The environmental risk factors evaluated in this study comprise urbanicity, migration, perinatal complications, physical maltreatment, sexual abuse, traumatic brain injury, cannabis consumption and alcohol abuse. Contacting male extreme group subjects of GRAS (with low vs. high environmental risk before age 18; discovery sample; N = 134) [18] for a planned epigenetic follow-up, we found 27% of high-risk individuals in forensic units in contrast to only 6% of low-risk subjects (p < 0.001; χ²-test, two-sided). This finding was replicated in the remaining GRAS-I sample (GRAS-I males and females minus extreme group subjects), where a stepwise increase in lifetime prevalence of forensic hospitalization was seen upon risk accumulation (Fig. 1a).

This observation made us wonder whether we would find a strong intercorrelation between the here investigated environmental risks. To test for multicollinearity between the risk factors included in the accumulation models, we calculated the variance inflation factor (VIF) for each sample. Our results suggest that none of the included factors significantly collinears with any other (for each sample VIF ≤ 1.28), allowing us to include them in our models.

We hypothesized that forensic hospitalization reflects violent aggression. To quantify this trait, and in absence of established instruments for comprehensive retrospective analysis of violent aggression, we generated the VASS (Fig. 1b). Information for VASS was extracted for all discovery individuals (N = 134) from detailed history, available in the GRAS database [19, 20], and additional extensive chart study based on original medical documents over lifetime. VASS ranges from no documented aggression to lethal consequences of violent aggression. Relative weight is given to severity of aggression and number of registered re-occurrences. For first cross-validation of this new tool, an intra-sample expert ranking of relative aggression severity in the discovery sample was performed by 2 independent psychologists (unaware of environmental risk status of subjects under study), yielding Spearman’s rho = 0.97 for interrater reliability and rho = 0.96 for intercorrelation with VASS (Fig. 1c). Inspection of VASS values in the discovery sample upon risk accumulation again demonstrates the ‘stair pattern’ (Fig. 1d).

Since not all information was available as detailed for the schizophrenia replication samples of GRAS-I and -II as for the discovery sample, we introduced a dichotomous aggression proxy, including history of forensic hospitalization and/or conviction for battery, sexual assault, manslaughter or murder (at least once in lifetime). Intercorrelation with VASS and expert ranking, respectively, resulted in r_pb = 0.63 (point-biserial) and r_rb = 0.62 (rank-biserial) (Fig. 1c). Applying this proxy to replication samples I-III (GRAS-I males without discovery sample, GRAS-II males, GRAS-I&II females), consistently yielded the ‘stair pattern’ upon risk accumulation, even though at slightly lower level in females (Fig. 1e–g). The percentage of subjects with documented aggression increases with the number of risk factors, strikingly similar in all schizophrenia cohorts (Fig. 1h). Important for future preventive measures in at-risk subjects is the observation that a single risk factor (independent of its kind) is still compensated for (Fig. 1h). When comparing subjects with 0 vs. ≥3 environmental factors over all schizophrenia samples, the odds ratio for violent aggression (based on aggression proxy) amounts to 10.5. Details on sociodemographic and disease-related variables, as well as on the various highly intercorrelating measures of violent aggression in the environmental risk accumulation groups in discovery and replication samples are given in Tables 1 and 2. Whereas no consistent differences in premorbid intelligence, present cognition (cognitive composite), and chlorpromazine equivalents (relative amount of antipsychotics) emerge among groups, age tends to be lower and suicidality to occur more frequently with increasing pre-adult environmental risk exposure in the schizophrenia cohorts, which is not unexpected considering our previous report [18] (Table 1). A remarkable increase in all available measures of violent aggression becomes obvious upon accumulation of environmental risk (final model consisting of urbanicity, migration, physical and sexual abuse, alcohol and cannabis), reflected by highly significant p values in group and trend statistics throughout samples (Table 2).

For analyzing onset of aggressive behavior, the extensive information on aggression available in the discovery sample was exploited. Early aggression (any aggression documented before age 18 years and well before schizophrenia onset) clearly increased upon ≥2 risk factors, whereas aggression seen only later in life seemed independent of early environmental risk (Fig. 1i). Therefore, we hypothesized that violent aggression upon risk accumulation may be unrelated to mental disease.

To test this hypothesis, we had the chance to analyze 2 well-characterized independent samples (replication IV and V; Tables 1 and 2) of young individuals from the Spanish general population. Since data on criminal conduct could not be obtained in these cohorts, we had to use alternative, psychometrically validated instruments as aggression proxies, namely LSRP secondary psychopathy score [26], measuring rule-breaking and lack of effort towards socially rewarded behavior, and the aggression-hostility factor of ZKPQ-50-CC [27]. Urbanicity as risk factor was unavailable in these samples (reducing the model to 5 of the 6 risk factors explored in schizophrenia, that is migration, physical maltreatment, sexual abuse, alcohol and cannabis). We also note that subjects were younger and as academics probably higher educated as compared to the disease cohorts. Despite these mitigating facts, and despite employing individuals of another country, the expected ‘stair pattern’ still emerged clearly for both proxies, likely suggesting generalizability of these findings (Fig. 1j,k; Tables 1 and 2). Data given here for the general population samples (replications IV and V) are based on both males and females. In addition, evaluating men and women separately (taking both general population cohorts together) yielded significant results for both genders (Table 2 bottom).

Addressing the composition of risk factors among groups across cohorts, we obtained a comparable pattern throughout schizophrenia samples (stacked-charts; Fig. 1a,d-g). In the general population subjects, particularly alcohol and cannabis consumption (classified as ‘secondary hits’) predominated (Fig. 1j,k) which also seem to play an appreciable role in schizophrenia cohorts. Therefore, we wondered whether separate analysis of risk accumulation, integrating only primary vs. only secondary hits, would still result in significant effects on aggression. For all schizophrenia samples individually, group difference and trend remained highly significant (not shown). Taking all schizophrenia subjects together (N > 1200), the aggression proxy yields for the accumulation model, built on primary risks only (urbanicity, migration, physical and sexual abuse), p = 4.5 × 10⁻¹⁷ (χ² = 75.28) and p < 2.2 × 10⁻¹⁶ (χ² = 68.28), for group differences and trend, respectively. The corresponding results for secondary risk factors (alcohol, cannabis) in schizophrenia are p = 6.6 × 10⁻¹⁹ (χ² = 83.71) and p < 2.2 × 10⁻¹⁶ (χ² = 83.40). Analogously, taking all general population subjects together (N > 530), we obtain for LSRP with primary risks (urbanicity not available) p = 0.002 (H = 12.65) and p = 0.0003 (J = 33774.5), and with secondary risks p = 1.3 × 10⁻⁴ (H = 17.92) and p = 5.3 × 10⁻⁵ (J = 42412.5) for group differences and trend. Also here, significance was already reached with separate analysis of both cohorts (not shown).

For deciding on the accumulation model, we had initially screened all individual risk factors of our ‘primary plus secondary risk factor model’ separately in both schizophrenia and general population cohorts to get an estimation of their relative impact (Tables 3a,3b,3c). Perinatal complications and neurotrauma before the age of 18 years were unavailable for general population subjects. Since these risks showed the lowest overall impact on aggression proxies in schizophrenia, we decided not to include them in our present accumulation model.

Finally, we performed an epigenome-wide association scan to detect differential methylation of blood-derived DNA of selected extreme group individuals (discovery sample; N = 134; Fig. 1a), originally planned as epigenetic follow-up study [18]. This scan turned out to be negative. In fact, contrasting subjects either with high vs. low number of environmental hits or according to VASS median split yielded a single methylation difference upon lowering the Bonferroni threshold to 10⁻⁶ (Table 4). Similarly, when looking in an exploratory fashion (small/unbalanced group sizes) at individual risk factors separately, results were essentially negative (Table 4). Hits associated with migration were likely related to ethnicity rather than environmental risk, as reported recently [31]. The power of our sample size - even though in the range of suggestions [32] and despite extreme group comparison - may not have been sufficient to detect differences, also due to a vast underlying heterogeneity of individual methylation sites. Even the search for methylation differences of aggression-related candidate genes [33–35] turned out negative (not shown), putting the relative weight of phenotypical consequences (here violent aggression) vs. common methylation results in humans into perspective. In contrast, determining HDAC1 mRNA levels in PBMC available from male extreme group subjects (≤1 vs. ≥3 risks) revealed a highly significant difference (p = 0.001), with higher levels in the high-risk (N = 33) compared to the low-risk group (N = 109) (Fig. 1l). This transcript encodes an enzyme of the histone deacetylase complex which serves as an overarching regulator of epigenetic processes. Indeed, peripheral HDAC1 mRNA levels seem to be a more robust readout of epigenetic modifications in small sample sizes [36] as compared to specific methylation sites in the epigenome-wide association scan, and suggest lasting epigenetic alterations.