related to Figure 4.
Likelihood-ratio tests assessing whether the full three-predictor model (presented in Table S37) was superior to nested models based on two predictors only (iron-to-ferritin ratio and either [a] ferritin or [b] iron alone). N=167.

Table S44
Deriving an optimal cutoff for a logistic regression model combining all three predictors, visualized in Figure 4.
Optimal cutoff (Youden method) was empirically derived with the classification variable (xbi) representing linear prediction from a logistic regression model predicting disease based on a combination of three predictors (Table S37). N=167.

Table S45
Genome-wide association of iron regulatory proteins and schizophrenia based on GWAS results of CLOZUK.
A list of iron regulatory genes was based on the study by McAllum et al 14 . NCOA4, coding for the cargo receptor mediating ferritinophagy 26 , and TAX1BP1, implicated in NCOA4-mediated ferritinophagy 27,28 , were added to the list. Based on the CLOZUK GWAS 15 , for each iron regulatory gene, the single nucleotide polymorphism (SNP) displaying the highest significance of association with schizophrenia is provided, alongside disease odds ratio (OR), association significance and the RegulomeDB and LINSIGHT scores. Except for IREB2 (presented in Table S46), none of the iron regulatory genes displayed a genome-wide significant association with schizophrenia, defined as a P < 5 × 10 -8 . Data were extracted using SZDB2.0: an updated comprehensive resource for schizophrenia research 13 .
a RegulomeDB is a database that annotates SNPs with known and predicted regulatory elements in the intergenic regions of the H. sapiens genome 29 . Lower scores are associated with stronger supporting data.
b LINSIGHT combines a generalized linear model for prediction of deleterious noncoding variants from functional and population genomic data, with scores ranging 0-1 (higher scores are more likely to have deleterious fitness consequences) 30 .  Table S46 IREB2 -Association with schizophrenia and integrative (Mendelian randomization) analysis.
(a) Based on the CLOZUK and PGC2, the index SNP (SNP with smallest P-value in this loci) alongside P-values of the index SNP in each GWAS are denoted. In both GWASs, genome-wide significant association with schizophrenia, defined as a P < 5 × 10 -8 , was detected.
Significance for the Summary data-based Mendelian (SMR) randomization analysis (PSMR) 16 was mined using SZDB2.0: an updated comprehensive resource for schizophrenia research 13 .   (a) Normal Q-Q plot depicting actual (x-axis) vs. predicted (y-axis) iron values (presented as z-scores derived from the control distribution), for 85 control subjects (blue-border circles) and 86 Scz cases (red-border circles) in the combined cohort. Assumptions of normality were not violated, based on robust tests of normality (Table S4)  (a) Frequency histogram. Relative frequency (y-axis) of iron levels (binned control-derived z-scores, x-axis ) among control subjects (brown boxes) and Scz patients (red boxes) is presented. Patients' distribution is wider. (b) Homoscedasticity plot. Scatter plot depicting absolute residuals (±SD) derived from an unpaired t-test of iron by diagnosis.
Significance of F test to compare variances (F=2.29, DFn=85, Dfd=84) is denoted (see also Table S6).   Figure S5 Balancing plots for PFC iron across diagnosis based on regression-adjusted propensity-score matching, including mode of death as a covariate, related to Figure 1.
Following propensity-score kernel matching with age, pH and PMI as covariates, and including also mode of death as a covariate, based on a logit model where diagnosis of schizophrenia was defined as Treatment, balancing plots depicting raw (left) and PS-matched (right) data are presented, based on KMATCH, Stata's module for multivariate-distance and propensityscore matching. Note marked separation between raw and adjusted data (compared to Figure S4), largely reflecting gross inequality in mode of death across groups.   (a) Comparison of copper levels across diagnostic groups. Scatter plot with bars depicting marginal means (±95%CI) of PFC copper (presented as z-scores) among control individuals and schizophrenia cases. Value above the bar represents significance of diagnosis effect. Outliers downweighed in the analysis appear as transparent data points. Based on a robust (IRWLS MM-estimators regression model, Table S19). (b) Effects of copper on iron in controls and cases. Predicted marginal means (±95%CI) of iron (y-axis, control derived zscores) according to copper (x-axis, control derived z-scores) plotted for each diagnostic group. Based on a robust (IRWLS MM-estimators) regression model with copper, diagnosis and their interaction term as predictors (Table S20). **P<0.01 for between-group difference in iron at control-derived mean copper levels. n controls =66, n scz =67. Figure S8 Prefrontal zinc across diagnostic groups and its dependence on iron.
(a) Comparison of zinc levels across diagnostic groups. Scatter plot with bars depicting marginal means (±95%CI) of PFC zinc (presented as z-scores) among control individuals and schizophrenia cases. Value above the bar represents significance of diagnosis effect. Outliers downweighed in the analysis appear as transparent data points. Based on a robust (IRWLS MM-estimators regression model, Table S22). (b) Effects of iron on zinc in controls and cases. Predicted marginal means (±95%CI) of zinc (y-axis, control derived z-scores) according to iron (x-axis, control derived z-scores) plotted for each diagnostic group. Based on a robust (IRWLS MMestimators) regression model with iron, diagnosis and their interaction term as predictors (Table S23). Figure S9 Exploring the effect of iron on diagnosis through zinc.
(a) Following a logistic regression predicting diagnosis by iron and zinc (N=171) while controlling for sex, ethnicity, sample pH and post-mortem interval, each observation's influence on the regression coefficient (Pregibon's dbeta) is depicted (yaxis) against the predicted probability for diagnosis of schizophrenia (x-axis). As visualized, the influence exerted by sample #41 was unusually high so this sample was excluded. Thus, the structural equation model presented in panel b was based on a total of 170 samples. (b) Exploring the effect of iron on diagnosis through zinc. With iron and zinc as continuous variables (z-scores adjusted for the covariates sex, age, race, pH and log 2 PMI) and diagnosis as a binomial variable, the direct effect of iron on diagnosis and the effects of iron on zinc and zinc on diagnosis were quantified using a general structural equation model. For each effect, coefficient is provided alongside standard error (SE), z-score and significance in brackets. Significance of indirect effect (iron on diagnosis via zinc) was estimated using bootstrap-generated confidence intervals. The highly significant direct effect of iron on diagnosis was over five times larger than the non-significant indirect effect of iron on diagnosis, indicating that from a statistical viewpoint, the risk for a schizophrenia diagnosis conferred by elevated iron was unlikely to be mediated via zinc.  Figure. z-scores were derived using controls' distribution. N controls =85, N schizophrenia =86. Related to Figure 1b. (c) Robust comparison of iron-to-ferritin ratio across diagnostic groups. Individual data points depicting prefrontal iron-toferritin ratio (y-axis, z-scores) against diagnosis (x-axis) are presented. Symbol size corresponds to analytic weight derived using a robust (IRWLS MM-estimators) regression model. Outliers markedly downweighed in the analysis (i.e. weight<0.3) are overlayed with a gray cross, and while included in the robust regression model, these points were excluded from the graphical display in the Main Figure. z-scores were derived using controls' distribution. N controls =85, N schizophrenia =86. Related to Figure 1c.  (a) Following a logistic regression predicting diagnosis by iron-to-ferritin ratio (N=171) while controlling for age, sex, ethnicity, sample pH and log-post-mortem interval, each observation's influence on the regression coefficient (Pregibon's dbeta) is depicted (y-axis) against the predicted probability for diagnosis of schizophrenia (x-axis). As visualized, the influence exerted by samples #8, #71, #146 and #160 was unusually high so these samples were excluded. Thus, the analyses presented in Figure 4 and in the subsequent panel was based on a total of 167 samples. (b) Odds ratio of having schizophrenia per unit iron/ferritin increase, according to age of death. Based on serial logistic regression analyses, the odds ratio of having a diagnosis of schizophrenia (as compared to being a control) attributed to a 1 SD increase in covariate (sex, ethnicity, sample pH and post-mortem interval)-adjusted iron-to-ferritin ratio (y-axis) is plotted against age cutoff (x-axis, years). Using ten-year steps beginning at age 25, each analysis included only individuals younger than the designated age cutoff. For each analysis, symbol size represents number of individuals included (n, upper-center legend), and symbol colour represents significance of iron in predicting diagnosis (-log10-p-value, righthand legend). While the odds ratio was maximal when the analysis included only individuals younger than 35, at this age cutoff the model did not provide superior discrimination compared to a model based on iron alone. a Genome Position, based on hg19 genome assembly. b logFC, log2(Fold Change). Fold change is the ratio of the expression value of a gene in schizophrenia cases to the expression of healthy controls. If logFC > 0, this gene is highly expressed in schizophrenia. Otherwise, this gene is highly expressed in controls.
c T Statistic, the T Statistic is used in a T test when deciding whether we should support or reject the null hypothesis. a Genome Position, based on hg19 genome assembly. b logFC, log2(Fold Change). Fold change is the ratio of the expression value of a gene in schizophrenia cases to the expression of healthy controls. If logFC > 0, this gene is highly expressed in schizophrenia. Otherwise, this gene is highly expressed in controls.
c T Statistic, the T Statistic is used in a T test when deciding whether we should support or reject the null hypothesis. a Genome Position, based on hg19 genome assembly. b logFC, log2(Fold Change). Fold change is the ratio of the expression value of a gene in schizophrenia cases to the expression of healthy controls. If logFC > 0, this gene is highly expressed in schizophrenia. Otherwise, this gene is highly expressed in controls.
c T Statistic, the T Statistic is used in a T test when deciding whether we should support or reject the null hypothesis. a Genome Position, based on hg19 genome assembly. b logFC, log2(Fold Change). Fold change is the ratio of the expression value of a gene in schizophrenia cases to the expression of healthy controls. If logFC > 0, this gene is highly expressed in schizophrenia. Otherwise, this gene is highly expressed in controls.
c T Statistic, the T Statistic is used in a T test when deciding whether we should support or reject the null hypothesis. d FDR, the false discovery rate (FDR) is a method of conceptualizing the rate of type I errors in null hypothesis testing when conducting multiple comparisons.