Major neurocognitive psychosis: a novel schizophrenia endophenotype class that is based on machine learning and resembles Kraepelin’s and Bleuler’s conceptions

Abstract

The purpose of this study is to describe how to use the precision nomothetic psychiatry approach to (a) delineate the associations between schizophrenia symptom domains, including negative symptoms, psychosis, hostility, excitation, mannerism, formal thought disorders, psychomotor retardation (PHEMFP), and cognitive dysfunctions and neuroimmunotoxic and neuro-oxidative pathways and (b) create a new endophenotype class based on these features. We show that all symptom domains (negative and PHEMFP) may be used to derive a single latent trait called overall severity of schizophrenia (OSOS). In addition, neurocognitive test results may be used to extract a general cognitive decline (G-CoDe) index, based on executive function, attention, semantic and episodic memory, and delayed recall scores. According to partial least squares analysis, the impacts of adverse outcome pathways (AOPs) on OSOS are partially mediated by increasing G-CoDe severity. The AOPs include neurotoxic cytokines and chemokines, oxidative damage to proteins and lipids, IgA responses to neurotoxic tryptophan catabolites, breakdown of the vascular and paracellular pathways with translocation of Gram-negative bacteria, and insufficient protection through lowered antioxidant levels and impairments in the innate immune system. Unsupervised machine learning identified a new schizophrenia endophenotype class, named major neurocognitive psychosis (MNP), which is characterised by increased negative symptoms and PHEMFP, G-CoDe and the above-mentioned AOPs. Based on these pathways and phenome features, MNP is a distinct endophenotype class which is qualitatively different from simple psychosis (SP). It is impossible to draw any valid conclusions from research on schizophrenia that ignores the MNP and SP distinctions.

Summations

• The current gold standard approach to schizophrenia, as illustrated by DSM and ICD case criteria, is incorrect and leads to errors in scientific studies, while the label schizophrenia is stigmatising.

• One wonders how many incorrect results have been published in schizophrenia research because the new endophenotype class ‘major neurocognitive psychosis (MNP)’ was not taken into account.

• Clinicians and researchers must always employ the machine learning-based model of MNP rather than the flawed DSM/ICD schizophrenia diagnosis, as the endophenotype class MNP is distinguished by a unique pathophysiology, a severe generalised cognitive deficit, and a worse outcome.

Perspectives

• The endophenotype class MNP model should be cross-validated in other countries and ethnicities.

• Future precision medicine research should incorporate further pan-omics and brainome data into the immune and oxidative neurotoxicity model of MNP.

Introduction

In 2017, an estimated 1.1 million new cases of schizophrenia were reported, with a total of 20 million cases reported worldwide in 2019 (GBD, 2018; Javitt, 2014). Between 0.3% and 0.7% of all individuals are diagnosed with schizophrenia (GBD, 2018; Javitt, 2014; Jablensky, 2000). About half of them will improve significantly over time, with no more relapses, and a small percentage will recover entirely, while the other half will be disabled for the rest of their lives (Jablensky, 2000; Vita & Barlati, 2018; Maes et al., 2021). People with schizophrenia frequently have long-term unemployment, poverty, and homelessness, and they have a higher suicide risk and more physical health issues than the general population, resulting in a 20-year drop in life expectancy on average (Charlson et al., 2018; Hor & Taylor, 2010).

Dr Emil Kraepelin published the first official description of schizophrenia as a mental disorder in 1887. He coined the label ‘dementia praecox’ or ‘early dementia’ to characterise the symptoms that are now recognised as schizophrenia. The cornerstone of Kraepelin’s dementia praecox was a general ‘weakening’ of mental processes that resulted in a ‘defect’ coexisting with ‘productive’ or ‘florid’ symptoms (Shepherd, 1995; Engstrom et al., 2006; Ebert & Bar, 2010; EBO, 2022; Berrios et al., 2003; Decker, 2007). In 1911, Eugen Bleuler coined the term ’schizophrenia’ a combination of the Greek terms schizo (split) and phrene (mind) to describe the mental disorder and fragmented thinking that affect people with schizophrenia. Bleuler considered schizophrenia to be a psycho-organic disorder comprising two symptom clusters, namely a core with four primary symptoms, namely disordered associations, autistic behaviour and thinking, aberrant emotion, and ambivalence, and accessory symptoms including hallucinations, delusions, social disengagement, and reduced desire (Berrios, 2011; Moskowitz & Heim, 2011; Jablensky, 2010).

Kurt Schneider made the next significant advance in 1959 when he enumerated his ‘first rank’ characteristics of the disorder, including ‘auditory hallucinations; thought withdrawal, insertion, and interruption; thought broadcasting; somatic hallucinations; delusional perception; feelings or actions as made or influenced by external agents’ (Jablensky, 2010; Soares-Weiser et al., 2015).

The Russian school lead by A. V. Snezhnevsky was essential in the conception of negative symptoms in schizophrenia. These authors described the entire spectrum of deficiency symptoms, ranging from barely noticeable to profound deficits in mental function, including mental fatigue and diminished energy potential, a decline in intelligence and social disengagement, social withdrawal, frailty, excessive vulnerability, asthenisation of mental activity, mental marasmus, dulled affect, autism, personality regression, and dementia (Snezhnevsky, 1971; Snezhnevsky & Vartanyan, 1970). Importantly, the Russian school believed that this loss of mental capacity was caused by damage to the central nervous system (Mosolov & Yaltonskaya, 2022).

Since the 1970s, the labels ‘defect’ and ‘productive symptoms’ have mostly been replaced by the terms ‘negative’ and ‘positive’ symptoms (Jablensky, 2010). Based on this new knowledge, new diagnostic criteria were constructed, which are still used today by the International Classification of Diseases (ICD) and the Diagnostic and Statistical Manual of Mental Disorders (DSM).

Crow described a subclassification of schizophrenia depending on whether positive or negative symptoms predominate. ‘Type I’ (positive) schizophrenia as defined by hallucinations, delusions, and formal thought disorder (FTD), with an underlying dopaminergic dysfunction, whereas ‘Type II’ (negative) schizophrenia was defined by social withdrawal, loss of volition, affective flattening, and poverty of speech, all of which were presumed to be associated with structural brain abnormalities (Crow, 1980). While positive symptoms may be related to dopaminergic aberrations in the mesolimbic circuits, negative symptoms are thought to be associated with dopaminergic aberrations in the mesocortical circuits.

Kirkpatrick et al. (2000, 2001) suggested the definition of a subtype of schizophrenia defined by persistent ‘primary’ negative symptoms that cannot be attributed to other psychiatric disorders. This clinical construct which was based on Kraepelin’s dementia praecox was dubbed ‘deficit schizophrenia’ and was speculated to represent a unique ‘disease’ within the schizophrenia spectrum (Kirkpatrick et al., 2000, 2001; Kaiser et al., 2011). When this negative symptom cluster is present throughout severe psychotic exacerbations and the more stable inter-episode stages of illness, it is referred to as deficit schizophrenia.

The NINH and NHS continue to categorise schizophrenia symptoms into three domains: positive and negative symptoms, as well as cognitive abnormalities (NIHM, 2019; NHS, 2019). Cognitive deficiencies associated with schizophrenia include impairments in working memory, executive functioning, fluency, list learning, attention and processing speed and those cognitive deficits contribute to lowered health-related quality of life (HR-QoL) and impairments in social functioning, including the ability to work, find work, live independently, and function normally (Maes & Kanchanatawan, 2021; Ueoka et al., 2011; Alptekin et al., 2005; Tolman & Kurtz, 2012; Mohamed et al., 2008; Keefe & Harvey, 2012).

Nonetheless, we discovered recently that: (a) negative and positive symptoms cannot be regarded as independent dimensions and that a latent vector underpins both positive and negative symptoms, dubbed the overall severity of schizophrenia (OSOS) latent vector; (b) a common core of cognitive deficits underpins dysfunctions in episodic, semantic, and working memory, executive functions, and attention, dubbed the generalised cognitive decline (G-CoDe); (c) the negative, positive and cognitive symptoms are strongly associated with neuroimmunotoxic pathways including increased levels of neurotoxic cytokines, chemokines, and oxidative and nitrosative stress compounds; and (d) increased severity of OSOS, G-CoDe and neuroimmunotoxic pathways shape a new endophenotype class, namely major neurocognitive psychosis (MNP), largely overlapping with deficit schizophrenia (Maes et al., 2020c, 2021; Kanchanatawan et al., 2018a, b; Al-Hakeim et al., 2020a, b). Such data show that the common approach of examining the correlations between positive and negative symptoms, cognitive dysfunctions and biomarkers is inadequate and that these associations should be examined using the new precision nomothetic approach combining factor analysis and multiple regression analysis in partial least squares (PLS) analysis (Maes et al., 2020c; Stoyanov & Maes, 2021).

Aims and methods

The present study aims to explain how to use the precision nomothetic psychiatry approach to analyse the interconnections between positive and negative symptoms, cognitive dysfunctions and biomarkers in schizophrenia. Towards this end, we review our data obtained in different study samples of patients with schizophrenia and deficit schizophrenia and using examples extracted from these studies we show how PLS path analysis should be used to examine these complex associations. We will review how to use machine learning to examine whether a common core underpins positive and negative symptoms and cognitive dysfunctions, to examine how biomarkers and cognitive deficits predict the symptomatome of schizophrenia, and how these data can be employed to discover new endophenotype classes.

PLS modelling

PLS modelling allows for (a) the development of causal models that link causal factors including genetic and environmental factors (dubbed the ‘causome’) to adverse outcome pathways (AOPs), cognitive dysfunctions (dubbed the ‘cognitome’) and schizophrenia symptoms (dubbed the ’symptomatome’); (b) the development and inclusion in the model of latent vectors based on unobservable variables, such as positive and negative symptom constructs; and (c) the delineation of mediation effects, such as the effects of AOPs on the symptomatome which are mediated by the cognitome. As input variables, different single indicators (age, gender, and genomic data) and latent vectors derived from a set of highly connected indicators (e.g. a set of interconnected biomarkers or cognitive test results) may be used to predict the final outcome variable, namely the symptomatome and phenomenome (Maes et al., 2020c; Stoyanov & Maes, 2021).

The most prominent feature of PLS (Ringle et al., 2014; Hair et al., 2022) is that linear composites or factors of observed variables are constructed which act as proxies for latent variables that are difficult to measure directly, such as the ’symptomatome’. A latent construct could be thought of as something derived from empirical data that allow for empirical testing of the underlying constructs (Ringle et al., 2014; Hair et al., 2022; Rigdon, 2012). However, the model associations among the constructed latent vectors and single indicators can only be meaningfully evaluated if the PLS model’s construct validity is demonstrated. Therefore, the concept validity of the PLS model should be examined, including criterion validity, convergent validity and discriminant validity (Ringle et al., 2014; Hair et al., 2022; Henseler et al., 2015). Only when the model construct validity check meets predefined quality standards is a complete PLS analysis with 5000 or more bootstrap samples performed. The most critical conditions are that the model’s overall quality is acceptable, as indicated by a Standardized Root Mean Squared Residual (SRMR) < 0.08, that the latent vectors demonstrate adequate reliability validity, as indicated by appropriate composite reliability > 0.8, rho A > 0.8, Cronbach alpha > 0.7, adequate convergence as indicated by an average variance extracted > 0.5 and high factor loadings, namely > 0.6 at p < 0.001, and adequate prediction performance. The latter may be evaluated using PLS_Predict and a tenfold cross-validation technique (Stoyanov & Maes, 2021; Ringle et al., 2014; Hair et al., 2022). Moreover, confirmatory tetrad analysis (CTA) should be conducted to confirm that the reflective models are not misspecified. Compositional invariance can be investigated using predicted-oriented segmentation analysis and measurement invariance assessment.

Furthermore, the produced latent vectors and indicators must have appropriate discriminant validity, which refers to how independent various latent components are different from one another. Discriminant validity assures that a construct is empirically unique and represents a reality that is not captured by other measures in the PLS model (Hair et al., 2022). Checking for discriminant validity is a way to assure that the model’s latent vectors measure what they are supposed to measure (Henseler et al., 2015). Failure to disclose discriminant validity issues might lead to skewed structural parameter estimations and incorrect inferences regarding the associations between constructs. If the constructs do not have enough discriminant validity, they can be combined into a more generic construct that replaces the problematic constructs in the model, and researchers should re-evaluate the newly formed construct’s discriminant validity with all opposing constructs.

The Fornell-Larcker criterion, cross-loadings and the Heterotrait-Monotrait (HTMT) ratio are three ways of determining discriminant validity (Henseler et al., 2015). The Fornell-Larcker criterion is used to assess the degree of shared variance between the model’s latent variables by comparing a construct’s AVE to its shared variance with other constructs, whereby its square root AVE (SRAVE) should exceed the correlations between the constructs. When each measured item has a poor association with every other construct except those to which it is conceptually related, discriminant validity is demonstrated (Gefen & Straub, 2005). A large correlation between items belonging to the same construct and a low correlation between items belonging to distinct constructs is necessary to demonstrate discriminant validity at the item level. Calculating cross-loadings, commonly known as ‘item-level discriminant validity’, is the second approach for evaluating discriminant validity. Each indicator loading should be bigger than all its cross-loadings, or the construct in question will be unable to distinguish between the construct it was designed to assess and another (Chin, 2010). Double-loaders (variables that load heavily on two different factors) are a sign of poor discrimination validity. The most accurate method for measuring discriminant validity is perhaps the HTMT ratio (Henseler et al., 2015) whereby an HTMT ratio >0.85 or >0.9 suggests a lack of discriminant ability (Kline, 2011; Teo, 2011).

Once the construct validity has been established, a full PLS path analysis can be performed using 5000 or more bootstrap samples, allowing for the computation of path coefficients with exact p-values, as well as specific indirect (or mediated), total indirect (mediated) and total (direct and indirect) effects, all of which are used to determine the significance of the (mediated) paths. As a result, PLS path modelling is an effective approach that allows the construction of novel models of a complex illness such as schizophrenia described by causal links between the causome, AOPs, cognitome, symptomatome and phenomenome (Maes et al., 2020c; Stoyanov & Maes, 2021). The PLS analyses we conducted on different Thai and Iraqi study samples showed that the classical concepts of negative symptoms and schizophrenia as a unitary disorder need revision.

Results

The labels ‘negative and positive symptoms’ are confusing

The current standard view of schizophrenia is that it is divided into positive and negative symptoms (Jablensky, 2010; Crow, 1980; NIHM, 2019; NHS, 2019). Negative symptoms like emotional flatness, avolition, alogia, and anhedonia, as well as positive symptoms like delusions, hallucinations, excitement, and disordered thinking (Andreasen, 1989), separate schizophrenia from other mental illnesses. The patient’s loss of emotions (anhedonia), cognitive processes (logic thinking) and behaviours (social isolation) are classified as negative symptoms (Jablensky, 2010; Andreasen, 1989; Early Psychosis Intervention, 2022), whereas positive symptoms are assumed to be new and maladaptive mental processes and behaviours that did not exist before the onset of schizophrenia and have developed as symptoms of the illness (Maes & Anderson, 2021; Jablensky, 2010; NIHM, 2019; NHS, 2019; Early Psychosis Intervention, 2022). Positive symptoms are changes in thoughts and feelings that are ‘added on’ to a person’s experiences, whereas negative symptoms are ‘losses’ that are ‘taken away’ or diminished (e.g. lower motivation or reduced intensity of emotion) (e.g. paranoia or hearing voices) (Early Psychosis Intervention, 2022).

Initially, Maes et al. (Kanchanatawan et al., 2018a, b, c; Noto et al., 2016) distinguished between positive and negative symptoms using standard rating scales such as the Positive and Negative Syndrome Scale (PANSS) (Kay et al., 1987) and Scale for the Assessment of Negative Symptoms (SANS) (Andreasen, 1989) as this is the gold standard method when analysing schizophrenia data (Maes & Anderson, 2021). Nonetheless, Maes laboratories (Kanchanatawan et al., 2018c) thought it would be fascinating to divide positive symptoms into their subdomains, such as psychosis (hallucinations and delusions), hostility, excitement and mannerism, just as they thought it would be fascinating to investigate the subdomains of negative symptoms, such as anhedonia, avolition and flattening. The central idea was that different symptoms may have distinct associations with biomarkers and molecular pathways. Consequently, we created composite scores expressing psychosis, hostility, excitation and mannerism (PHEM) based on all relevant z-transformed item scores of the PANSS and the Brief Psychiatric Rating Scale (BPRS) (Overall & Gorham, 1962), whilst the negative symptom categories of the SANS were deemed appropriate (Kanchanatawan et al., 2018c). In addition, formal thought disorders (FTDs) and psychomotor retardation (PMR) were identified as two additional symptom domains that significantly contribute to schizophrenia symptomatology (Sirivichayakul et al., 2019; Maes et al., 2020a). FTD is marked by abnormalities in abstract and concrete cognition, including disorganised, illogical, and inadequate mental processes, as well as intrusions, fluid thought, and weakened connections (Sirivichayakul et al., 2019). PMR is characterised by deficits in gross and fine motor function, sluggish motor responses, and slow movements (Maes et al., 2019b, 2020a). The computed PMR score was based on the summed z-transformed PMR item scores (Maes et al., 2019b, 2020a) from the PANSS, BPRS and the Hamilton Depression Rating Scale (HDRS) (Hamilton, 1960). Importantly, Cambridge Neuropsychological Test Automated Battery (CANTAB) assessments of the Motor Screening Task (MOT) (CANTAB, 2018), a psychomotor function index, externally confirmed this PMR score (Maes et al., 2020a).

Nonetheless, a primary concern was that certain symptom domains, such as FTD and PMR, could not be properly classified as positive or negative. It is unclear, for instance, whether PMR is ‘added on’ to a person’s normal motor behaviour (a new motor dimension) or should be considered a ‘loss’ (loss of movements). Similarly, the status of FTD is ambiguous: is it a novel thought process (the onset of formal cognitive disorders) or a loss of functions (loss of thought processes)? However, the same concerns can be asked regarding negative symptoms: are anhedonia and avolition, for example, really losses, or are they new ’sensations’ added to human sentience experiences (‘loss of feelings’ is also a feeling)? And what about delusions? Do they add to the way people normally think or do they show a loss of normal human thought processes?

Schizophrenia and range restriction

A second issue is that researchers, including Maes’ laboratories (Noto et al., 2016), computed the associations between biomarkers and rating scale scores of positive and negative symptoms in schizophrenia study samples. Nevertheless, this study group is a restricted study sample of the population. Figure 1 (hypothetical data) shows the regression of severity of illness (the severity of the symptomatome or OSOS) on biomarker pathways (the APOs). This regression, for example, shows that around 63.6% of the variance in OSOS is explained by the severity of the AOPs when the regression is computed in controls (HC in this figure) and schizophrenia (SCZ in the figure) patients combined (F = 44.79, df = 2/51. p < 0.001). However, if this regression is calculated in the restricted study sample of patients only, the association has less impact accounting for only 29.9% of the variance (F = 7.69, df = 2/36, p = 0.002). Thus, when we restrict the range of OSOS (e.g. only patients are included), the correlation coefficient is reduced. Furthermore, if we narrow the range of OSOS even further (e.g. only patients with very high OSOS scores, SCZ2 in Fig. 1), the correlation may no longer be significant (F = 0.23, df = 2/18, p = 0.796). This is a normal phenomenon resulting from restricting the range of the data (Bland & Altman, 2011; Karsner, 2022).

Fig. 1. Regression analysis in (range restricted) study samples including healthy controls (HC) and different schizophrenia (SCZ) endophenotype classes (SCZ1 and SCZ2).

By analogy, it is common practice to compute differences in the biomarkers and severity of the symptomatome between controls and patients and to compute associations between biomarkers and severity in the restricted group of patients (Maes & Anderson, 2021). In fact, this is both illogical and incorrect. Thus, the regression of severity of illness on biomarkers is computed in a restricted sample, but the regression of the biomarkers on the classification (principle of ANOVA for regression) is carried out without any range restriction. In the latter case, biomarkers are considered to be the dependent variable, whereas, in reality, biomarkers are the explanatory variables that predict the severity of illness (dependent variable) which, in turn, is used to categorise the subjects into diagnostic classes (Maes & Anderson, 2021). Thus, the diagnosis of schizophrenia is a post hoc man-made higher-order construct, which is the outcome variable but is frequently used incorrectly as an explanatory variable in analysis of variance or GLM analysis. It becomes even surreal when this outcome variable is employed as an explanatory variable whilst it shows very low reliability and validity, as reviewed previously (Maes & Anderson, 2021). As such, the majority, if not all, researchers use invalid higher-order constructs in incorrect statistical models (explanatory and dependent variables are switched) and use this category to range restrict when investigating covariations between features. Restricting sample variance artificially weakens existing correlations and generalisability and, consequently, correlation coefficients derived from an unrestricted sample should be corrected for range restriction using specific formulas to re-compute the actual correlation coefficient in an unrestricted sample (Karsner, 2022; Sackett & Yang, 2000). Examining these associations in schizophrenia models, we have repeatedly shown that PLS-POS, a method to disclose heterogeneity in PLS path models, did never retrieve the diagnosis as a source of heterogeneity or segmentation.

PHEMFP (PHEM, PMR, FTD) and negative symptoms do not show discriminant validity

Figure 2 illustrates an example of a PLS model in which the effects of biomarkers on negative symptoms and PHEMFP symptoms are investigated. Given that we discovered strong associations among the six PHEMFP symptom domains and the negative symptom domains of the SANS, we first investigated whether it was possible to derive reliable latent vectors from the PHEMFP and negative symptom domains. This model is based on the data presented in Maes et al. (2019a) and is, therefore, based on real data collected from schizophrenia patients and controls.

Fig. 2. Results of Partial Least Squares (PLS) analysis. The neuroimmunotoxic adverse outcome pathways predict two different latent vectors extracted from either negative symptom domains (analogia, anhedonia, avolition, attention, flattening and PANNS negative scale score) or PHEMFP (psychosis, hostility, excitation, mannerism, formal thought disorders (FTDs) and psychomotor retardation (PMR)). Shown are the loadings (with p values) of all indicators of the latent vectors and the path coefficients (with p values). Figures in the blue circles indicate explained variance. PANSneg: negative domain score of the Positive and Negative Syndrome Scale (PANNS).

Firstly, the data demonstrate that we were able to extract a reliable vector with adequate construct validity that underpins the negative symptom domains because (a) the composite reliability (=0.963), rho A (=0.959) and Cronbach’s alpha (=0.954) values are more than adequate, demonstrating accurate internal consistency or composite reliability; (b) the AVE value was 0.814 indicating sufficient convergent validity; (c) all loadings on the latent vector are significant and >0.849; and (d) CTA confirms that the vector is not misspecified as a reflective model. As a result, the six negative subdomains are manifestations of this latent vector known as the ‘negative symptom domain’. In addition, the first latent vector extracted from the PHEMFP symptoms demonstrates (a) adequate composite reliability (=0.945), rho A (=0.946) and Cronbach’s alpha (=0.929); (b) adequate convergence with AVE of 0.741 and significant loadings all greater than 0.767; and (c) CTA results confirming that this factor is not misspecified as a reflective model. Consequently, the PHEMFP symptom domains should also be viewed as manifestations of this latent single trait. In addition, PLS blindfolding revealed that the negative (0.253) and PHEMFP (0.187) latent vectors exhibit appropriate construct cross-validated redundancies, indicating that both constructs have predictive significance. Lastly, the model quality data are more than adequate, with SRMR = 0.041, whilst PLS predict demonstrates that the Q2 predict scores for all indicators are positive, indicating that they outperform the most naive benchmark and, therefore, that the model has significant predictive performance.

Based on this high-quality model data, we performed a complete PLS analysis using 5000 bootstrap samples to investigate the pathways from biomarkers to both positive and negative symptoms. To display a more parsimonious model, we entered the biomarkers as a single indicator, namely a z unit-based composite score computed as the sum of the z transformations of various cytokines + z chemokines + z oxidative stress biomarkers + z lipopolysaccharides + z leaky gut indicators + z BBB permeability indicators (Maes et al., 2020b, c, 2021). The final PLS model reveals that the combined effects of neuroimmunotoxic pathways account for 32.2% of the variance in the domain of negative symptoms and 26.7% of the variance in the domain of PHEMFP.

Nonetheless, this model has significant flaws because both latent vectors lack discriminatory power. First, the Fornel-Larcker criterion (see Table 1) demonstrated that the SRAVE of the PHEMFP construct was less than its correlation with the negative symptom subdomain. Second, the cross-loadings (Table 2) demonstrate that all negative and PHEMFP domain scores have a high cross-loadings on both factors, and thus that all subdomains are double-loaders. All indicator loadings are indeed greater than their cross-loadings (except PMR), but all PHEMFP symptoms loaded highly on the negative latent vector (e.g. excitement loaded at 0.948 on the PHEM vector and 0.856 on the negative domain vector) and all negative subdomains loaded highly on the PHEMFP subdomain (e.g. avolition showed a loading of 0.886 on the negative vector and of 0.799 on the PHEMFP vector). Thirdly, the HTMT ratio between PHEMFP and the negative domain was 0.906, indicating a lack of discriminatory ability. We have rerun these tests of discriminant validity with the same six negative subdomains and the four PHEM subdomains (psychosis, hostility, excitation and mannerism). Similar results to those described in the preceding section indicate that these constructs lack sufficient discriminant validity with, for example, an HTMT ratio of 0.862. In such situations, it is recommended to reassign problematic indicators or combine problematic latent vectors. Since most of the indicators appear problematic, we looked into whether or not both vectors could be combined into a single overarching latent structure.

Table 1. Discriminant validity (Fornel-Larcker criterion) among features of schizophrenia, namely negative and PHEMFP (psychosis, hostility, excitation, mannerism, formal thought disorders and psychomotor retardation) symptom domains and neuroimmunotoxicity (NIT)

Table 2. Discriminant validity (cross-loadings) among negative and PHEMFP (psychosis, hostility, excitation, mannerism, formal thought disorders and psychomotor retardation) symptom domains

A common latent trait underpins all symptom domains

Figure 3 shows a new PLS model with the six negative and six PHEMFP subdomain scores combined. Surprisingly, one vector could be extracted from the 12 subdomains, and this factor showed adequate construct validity with accurate composite reliability (=0.969), rho A (=0.969), Cronbach’s alpha (=0.964) and AVE (=0.723) values, and significant loadings (>0.707) that were all significant (p < 0.001). Furthermore, CTA confirmed that the vector is not misspecified as a reflective model, whilst blindfolding showed an appropriate cross-validated redundancy (0.408) indicating that this latent vector has good predictive relevance. As such, we concluded that negative and PHEM symptoms lack sufficient discriminatory power and probably do not constitute distinct constructs. All negative and PHEM(PF) subdomains are manifestations of the same common latent trait.

Fig. 3. Results of Partial Least Squares (PLS) analysis. The neuroimmunotoxic adverse outcome pathways and a latent vector extracted from cognitive test scores (dubbed the cognitome) predict a latent vector extracted from negative (analogia, anhedonia, avolition, attention, flattening and PANNS negative scale score) and PHEMFP (psychosis, hostility, excitation, mannerism, formal thought disorders (FTDs) and psychomotor retardation (PMR)) symptom domains. The cognitive test scores are as follows: Mini Mental State Examination (MMSE), executive functions tests, verbal fluency test (VFT), Word List Memory (WLM) and True Recall. Shown are the loadings (with p values) of all indicators on the latent vectors and the path coefficients (with p values). Figures in the blue circles indicate explained variance. P1: direct effects, P1 and P2 mediated (indirect) effects. PANSneg: negative domain score of the Positive and Negative Syndrome Scale (PANNS).

Almulla et al. (2021a) using a study sample consisting of Iraqi schizophrenia patients and controls, examined the factor structure of the six PHEM symptoms, the total SANS score and the negative subscale of the PANNS using exploratory factor analysis (EFA) including dimensionality tests which allow estimating the number of factors to be retained including parallel analysis (optimal implementation), the Schwartz’s Bayesian information criterion, and the Hull test (Almulla et al., 2021a). Furthermore, these authors assessed closeness to unidimensionality utilising unidimensional congruence (UNICO), explained common variance (ECV) and mean of item residual absolute loadings (MIREAL), whereby the data should be treated as essentially unidimensional when UNICO > 0.95, ECV > 0.85 and MIREAL < 0.300. The EFA results indicated that all subdomains loaded highly on the first factor (all > 0.660), while the dimensionality tests showed that only one factor should be retained. The closeness to unidimensionality tests indicated that the PHEMFP and negative symptoms should be treated as essentially unidimensional. Moreover, model fit indices showed excellent model fit, good construct replicability, excellent performance across studies and good quality of the factor score estimates.

All in all, the above results of different studies (Maes et al., 2020b; Almulla et al., 2021a) show that a single latent trait, which is essentially unidimensional, underpins the six PHEMFP and negative symptoms of schizophrenia and, therefore, that the latent variable score of these subdomains may serve as a validated and reproducible index of OSOS (Maes et al., 2019b; Almulla et al., 2021a). Our results indicate that OSOS can be reflectively quantified using PHEMFP and negative subdomain scores, and that this reflective latent construct serves as the common denominator for the manifestations that are largely determined by OSOS. It follows that studies reporting on differential correlations between biomarkers and positive and negative symptoms are not very relevant, especially not when computed in the restricted study group of schizophrenia patients.

A bifactoral model

In our studies described above, we also performed PLS or EFA using the same variables in the restricted study group of schizophrenia patients (Maes et al., 2019b; Almulla et al., 2021a). These results showed again that one factor albeit less significant may be extracted from the PHEMPF subdomain and negative PANSS and SANS scores, indicating that even in a restricted study sample, the same latent trait could be established. These findings again show that selecting only schizophrenia to compute correlations between OSOS and biomarkers restricts the range, and thus may artificially weaken existing correlations, thereby hampering generalisability.

Nevertheless, we also analysed our data using bifactorial direct hierarchical EFA, which allows us to define a first generalised factor and additionally one or more single-group factors (SGFs) (Maes et al., 2019b). A bifactor model differs considerably from classical factor analysis with rotated solutions, because a bifactor model method defines a general factor (GF) and subordinate or SGFs. The former is a broader factor which is the source of common variance running through all items in the EFA and the SGF reflects the coherency among a subgroup of items not accounted for by the GF. Maes et al. (2019b) detected that the PHEMFP and total SANS and PANSS negative symptom scores were best modelled by employing a bidimensional oblique solution consisting of (a) a GF reflecting OSOS; and (b) a SGF reflecting negative symptoms and PMR combined. Importantly, the same bifactorial exploratory solution with a GF reflecting OSOS and a SGF reflecting a subordinated factor of negative symptoms and PMR was detected in the restricted sample of schizophrenia patients. Moreover, the general and SGF scores were differentially associated with biomarkers pathways indicating that both constructs are mediated by different albeit partially overlapping pathways. This shows that it is also important to compute the associations between biomarkers, OSOS and putative SGFs in the restricted study group of schizophrenia patients.

Concluding remarks on the symptomatome

All in all, computing the associations between dependent symptomatome (e.g. OSOS and the SGF) and independent (causome, environmentome, AOPs, cognitome and connectome) variables in the unrestricted study group of controls and schizophrenia patients combined allows defining the covariation between those components all over the spectrum from controls to the schizophrenia spectrum including from mild to the most severe phenotypes. This approach should be complemented by the computation of the regression between the features in the restricted sample of schizophrenia patients to delineate schizophrenia-specific covariations among those variables.

Our findings that a unidimensional or bifactorial model is the best way to explain the symptomatome of schizophrenia contradicts the conventional gold standard, which holds that a two-dimensional construct consisting of positive and negative symptoms is the gold standard. Figure 4 shows the covariation between OSOS and AOPs all over the spectrum from controls to schizophrenia patients and the more severe phenotypes as well as the association between the SGF and AOPs which together shape a distinct more severe phenotype characterised by increased OSOS and SGF scores and more severe AOPs (see below).

Fig. 4. Position of patients belonging to qualitatively distinct schizophrenia endophenotype classes in a spectrum of increasing adverse outcome pathways and overall severity of schizophrenia from healthy controls (HCs) to simple neurocognitive psychosis (SNP) and the more severe endophenotype class MNP. SGF: single-group factor consisting of negative symptom and psychomotor retardation. Adapted from Maes et al. (2019b).

The G-CoDe

We and others have previously discussed how deficits in executive functions, learning and working memory, as well as semantic and episodic memory, can result in the formation and recall of false memories and thus may partially explain FTDs, disorganised thought processes, paranoia, and other schizophrenia symptoms (Maes & Kanchanatawan, 2021; Keefe & Harvey, 2012; Orellana & Slachevsky, 2013; Corlett et al., 2007). Cognitive impairments that reflect central neurocircuitry dysfunctions frequently precede the onset of acute psychotic episodes, suggesting that cognitive abnormalities contribute to the schizophrenia symptomatology (Maes & Kanchanatawan, 2021; Keefe & Harvey, 2012). For ultra-high-risk individuals, verbal memory deficits and attentional deficits, for example, are predictive of psychotic symptoms (Brewer et al., 2005; Hawkins et al., 2004).

Given that cognitive impairments may precede and explain at least a part of the schizophrenia symptomatology, we have incorporated cognitive impairments into our PLS models as shown in Fig. 3 and allowed cognitive impairments to predict the symptomatome and AOPs to predict cognitive deficits and the symptomatome. As a result, a mediation model is developed in which cognitive deficits fully or partially mediate the effects of AOPs on the symptomatome. In our initial attempts to investigate such mediated effects, we separately analysed cognitive disorders as measured by different probes of the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) (Fillenbaum et al., 2008) and the CANTAB (2018).

Nevertheless, when examining the highly significant intercorrelations and cross-loadings between the cognitive test scores of CERAD and CANTAB, we combined different executive CANTAB test scores (spatial working memory, one-touch stockings of Cambridge and intra-extra dimensional set-shift) with the CERAD tests’ scores (verbal fluency, word list memory, recall and the Mini-Mental State Examination (MMSE) scores) in factor analyses. Figure 3 demonstrates that we were able to extract one latent vector from those multiple cognitive test scores that showed adequate psychometric properties, namely: (a) adequate internal consistency with composite reliability of 0.905, rho A of 0.875 and Cronbach’s alpha of 0.869; (b) adequate convergent validity with AVE = 0.657 and all loadings on the latent vector are >0.741 at p < 0.001; (c) adequate construct cross-validated redundancy of 0.207; and (d) confirmation that the vector is not misspecified as a reflective model. Impairments in semantic and episodic memory, recall, MMSE, and executive functions are thus all manifestations of a single trait that is the cause of these cognitive dysfunctions. As a result, we coined this latent cognitive vector ‘generalized decline’ (G-CoDe), because the severity scores indicate gradual and more generalised deterioration of cognitive functions (Maes & Kanchanatawan, 2021). Importantly, those neurocognitive deficits in schizophrenia reflect dysfunctions in brain circuits, including ‘the prefronto-parietal, prefronto-temporal, prefronto-striato-thalamic, and dorsolateral prefrontal cortex, amygdala, and hippocampus’ (Maes & Kanchanatawan, 2021; Orellana & Slachevsky, 2013).

Figure 3 depicts the final PLS model, which has more than adequate model quality, with an SRMR of 0.043. We found that 57.9% of the variance in the symptomatome was explained by the regression on the G-CoDe and the AOPs, while 32.6% of the variance in the cognitome was explained by the regression on the AOPs. This PLS graph depicts the path coefficients (with exact p-value) of the causal relationships between the various indicators. This figure also displays the direct effects of AOPs on the symptomatome (P1), the indirect (or mediated) effects of AOPs on the symptomatome through the cognitome (P2. P3) and the total effect, namely P1 + (P2. P3). Significant specific indirect effects (t = 5.70, p < 0.001) of the AOPs on the symptomatome were mediated by the cognitome, which serves as a successful partial mediator (partly because there are also direct effects of the AOPs on the symptomatome). The total effects of the AOPs on OSOS are highly significant (t = 8.55, p < 0.001). As a result, a simple mediator model is constructed, indicating that the AOPs have both direct and indirect (or mediating) effects on the symptomatome. This is an example of complementary mediation in which the direct and indirect effects are complementary because they are both significant and point in the same direction. In addition, this is a distal-mediated model because the path from the cognitome to the symptomatome is greater than the path from the AOPs to the cognitome.

Importantly, such results demonstrate that neuroimmunotoxic AOPs cause aberrations in prefronto-parietal, prefronto-temporal, prefronto-striato-thalamic, and dorsolateral prefrontal cortical, amygdaloid, and hippocampal circuits, and that these circuits partially determine OSOS. In addition, the direct effects of AOPs on the symptomatome demonstrate that other pathways unrelated to the cognitome are also involved in the AOPs’ effects on OSOS. It should be emphasised that the Fornell-Larcker criterion revealed that the cognitome and OSOS have sufficient discriminant validity, although the HTMT was 0.850, indicating that discriminant validity may not be achieved. Nonetheless, the cross-loadings showed that each cognitome indicator loaded highly on OSOS and that each of the 12 OSOS indicators loaded highly on the cognitome factor. Therefore, we deemed it essential to determine whether a single latent trait could be extracted from the 5 cognitome and 12 OSOS domains.

Figure 5 demonstrates that one latent trait could be extracted from these 17 indicators with sufficient construct validity (all loadings > 0.610 at p < 0.001, AVE = 0.632, composite reliability = 0.966, rho_A = 0.967 and Cronbach’s alpha = 0.962, construct cross-validated redundancy = 0.215). In addition, CTA demonstrated that this construct was not incorrectly specified as a reflective model. This demonstrates that a single common core underlies the cognitive deficits, PHEMFP and negative symptoms of schizophrenia, indicating that objective neurocognitive deficits and the symptomatome are strongly intertwined indicators of the core of schizophrenia. Since the neurocognitive impairments measured in schizophrenia reflect aberrations in cortico-amygdala-hippocampal circuits, it follows that also OSOS is largely mediated by dysfunctions in these circuits. Moreover, PLS path analysis revealed that the regression on the AOPs explained 35.8% of the variance in this common core. Inferentially, the neuroimmunotoxic AOPs contribute to aberrations in the cortico-thalamo-hippocampal circuits leading to the cognitome and OSOS, whereby the pathway from AOPs to cognitome and OSOS is a key component of schizophrenia.

Fig. 5. Results of Partial Least Squares (PLS) analysis. The neuroimmunotoxic adverse outcome pathways predict a latent vector extracted from cognitive test scores (Mini Mental State Examination (MMSE), executive functions tests, verbal fluency test (VFT), Word List Memory (WLM) and True Recall), negative symptom domains (analogia, anhedonia, avolition, attention, flattening and PANNS negative scale score) and PHEMFP symptom domains (psychosis, hostility, excitation, mannerism, formal thought disorders (FTDs) and psychomotor retardation (PMR)). Shown are the loadings (with p values) of all indicators of the latent vector and the path coefficient (with p values). Figures in the blue circles indicate explained variance. PANSneg: negative domain score of the Positive and Negative Syndrome Scale (PANNS).

The new endophenotype class MNP

Construction of MNP, a new endophenotype class

The results of the aforementioned PLS analyses are crucial for understanding the phenome of the newly discovered endophenotype classes of schizophrenia, namely ‘Major Neurocognitive Psychosis’ and ’Simple Psychosis’ (SP) (Kanchanatawan et al., 2018a; Almulla et al., 2021a; Al-Hakeim et al., 2020b). Using unsupervised learning (cluster analysis) on Thai and Iraqi study groups with clinical, cognitive and AOP data as input variables, a two-cluster solution was discovered with the separation of stable-phase schizophrenia into two clinically and biologically meaningful clusters (Kanchanatawan et al., 2018a).

Using conventional statistical tests, we determined that MNP is distinguished from SP and healthy controls not only by increased negative and PHEMFP subdomain scores but also by increased affective symptoms, more cognitive deficits and aberrations in a multitude of AOPs (Kanchanatawan et al., 2018c; Almulla et al., 2021b; Al-Hakeim et al., 2020a). In addition, MNP is accompanied by highly specific abnormalities in protective AOPs (dubbed the ‘protectome’) including natural IgM-mediated autoimmune responses, antioxidant gene variants and diminished antioxidant defences, which are absent in SP and thus are distinguishing features of MNP (Maes et al., 2019a). Therefore, MNP is a clinically and biologically distinct group that can be differentiated from SP (see below). Furthermore, using a specific supervised learning technique that allows for the examination of whether a class is qualitatively distinct, namely soft independent modelling of class analogy (SIMCA), we were able to confirm that MNP is qualitatively distinct from SP based on negative, PHEMFP and neurocognitive scores and AOPs as well (Kanchanatawan et al., 2018a; Almulla et al., 2021a; Al-Hakeim et al., 2020b), indicating that MNP is validated as a distinct phenotype in the schizophrenia spectrum. Figure 4 depicts the position of MNP versus SP patients in the schizophrenia spectrum.

Intriguingly, the machine learning-derived cluster MNP resembles Kraepelin’s dementia praecox and Snezhnevsky’s deficit model because this phenotype is characterised by a ‘defect’ or deficit (which we have quantified by the G-CoDe and negative symptom scores) coexisting with ‘productive’ or ‘florid’ symptoms (which we have quantified by PHEMFP scores). The MNP phenotype is also reminiscent of the case definition of Eugen Bleuler and Snezhnevsky who both conceptualised ’schizophrenia’ as a psycho-organic disorder (which we have delineated as a neurological disease caused by neuroimmunotoxicity) characterised by primary symptoms (which we have quantified by computing the negative symptom and G-CoDe scores) and accessory symptoms (which we have quantified as psychotic domain scores), such as hallucinations and delusions. Even so, the classification of MNP and SP based on machine learning does not match Crow’s theory. This is because our machine learning shows that positive (Type I) or negative (‘Type II’) symptoms belong to a common core called OSOS and cannot be regarded as separate entities.

MNP and deficit schizophrenia

MNP versus deficit schizophrenia

Deficit schizophrenia was conceptualised by Kirkpatrick et al. (2000, 2001) as a subtype of schizophrenia defined by persistent ‘primary’ negative symptoms. The diagnostic criteria required the presence of schizophrenia and at least two negative symptoms from a list of six (restricted affect, diminished emotional range, sense of purpose and social drive, poverty of speech, and repression of interests) during clinically stable periods within the previous year. In addition, the negative symptoms must be idiopathic and unrelated to other conditions such as anxiety, depression, drug effects, suspicion, mental retardation or psychotic symptoms.

The distinct category of MNP has a very ambiguous relationship with ‘deficit schizophrenia’. On the one hand, we derived the model of MNP using the diagnostic criteria of deficit schizophrenia, with which it exhibits a substantial overlap (Kanchanatawan et al., 2018a, b), but on the other hand, the machine learning model of MNP differs significantly from the clinical concept of deficit schizophrenia. While the diagnostic criteria of deficit schizophrenia are solely based on negative symptoms, MNP is characterised by more pronounced PHEMFP symptoms, OSOS and SGF, the G-CoDe, and more severe neuroimmunotoxic and neuro-oxidative pathways as compared with SP (see below, section ‘A full description of all MNP features’) (Kanchanatawan et al., 2018a; Maes et al., 2019b; Almulla et al., 2021a; Al-Hakeim et al., 2020a). Moreover, there are significant issues with the diagnosis of deficit schizophrenia as reviewed in the next section.

Major concerns with Kirkpatrick’s deficit schizophrenia

First, previous reports have concluded that it may be more accurate to view negative symptoms and thus deficit schizophrenia as a dimension rather than a separate disease (Roy & DeVriendt, 1994). Additionally, negative and positive dimensions may coexist in the same individual, and there is also disagreement regarding whether negative symptoms intensify from a healthy state to schizophrenia with a ‘fully formed illness’ (dimensional theory) or whether type II or deficit schizophrenia is a distinct nosological category (categorical theory) (Kanchanatawan et al., 2018b; Takahashi, 2013).

Second, as described in the previous sections, negative and PHEMFP symptoms not only coexist and covary during the stable phase of schizophrenia but also share a common core that underlies those subdomains. Furthermore, in the stable phase of schizophrenia, negative and PHEMFP symptoms correlate strongly with anxiety and depression to the extent that they all appear to belong to the same latent construct (Almulla et al., 2021b). All of these symptom domains are, therefore, strongly interconnected manifestations of the illness’s symptomatome core, indicating that negative symptoms are not the result of psychosis, depression or anxiety, but rather are manifestations of the single-trait OSOS. Moreover, as indicated by our PLS models, neuroimmunotoxic AOPs explain a large part of the variance in the cognitome and OSOS, as well as the strong covariation between these strongly related components, despite their distinct presentation. Inference shows that negative symptoms should be thought of as a sign of OSOS that is partially determined by AOPs.

Third, Kirkpatrick’s exclusion criteria that negative symptoms may not be caused by depression, anxiety or psychotic symptoms are nonetheless remarkable. How could one decide whether the negative symptoms which occur in schizophrenia are idiopathic and not the consequence (or cause for that matter) of psychosis, depression or anxiety, which are other features of schizophrenia? Excluding features of an illness to define that interrelated features of the same illness are idiopathic is at least a subjective method. Such subjective criteria have no place in the precision psychiatry approach.

Fourth, both ’schizophrenia’ and ‘deficit’ are utterly stigmatising labels and the label ‘deficit schizophrenia’ does not even describe the disease’s AOP, neurocognitive and symptomatome features (Kanchanatawan et al., 2018a). Additionally, our machine learning-derived model is conceptually quite different from the clinical case definition of deficit schizophrenia and, therefore, we labelled the new endophenotype class MNP (Kanchanatawan et al., 2018a). In fact, the ‘neurocognitive psychosis’ label describes the results of our machine learning results that both negative and PHEMFP symptoms are psychotic symptoms that to a large extent are explained by neuroimmune aberrations and cognitive impairments, while the label ‘major’ is used to denote that MNP is the full-fledged, qualitatively distinct phenotype as opposed to SP, which is a less severe phenotype.

A full description of all MNP features

Biomarkers of MNP

Table 3 shows the biomarker, clinical, cognitive and phenomenome features of MNP versus SP and controls as determined in our studies. First, MNP is characterised by significantly activated neuroimmunotoxic pathways including neurotoxic cytokines particularly tumour necrosis factor (TNF)-α, interleukin (IL)-1β and IL-6/IL-23/Thelper (Th)-17 signalling, and neurotoxic chemokines (particularly eotaxin), increased oxidative damage to proteins and lipids, damage to paracellular pathways including tight and adherens junctions with increased translocation of Gram-negative bacteria and, therefore, increased lipopolysaccharides (LPS) load, and breakdown of the blood–brain barrier (Maes & Anderson, 2021; Kanchanatawan et al., 2018a, b, c; Al-Hakeim et al., 2020a, b; Sirivichayakul et al., 2019; Maes et al., 2019a, 2020a, b, c, 2021). In addition, MNP is characterised by diminished antioxidant defences, particularly diminished sufhydryl (-SH) groups and paraoxonase1 (PON1) activity, which results in a highly substantial rise in the oxidative stress toxicity/antioxidant ratio (Maes et al., 2020b). Significantly, MNP is associated with changed frequencies of the Q192R PON 1 genotope, with a higher frequency of the QQ genotype and Q allele in MNP compared to SP and controls (Maes et al., 2020b). PLS analysis demonstrates that the QQ genotype increases MNP risk via mediated effects on PON1 activity. In addition, MNP is distinguished by diminished IgM responses to a number of oxidation-specific epitopes, such as malondialdehyde, azelaic acid, and TRYCAT-adducts (Maes et al., 2019a). These results demonstrate that MNP, but not SP, is associated with a malfunction in IgM production by B1 cells, which is an essential component of the innate immune system that protects against microbial infections, oxidative stress, and inflammation and strives to preserve homoeostasis (Maes et al., 2019a). Computing the ratio of the multiple immune and oxidative toxicities (MITOTOX, namely neuroimmune + neuro-oxidative + bacterial translocation + damage to the paracellular pathway + increased neurotoxic levels) on the protective resilience biomarkers (PRORES, namely antioxidant levels and IgM-mediated natural immunity) reveals that an increased MITOTOX/PRORES ratio is a defining feature of MNP (Maes et al., 2019a, 2020b).

Table 3. Key features of major neurocognitive psychosis (MNP) versus simple psychosis (SP) and healthy controls

MITOTOX/PRORES: ratio of multiple immune and oxidative toxicities (MITOTOX) on protective resilience against neuroimmune, neuro-oxidative and bacterial stress (PRORES).

MITOTOX is computed as a z unit-based composite score based on key oxidative stress biomarkers + key cytokines + IgA to tryptophan catabolites + lipopolysaccharides. PRORES is computed as a z unit-based composite score based on paraoxonase 1 activity + sulfhydryl groups + total radical antioxidant trapping parameter + IgM directed to oxidative-specific epitopes.

Neurocognitive features of MNP

Second, MNP is characterised by multiple dysfunctions in neurocognitive functions including verbal fluency, word list memory, recognition and delayed recall, delayed memory savings, visual sustained attention, attentional set-shifting, spatial planning, strategy use, rule acquisition, emotional recognition, and a more generalised impairment as measured with the MMSE (Maes & Anderson, 2021; Maes & Kanchanatawan, 2021; Kanchanatawan et al., 2018a, b; Maes et al., 2020a, b, 2021; Sirivichayakul et al., 2019; Almulla et al., 2021b). This suggests that MNP is associated with a generalised deficit in the major cognitive domains of perception, recognition, attention, learning, semantic and episodic memory, decision-making, thought consciousness, and sensory input. In contrast to SP, MNP is accompanied by neurocognitive deficits that are more severe than those observed in mild cognitive impairment but significantly less severe than in Alzheimer’s disease (Kanchanatawan et al., 2018d). Consequently, the findings of our studies support earlier theories that schizophrenia is characterised by a generalised ‘defect’ or ’scar’ that previously was mistakenly referred to as ‘dementia’ (Jablensky, 2000; EBO, 2022; Decker, 2007; Berrios, 2011; Moskowitz & Heim, 2011). Moreover, there is now evidence that the neuroanatomical basis of such neurocognitive impairments may be ascribed to dysfunctions in neuronal circuits including ‘prefronto-striato-thalamic, prefronto-temporal, prefronto-parietal, amygdala and hippocampal circuits’ (Maes & Kanchanatawan, 2021; Orellana & Slachevsky, 2013).

Moreover, as explained above a common factor, the G-CoDe, may be extracted from these multiple cognitive functions and a large part in its variance is explained by the oxidative toxicity/antioxidant and MITOTOX/PRORES ratios. Since the decreased antioxidant and innate immune system protection and increased neuroimmunotoxic and neuro-oxidative toxicity in MNP are major predictors of the G-CoDe (Maes & Kanchanatawan, 2021; Maes et al., 2021), it is obvious that these neurotoxic pathways have affected the function of the neuroanatomical circuits underpinning the G-CoDe.

The symptomatome of MNP

Third, as reviewed above and shown in Table 3, diverse negative symptom domains (including blunting, alogia, avolition and anhedonia), PHEMFP symptoms as well as depressive and anxiety symptoms and the latent vector extracted from these symptoms scores (OSOS) are significantly greater in MNP than in SP and in controls (Maes & Anderson, 2021; Maes et al., 2020c, 2021; Kanchanatawan et al., 2018c; Almulla et al., 2021b; Sirivichayakul et al., 2019). Since increasing G-CODE can lead to false memories and recall, FTD and paranoia and frequently precedes the onset of acute psychotic episodes, it is safe to conclude that the severity of cognitive impairments contributes to the symptomatology on MNP.

Furthermore, we were able to extract one latent vector from the cognitive test results and the symptom domains (Maes et al., 2021; Almulla et al., 2021a), demonstrating that both domains are driven by a common core and, thus, that a shared substrate is the root cause of all cognitive and symptom manifestations established in MNP. Therefore, the same neuronal circuits that underpin the cognitome (namely prefronto-striato-thalamic, prefronto-temporal, prefronto-parietal, amygdala and hippocampal circuits) also underpin at least in part MNP.

Health-Related Quality of Life in MNP

Fourth, when compared to SP patients and controls, MNP patients have significantly lower scores on four different domains of HR-QoL, including physical, psychological, social and environmental scores, as well as on the overall decline in HR-QoL (Maes et al., 2021; Kanchanatawan et al., 2018a; Al-Musawi et al., 2022). Furthermore, MNP patients have more difficulties with daily activities than SP patients and controls, as well as more difficulties with self-care than normal controls (Kanchanatawan et al., 2018a). In various studies, we discovered that neuroimmunotoxic pathways (e.g. neurotoxic TRYCATs and the IL-6/IL-23/Th-17 axis) were strongly associated with lower HR-QoL, and that these effects were mediated via the AOPs, the cognitome and symptomatome (Al-Musawi et al., 2022).

Most importantly, by extracting a latent vector from OSOS (based on all symptom domains), G-CoDe (based on all neurocognitive test scores) and HR-QoL (based on all four domains), we were able to compute a new comprehensive index reflecting ‘behavioral-cognitive-physical-psychosocial’ (BCPS) worsening (Maes et al., 2020c, 2021). This BCPS-worsening index is much higher in MNP than in SP and is, therefore, another hallmark that MNP is a qualitatively distinct class. Interestingly, in first-episode schizophrenia (FES), the BCPS-worsening index is explained to a large extent by increased breakdown of the paracellular and vascular pathways, bacterial load, complement system alterations (namely increased IgA to circulating C1q immune complexes) and lowered PON1 activity (Maes et al., 2021), whereas in multiple-episode schizophrenia (MES), the BCPS-worsening score is largely predicted by increased TNF-α, oxidative toxicty and number of psychotic episodes. According to our results, patients with MES and FES with very high BCPS-worsening indices also experience MNP. This demonstrates that MNP can be present from the first psychotic episode or develop because of the cumulative effects of multiple episodes and that different neurotoxic pathways may contribute to its development.

While there is some evidence that the AOPs, cognitome, symptomatome, HR-QoL and BCPS-worsening features may be considered to be on a continuum from controls to SP to MNP, as shown in Fig. 4, these features may group together to form a new class with increasing severity of the pathways, G-CoDe, OSOS, HR-QoL and the BCPS-worsening index (Kanchanatawan et al., 2018a, b). As a result, quantitative differences may become qualitative, forming the new endophenotype class MNP. Furthermore, as shown in Table 3, some of the pathways and biomarkers, particularly those related to resilience and the protectome, differ qualitatively between MNP and SP/controls, namely (a) decreased IgM responses to oxidative-specific epitopes and TRYCAT adducts are a feature of MNP, whereas there is a trend towards increased IgM levels in SP; (b) PON1 activity is significantly reduced in MNP but not in SP; and (c) the QQ genotype is associated with MNP but not SP. These qualitative differences in the protectome coupled with increasing BCPS-worsening shape MNP as a discrete endophenotype class that is qualitatively different from SP as assessed by machine learning techniques like SIMCA (Kanchanatawan et al., 2018a, b).

As a consequence, failure to use these new endophenotype classes (MNP and SP) may result in false-positive outcomes (a pathway specific to only one phenotype is generalised to schizophrenia) and false-negative outcomes (existing phenotype-specific pathways for MNO or SP are not detected at all) (Maes & Anderson, 2021). One wonders how many erroneous results in schizophrenia research, including genetic research, have been published because the MNP and SP endophenotype classes were not taken into account. Future research should always include MNP and SP as key endophenotype classes, and clinicians should use these labels instead of schizophrenia because they are less stigmatising. Moreover, future research should enlarge our MNP and SP models with additional pan-omics data as well as brainome data including structural and functional (connectome) brain neuroimaging assessments in order to improve the diagnosis of MNP and SP (Stoyanov, 2020; Stoyanov et al., 2021).