Endophenotype 2.0: updated definitions and criteria for endophenotypes of psychiatric disorders, incorporating new technologies and findings

Abstract

Recent genetic studies have linked numerous loci to psychiatric disorders. However, the biological pathways that connect these genetic associations to psychiatric disorders’ specific pathophysiological processes are largely unclear. Endophenotypes, first defined over five decades ago, are heritable traits, independent of disease state that are associated with a disease, encompassing a broad range of neurophysiological, biochemical, endocrinological, neuroanatomical, cognitive, and neuropsychological characteristics. Considering the advancements in genetics and genomics over recent decades, we propose a revised definition of endophenotypes as ‘genetically influenced phenotypes linked to disease or treatment characteristics and their related events.’ We also updated endophenotype criteria to include (1) reliable measurement, (2) association with the disease or its related events, and (3) genetic mediation. ‘Genetic mediation’ is introduced to differentiate between causality and pleiotropic effects and allows non-linear relationships. Furthermore, this updated Endophenotype 2.0 framework expands to encompass genetically regulated responses to disease-related factors, including environmental risks, illness progression, treatment responses, and resilience phenotypes, which may be state-dependent. This broadened definition paves the way for developing new endophenotypes crucial for genetic analyses in psychiatric disorders. Integrating genetics, genomics, and diverse endophenotypes into multi-dimensional mechanistic models is vital for advancing our understanding of psychiatric disorders. Crucially, elucidating the biological underpinnings of endophenotypes will enhance our grasp of psychiatric genetics, thereby improving disease risk prediction and treatment approaches.

Endophenotype: classical definition, criteria, and successes

Gottesman and Shields [1] (1973) introduced endophenotypes as measurable phenotypes between illness and genotypes. Gershon, Rieder, and Goldin [2, 3] (1978, 1986) and Gottesman & Gould [4] (2003) specified the classical criteria of endophenotypes for psychiatric genetics listed in Table 1. Others further contributed to the discussion regarding its applications in psychiatric genetics [3, 5–7]. Neurophysiological, biochemical, endocrinological, neuroanatomical, cognitive, and other neuropsychological measures (including self-report questionnaire data) are commonly used endophenotypes. The commonly used endophenotypes for psychosis include measurements derived from electroencephalography (EEG), cognition, eye movements, and brain imaging. EEG responses to repeated external events (Event Related Responses; ERPs), such as auditory stimuli, have long been known as abnormalities in psychosis disorders and have been proposed as critical biological components of psychosis [8]. Impaired cognition is an important feature of schizophrenia (SCZ) [9, 10]. ERP and cognition abnormalities are present before the onset of psychosis symptoms [11–13].

Table 1.

Endophenotype 2.0 revisions: definition and criteria.

	Endophenotype 1.0	Endophenotype 2.0	Major changes from 1.0 to 2.0
Definition	Genotype: Not defined. Originally referred to concordance in relatives. Endophenotype: A phenotype that is heritable, co-segregates with a psychiatric illness, yet is present even when the disease is not (i.e. state independent), and is found in non-affected family members at a higher rate than in the population. (per Gottesman & Gould 2003 [4], Rieder, Goldin and Gershon [2, 3]) A heritable intermediate phenotype associated with a disease. [Intermediate phenotype is a mechanism-related manifestation of a complex phenotype.] (per Goldman & Ducci 2007 update [6]).	Genotype: state of variations in DNA sequence, including single nucleotide polymorphisms, copy number variants, tandem repeats, de novo mutations, and structural variants. Endophenotype: A genetically influenced phenotype between genotype and disease diagnosis. It is linked to disease or treatment characteristics and related events, which include disease symptoms, risk or resilience factors, and responses to treatment, with evidence from family or population data.	• Expanded the definition to include illness-, risk-, resilience-, and treatment response related features. • Emphasized genetic mediation. • Added molecular and cellular endophenotypes, and many other new phenotypes • Population evidence is emphasized
Criteria	1. The endophenotype is associated with illness in the population. 2. The endophenotype is heritable. 3. The endophenotype is primarily state independent (manifests in an individual whether or not illness is active). 4. Within families, endophenotype and illness co-segregate. 5. The endophenotype found in affected family members is found in non-affected family members at a higher rate than in the general population. (Per Rieder 1978 [3])	1. Reliable measurement in individuals 2. Associated with disease or treatment characteristics, and related events. 3. Genetically mediated	• Removed the requirement of state independence; • Removed the requirement of evidence from family data; • Added illness-related features, which include risk, resilience, and treatment response; • Response to risk environments, such as stress and inflammation, is specifically added; • Genetic mediation can be in many forms, including non-linear relationships; • Causation is emphasized
Examples of major types	Neurophysiological, biochemical, endocrinological, neuroanatomical, cognitive, or neuropsychological (including configured self-report data) measures	• Molecular endophenotypes: genomic and metabolic measurements (including expression, splicing, developmental, and age-related changes in these measurements) and findings derived or imputed from DNA, RNA, protein, or other molecular variations. • Cellular endophenotypes • Treatment- and course-related endophenotypes: response to drugs and response to non-drug treatment. • Response to risk-environment endophenotypes: psychological stress, toxin, insufficient nutrition, sleep condition, and inflammation.

Historically, the meaning of “genotype” has evolved from measuring similarity in relatives to DNA sequence information, including base-pair level variations (SNPs), insertions, deletions, and structural variants. “Phenotype” refers to any assessable characteristic of an individual beyond genotype. Phenotypes are the products of genotypes, environmental factors, and their interactions throughout the lifespan, ranging from molecular and cellular measurements to high-order behavioral traits.

Genetic study of endophenotypes has been pursued actively in the past decades and has generated some successful examples in studies of attention-deficit/hyperactivity disorder (ADHD), substance use disorders (SUD), and major depression (MDD), as reviewed by Sanchez–Roige and Palmer [14]. Behavioral traits of cognition (including intelligence), eye movement control, emotion, mood, executive function, social function, and personality are impaired in multiple psychiatric disorders. Cognitive deficits are present in children who later develop SCZ [11, 12]. Several notable examples of genetic analysis of behavioral endophenotypes across psychiatric disorders have been reported: shared genetics between education attainment and SCZ [15], neuroticism has been positively correlated with psychotic experiences and with the risk of SCZ [16–18], and depression [19]. Intelligence [20] and “Big Five” personality measures (openness [21, 22], conscientiousness [22], extraversion [23], and neuroticism [24]) have GWAS loci overlapping with loci associated with various psychiatric disorders. The N-back working memory test [25] is one example of a standardized behavioral measure. GWAS of N-back test in 2298 participants estimated SNP-based heritability ranging between 31 and 41%, but no genome-wide significant SNP was reported. Another endophenotype commonly linked to mood disorders is facial expression recognition [26]. A GWAS of facial emotion recognition in 4097 subjects failed to detect genome-wide significant SNP association [27]. Seven genome-wide significant genetic loci have been reported for endophenotypes of spatial processing, abstraction and mental flexibility, antisaccade task, sensorimotor dexterity, face memory, and continuous performance test by the Consortium on the Genetics of Schizophrenia (COGS) Study (N = 1533) [28]. COGA published GWAS of EEG on 8810 individuals and reported one locus on 18q23 associated with low theta EEG coherence [29]. In a GWAS of diffusion tensor imaging (DTI) model-derived parameters along 21 cerebral white matter tracts in nearly 44,000 individuals, 151 SNPs associated with white matter microstructure were discovered [30]. The Bipolar-Schizophrenia Network on Intermediate Phenotypes (B-SNIP) study (N = 849) identified three genome-wide significant associations for eye movement measures [31]. A GWAS of Prepulse inhibition (PPI) (N = 1493) identified two significant loci [32]. However, their meta-analysis (N = 2660) failed to capture any genome-wide significant loci for PPI [33].

Glahn et al. [7]. brought quantitative trait loci of gene expression (eQTL) or proteins (pQTL) into the scope of endophenotypes when there were very limited eQTL data in 2014. They discussed concerns about the classical assumptions regarding the genetics of endophenotypes as compared with the genetics of disease and proposed to study endophenotypes that are closer to gene action, like gene expression.

Consistent with Glahn et al., we propose that the genetic influence on endophenotypes will be stronger for traits closer to DNA than those more distal. Molecular QTLs exhibit significantly higher heritability than higher-order phenotypes, such as brain imaging, electrophysiology, or behavioral traits. The genetic contribution from individual SNPs tends to diminish as information progresses through the Central Dogma of molecular biology [34, 35]. Generally, as phenotypes become more distant from DNA sequences, more regulatory factors (including environmental factors) are involved, decreasing SNP heritability. Therefore, most endophenotypes are likely to exhibit genetic complexity. SNP heritability of these endophenotypes falls in the range between the two extremes represented by molecular traits and disorder diagnoses. Consequently, the sample size required to achieve sufficient statistical power may range from dozens to tens of thousands, even millions, depending on the specific characteristics of the endophenotypes, including their distance to DNA and the number of regulators involved. These values should be determined through population-based studies or large cohorts. Molecular endophenotypes such as gene expression and DNA methylation require much fewer samples for genome-wide study than are required for genome-wide associations with genotypes (GWAS) [36] (Table 2). Significant associations have been reported for eQTL with less than 100 samples, while disease GWAS typically requires tens of thousands.

Table 2.

Commonly used and emerging endophenotypes for psychiatric disorders.

Biological levels	Endophenotypes	Major existing data	Sample size range in existing data (N)
Molecular	Brain gene expression	26 studies including PsychENCODE, GTEx, ROSMAP	44–1866
	Brain DNA methylation	11 studies	125–494
	Brain protein abundance	Three studies, including PsychENCODE/BrainGVEX, ROSMAP	133–835
Brain structure and function	Structural and functional MRI (neurological, structure of the nervous system, structure of brain)a	373 studies including ENIGMA, BSNP	327–26,577
	EEGb	5 studies	191–6644
	Eye movementb	3 studies	60–6246
	Cognition (cognitive)a	78 studies	441–386,276
Individual differences in psychological characteristics	Personality (psychiatric, mental function, temperament, and personality)a	99 studies including UK Biobank	5228–1,067,913
	Sleep (psychiatric, mental function, sleep functions)a	55 studies	721–449,732
	Social interaction (social interactions)a	20 studies	86,941–452,302

^aAccording to GWAS Atlas [111] queried on 11/20/2024 (The query terms are quoted).

^bAccording to the GWAS Catalog [110] on 11/20/2024.

Some endophenotype data have been collected along with genotype data in large-scale population-based studies, including the UK Biobank [37, 38], the Adolescent Brain Cognitive Development (ABCD) study [39], and All of Us [40]. These data are primarily questionnaires and non-endophenotype diagnoses, but there is some inclusion of brain imaging, cognition tests, and medical history or records. UK Biobank has phenotypes of health-related outcomes, clinical coding, questionnaires, physical measures, imaging data, Big Five Personality Traits, subjective well-being, anxiety, loneliness, depressive symptoms, intelligence, and education attainment. ABCD collects brain imaging data and neurocognitive measures, including crystallized and fluid reasoning, reading ability, working memory, behavioral inhibition, verbal learning and memory, sequencing skills, processing speed, attention, and visuospatial reasoning. Environmental factors are also emphasized in ABCD, particularly those related to substance use. All of Us collects data from electronic health records and digital health records, substance use, surveys of social behavior, and optional social network data [40].

The UK Biobank has been one of the most productive and used data sources. In addition to brain imaging studies [41–43], GWASs of UK Biobank data have yielded significant genetic associations for traits of cognitive functions and educational attainment, which is pertinent to Schizophrenia and neurodegenerative diseases [44] and many other mental and physical health-related traits [45]. Interestingly, the SNP heritability of intelligence-related traits is particularly strong in socioeconomically deprived people [46]. Other large biobanks are emerging, such as Vanderbilt Biobank (BioVU), the eMERGE (Electronic Medical Records and Genomics) Network, Million Veteran Program (MVP), China Kadoorie Biobank, FinnGen, and Danish National Biobanks. Global Biobank Meta-analysis Initiative (GBMI) is a collaborative network of 23 biobanks representing more than 2.2 million consented individuals with genetic data linked to electronic health records. Although psychiatry is not the major focus of these large biobanks, these biobanks have some endophenotypes, like brain imaging and cognition measures. However, these existing biobanks were designed with the influence of endophenotype 1.0. Endophenotype 2.0 data are mostly not part of these existing biobanks.

Limitations of earlier endophenotype definition and the need for endophenotype 2.0 with new definition and criteria

The classic endophenotype definition and criteria were created before the era of genome-wide association study (GWAS) and genomics. New understandings about disorders, including mental illnesses, their phenotypes, and underlying genetics, have emerged. Here, we term the classic endophenotype definitions Endophenotype 1.0 and call attention to the following four major limitations:

1. The state independent criterion. The original criteria for endophenotypes required state independence, which excludes many important traits, such as early development stage-specific or age-dependent traits that may have permanent effects but are only detectable in a certain period of life. Temporally dynamic phenotypes (such as prenatal neural differentiation, growth, and migration) may prove critical for understanding disease susceptibility. The developmental hypothesis of psychiatric disorders has been proposed since the 1820s [47] and gradually matured [48–50]. Thousands of genes have very different expression profiles in prenatal development stages than in adults [51, 52].

2. The illness-association criterion. The original criterion is to co-segregate with illnesses. But resilience, response to treatment or risk environment, and course of illness are important research areas for psychiatric genetics. Treatment may or may not be directly associated with disease risk. However, even the side effects of treatment are translationally relevant. Genetic studies of the drug responses have developed into pharmacogenetics. Responses to non-medication treatments, even placebo effects, may prove equally interesting. Moreover, the genetics of these factors may or may not overlap with the genetics of illness.

   Human responses to environmental changes may be regulated by genetic factors. Cannabis use, environmental stress during pregnancy (including famine and certain viral infections), childhood, and adulthood, history of obstetric complications, and low serum folate level are recognized environmental risk factors for SCZ [53, 54]. The responses to these risk factors are known to vary among individuals, which implies possible genetic mediation [55].

   Statistical associations of endophenotypes with the risk of illness would include both positive and negative associations. Negative associations with risk would be related to resilience endophenotypes. By resilience, we refer to processes that reduce the risk of illness in persons with other predisposing risk genes or environmental factors. Resilience to schizophrenia has been reported to have associated loci [56], but resilience-related endophenotypes are less studied than risk-related phenotypes. Much of resilience might simply be the alternate alleles to genetic risk factors or lack of exposure, but there might also be specific genetic factors for protective effects, such as genes that can drive behaviors related to active stress-coping [57, 58], cognitive abilities [59], and flexibility [60–63], which can be related to most psychiatric disorders. Genetics of resilience related to stress response have also been explored using GWAS [64, 65].

3. Study in families or populations. The original criterion is about family study for historical reasons. However, many endophenotype studies are more frequently performed in populations than families because of more accessible access to large samples. This is particularly important for the weak genetic effects of endophenotypes. Requiring the phenotypes to co-segregate within families and be more common in well-family members than in the population has been largely replaced by case-control comparisons or analysis of quantitative traits in population samples.

4. Measurable distinctions between pleiotropy and causation were poorly developed when Endophenotype 1.0 was created. An endophenotype may be associated with illness, and genetic variants may be associated with both the endophenotype and the illness, but the variants do not have a causative effect on illness; such associations are called pleiotropy. Genetic mediation or causal relationship must be differentiated from association to make endophenotype useful for understanding disease mechanisms.

Genetic mediation refers to the phenomenon of shared genetic factors connecting an endophenotype to the development or causation of a disorder. It highlights the genetic influences that operate through the endophenotype, contributing to the manifestation of the disorder. Genetic factors of an endophenotype per se may not be directly observed to be associated with illness, as reviewed by Thomas [66]. When an SNP (or gene) is associated with both endophenotype and illness, the presence of causal relationships can be evaluated by statistical analyses such as Mendelian Randomization (MR) [67] or mediation analyses of multi-dimensional data. Moreover, genetic mediation can take the form of non-linear effects, which could fail correlation tests, like in the study of treatment changes [68]. Li et al. examined the non-linear association between sleep duration, cognition, and mental health, and uncovered underlying brain structure and genetic mechanisms missed by linear approaches [69]. A guideline for high-dimensional mediation analysis on high-throughput genomics studies is available [70].

One may ask, what are the consequences of using an outdated definition? Why must we update the definition when most of the proposed Endophenotype 2.0 is expensive to study? Using outdated definitions in research leads to inefficient use of resources, diminished statistical power, poor reproducibility, and, more importantly, missed opportunities to uncover critical links that could reveal true disease mechanisms. By incorporating advancements in omics, cellular biology, neuroscience, and computational modeling, updated definitions like Endophenotype 2.0 can provide deeper mechanistic insights and enhanced predictive accuracy. Although studying these phenotypes can be expensive at present, history shows that bold conceptual shifts often drive the development of cost-effective approaches, ultimately making advanced methods economically viable and more accessible over time.

The updated concept of Endophenotype 2.0 eliminates the restrictions and limitations of the original definition and criteria, allowing novel and broader phenotypes to be included within the unified framework of the endophenotype concept. When a phenotype is classified as an endophenotype, established methods for studying endophenotypes become applicable, fostering new opportunities for interdisciplinary applications. For example, QTL (quantitative trait loci) analysis demonstrates how methods developed for eQTL research have been successfully extended to other endophenotypes, including techniques for genetic mapping, fine-mapping, causal analysis, and multi-dimensional omics integration. The imputation of unmeasured endophenotypes in large datasets leveraging a genetic study of a small, deeply-phenotyped cohort, using methods like PrediXcan [71], UTMOST [72], BrainXcan [73], and others [74], has shown to be highly effective. Additionally, endophenotypes enable the exploration of their interconnected relationships through networked analyses. Comparative and integrative analyses across multi-dimensional endophenotypes are now feasible because they share similarly defined structural frameworks. This adaptability facilitates the introduction of new analytical methods that emphasize causal inference and non-linear genetic mediation. New mechanisms could also be revealed by including previously excluded phenotypes, particularly those associated with state-dependent traits. This adaptability underscores the value of updating the definition.

Endophenotype 2.0 fills in mechanism gaps between genetic variants and psychiatric disorders, capturing missing biology

Psychiatric genetics has achieved significant progress in the past fifteen years. Hundreds of common SNPs and dozens of rare copy number variants (CNVs) have been associated with major psychiatric disorders. However, a widely recognized gap exists between genetic associations of common diseases and biology [75–78]. The biological processes related to genotypic associations have not been unraveled, referred as “missing biology”. Disease-associated genetic variants frequently do not pinpoint specific genes, as most of the associated SNPs are non-coding and regulatory elements can be far away from the gene body. Even for associations mapped to genes through genetic, genomic, and epigenomic studies, their relationships to high-order phenotypes and functions at the levels of cells, circuits, brain, and behavior are mostly unknown. Functional annotation of GWAS signals and assigning genetic associations to specific mechanisms are the most important follow-ups of GWAS.

Linking genetic variants to disorder through endophenotypes contributes to the mechanistic interpretation of disease. Meyer-Lindenberg and Weinberger demonstrated how imaging genetics help relate genes to the mechanism, brain circuitry, emotion, and risk of schizophrenia and depression [79]. However, it should be noted that we expect pleiotropy to be common. Most regulations and actions are in networked systems and, therefore, multi-factorial and complex, involving many endophenotypes in between.

Different endophenotypes may interact to contribute to disease pathology, as illustrated in Fig. 1. At the same time, clinically distinct psychiatric disorders may share underlying mechanisms. The relationships among the endophenotypes will inform the mechanistic networks, pathways, and circuitries underlying diseases. Genomic Structural Equation Modelling (Genomic SEM) is a valuable tool that can be used to model associations among endophenotypes and to identify genetic variants that can affect cross-disorder liability [80].

Fig. 1. Environmental and Genetic Interactions with Endophenotypes 1.0 and 2.0.

The Hypothesized path analysis may resolve the mechanistic causal relationships from genetic variations (G) and environmental factors (E) to major endophenotypes (both 1.0 and 2.0) and, ultimately, disorders. The effect of each edge could be quantified in future studies.

Endophenotype 2.0 and the missing heritability gap

More than a decade ago, a gap in heritability estimates was termed “the missing heritability” in common diseases [78], where SNP-based heritability was found to be considerably smaller than heritability based on concordance of illness between relatives. For SCZ, twins and other concordances among relatives generate a heritability of 0.67 [81], but SNP-based heritability is 0.26 [82]. Now, heritability is empirically defined as the proportion of additive genetic variance in phenotype by case-control difference. The challenge of “missing heritability” can be reframed as a large part of case-control variance not being accounted for by additive genetic variance in the polygenic model. The additive genetic variance calculation is a “black box,” where effect sizes on illness and allele frequencies of all common SNPs enter the box, and the weighted summation of their effects exits the box.

One proposed explanation for missing heritability is that uncommon and rare variants contribute to familial concordance but only common variants are included in SNP-based heritability [83]. Another explanation is that SNP-based heritability only measures an event that occurs at conception (genotype generation), whereas illness concordance between relatives as adults is based on observed case-control status, which reflects the effects of common SNPs plus all the developmental and environmental exposure events that have occurred since conception. These events include gene-gene and gene-environment interactions, and their endophenotype products [84]. The additive effects model does not represent the indirect effects of genotypes interacting with other risk contributors.

Response to exposures, including protective environments and medication, is one major new category for Endophenotype 2.0. This new category allows us to explore mechanisms of a presumably major component of psychiatric genetics, genetic contributors that interact with the environment. Such endophenotypes may help uncover gene-gene interaction and gene-environment interaction. These non-additive interaction effects are not included in disease GWAS [85]. With the help of endophenotypes, some of the missed heritability accounted for by the non-additive effects may be recovered [84].

It is worth emphasizing that genes with strong effects on the endophenotype but weak effects on illness may nonetheless prove clinically useful, such as drug targets that can modify the endophenotype. A successful example is the drugs that modify a synthetic enzyme of serum cholesterol (HMG-CoA reductase, HMGCR) to prevent coronary artery disease (CAD). HMGCR is highly associated with LDL levels as identified by MR but only has a small effect on the disease CAD [86, 87].

Within the framework of endophenotype 2.0, the study of gene-environment interactions (GxE) opens new avenues for research. For example, genetic variants that regulate response to psychological stress may contribute to the risk of disorders. When these variants are mapped through endophenotypes, disease GxE analysis can be focused and powerful as the number of tests may be dramatically reduced. For example, an exposome score (ES), which estimates the sum of environmental risks [88], was tested for interaction with polygenic risk score (PRS) and confirmed GxE effects on psychosis [89]. In a similar vein, stress-related exposures were tested for interaction with PRS of depression and anxiety and confirmed the GxE effects on depression and anxiety [90]. While PRS and ES or equivalent summary scores are powerful measures for demonstrating the GxE effects on disease risk, they are blind to the granular mechanism of individual genes and pathways. There is still a significant opportunity for research into how particular genes interact with specific environmental risk factors.

Endophenotype 2.0 distinguishes causation from association

Association is not causation. Two parallel events stemming from a common cause can show an association without directly influencing each other. This non-causal relationship (horizontal pleiotropy) can complicate the interpretation of data and provide limited insight into the mechanisms underlying a disease. Understanding this distinction is crucial for accurate analysis and interpretation in scientific research. Estimating true causal effects can be achieved through vertical pleiotropy, where an SNP affects one endophenotype and, as a result, another. Conversely, genetic factors may directly influence the illness, bypassing the endophenotype in horizontal pleiotropy. These different genetic pathways can be analyzed using Mendelian Randomization (MR) methods [66, 67], enabling the assessment of pleiotropic and causal relationships.

MR provides tests of causality and can be applied to multi-dimensional data. Functional connections between genotypes and (epi-)genomic profiles have been studied in autopsy and fetal brains. Brain anatomic structures and physiology have also been associated with psychosis disorders. Nonetheless, most of these associations have yet to be tested for causation. An SNP associated with both illness and endophenotype can be used as an “instrument” to determine whether the association with illness is an indirect effect mediated by the endophenotype exposure or a direct effect on illness. The ratio of the regression coefficient for illness to the coefficient for the endophenotype is the MR measure of the SNP’s direct causality [66].

Genetic variants presumably exert their effects on illness through various endophenotypes. Thus, the challenging problem of determining the causality of an SNP on disease could be converted into two sub-problems (1) SNP to endophenotype and (2) endophenotype to disease causal relationships. MR can use inputs from either a single sample (“one-sample MR”) or two samples (“two-sample MR”) depending on the sources of the SNPs, exposure, and outcome data [91]. Two-sample MR is particularly useful because the endophenotype (exposure) data are often collected in a cohort different from the case-control (outcome) cohort. Applying two-sample MR to brain imaging and disease GWAS revealed significant causal relationships of structural features of the superior longitudinal fasciculus with anorexia nervosa (AN) and of the corpus callosum with Alzheimer’s disease [92].

Univariable MR studies the effect of a single exposure on an outcome such as disease diagnosis using methods such as Standard MR (Wald ratio) [93], inverse-variance weighted (IVW) [94], the weighted median [95], mode-based estimate (MBE) [96], contamination mixture [97], and the MR-Egger regression [95, 98]. IVW and MR-Egger use all available SNPs as instruments, assuming they are valid. Mendelian randomization pleiotropy residual sum and outlier (MR-PRESSO) [99], MR-Lasso [100], and MR-robust [100] select only a subset of SNPs that are likely to be valid instruments. The IVW method is sensitive but assumes no horizontal pleiotropy. MR-Egger can detect and adjust for horizontal pleiotropy but is less sensitive than IVW. It should be noted that the quality of MR relies on the instrument strength, measured by F-statistics, which indicates the association between the SNPs and the exposure of interest. F-statistic >10 is a general guideline. Sensitivity analyses should be performed using methods like MR-Egger regression, weighted median, or heterogeneity tests to evaluate the robustness of the results based on different assumptions about pleiotropy.

Multivariable MR (MVMR) [101] allows for the simultaneous analysis of multiple exposures (endophenotypes) to estimate their direct and indirect effects on an outcome. Multivariate generalized linear models [102] and multivariate functional linear models [103] can test for genetic effects on multiple traits. Pleiotropy-robust methods for MVMR [101] and Pleiotropy Estimation and Test Bootstrap (PET-B) [104] are designed to estimate and test the extent of pleiotropy between a variant and multiple phenotypes. All these methods aim to improve the detection and characterization of pleiotropy compared to univariate approaches. MVMR revealed complex pleiotropic causal relationships among intelligence, education, bipolar disorder (BD), and SCZ [105]. MVMR also discovered that the genetic liability of smoking had a deleterious association with longevity, while genetic liabilities for major psychiatric disorders had no independent associations [106].

Each method has its pros and cons. Combining multivariate and univariate GWAS models could be useful for identifying and narrowing down candidate loci with potential pleiotropic effects for downstream experiments [107]. Guidelines for performing MR are available and continuously updated [108].

Emerging endophenotypes of psychiatric disorders

Endophenotypes can be classified by biological levels from molecular to whole organism. Genetics have effects on structure, function, and physiology at each of the biological levels. Behavioral and neurophysiological phenotypes are high-end endophenotypes. Some endophenotypes are well-accepted, such as those in Endophenotype 1.0 for psychosis disorders. Most emerging endophenotypes have only fragmentary evidence suggesting they may meet the three criteria of Endophenotype 2.0. Population GWASs from the UK Biobank and others have uncovered genetic associations with many traits. Some of them are also associated with psychiatric disorders, such as educational attainment [15] and sleep chronotype [109].

Examples of commonly used and emerging endophenotypes are shown in Table 2, along with some existing data collections/studies in the public domain. We compiled data from the GWAS Catalog [110] and GWAS Atlas [111]. Molecular trait QTLs were collected directly from the literature, as they were not in the typical collections of the GWAS databases.

Since more than 80% of the genetic variants associated with psychiatric disorders are non-coding, dysregulated gene expression is currently the major suspect for the underlying molecular changes, and such dysregulation is defined here as a molecular endophenotype. Multiple postmortem brain studies support this expression dysregulation model. Differentially expressed genes and splicing isoforms have been reported in SCZ, BD, and ASD postmortem brains [112, 113]. SNPs that regulate expression (eQTLs), RNA splicing (sQTLs), and DNA methylation (mQTLs) are enriched in major psychiatric disorders [114–124].

Specific subtypes of brain cells have been implicated in psychiatric disorders. Neuropharmacological studies have implicated dopaminergic, glutamatergic, and GABAergic neurons in SCZ [125]. Analyses of coexpression modules in postmortem brains of psychiatric patients suggest that neurons, oligodendrocytes, astrocytes, and microglia are involved in SCZ, BD, and ASD [113]. With the technological advances of single-cell sequencing and algorithms for deconvolution [126] of cell type in genomic analyses, deeper understandings are expected for the roles of each subtype of cells in specific brain regions.

New endophenotypes are expected or under development

1. Molecular phenotype data has accumulated quickly in the past ten years. Molecular phenotypes such as gene expression in eQTLs have been proven effective in capturing genetic regulators by their associated SNPs. Gene expression, splicing, gene networks, protein abundance, and chromatin accessibility have associations with disease and are thus molecular endophenotypes. Quantitative trait loci (QTL) for eQTLs, sQTLs, and mQTLs are commonly used to capture genetic regulators for thousands of genes and to explain GWAS results.

   We expect molecular endophenotypes to snowball in the coming years with the fast-evolving sequencing technologies. These would include more molecular trait (-omics) QTLs mapped to early brain development and cell subtypes. eQTLs from postmortem tissues only explain a small portion of GWAS signals [127]. Dynamic (conditional or context-dependent) eQTLs that may capture regulators of gene expression under specific developmental stages or environments are expected to recover more functional annotation of GWAS. Recently, 672 fetal brain samples have been used to generate early developmental eQTL, sQTL, and isoform QTL data [128]. Single-cell RNA sequencing (scRNA-seq) profiled expression across three differentiation stages on 215 induced pluripotent stem cell (iPSC) lines that were differentiated toward a midbrain neural state and generated cell-type-specific eQTLs for early developmental stages [129].

2. Cellular phenotypes are not yet practical measurements but are conceivable future developments. Cellular function studies in cultured cells have been performed on some candidate genes, including POU3F [130, 131], DGCR5 [132], TRIM8 [131], DISC1 [133, 134], NRXN1 [135], SETD1A [136], TCF4 [137], ZNF804A [138], PCCB [139], GLT8D1, and CSNK2B [140], and others [141]. Expression dysregulation of these genes was found to affect cell differentiation [131, 135, 140], proliferation [131, 137, 140], electrophysiology [131, 140], and expression of other risk genes [132, 136–139]. More associated genes need to be studied for cellular phenotypes in both early developmental and matured brain subtype cells, which will be more challenging. Development and neural circuitry are also important for the etiology of psychiatric disorders. Fetal brain eQTLs showed unique contributions to the risks of SCZ and autistic spectrum disorder (ASD) [128, 142, 143], further emphasizing the importance of developmental biology.

   iPSC-derived neuronal, glial cells and various organoids are used for modeling phenotypes of early-developing brains and have the advantage of preserving each donor’s genetic background, diagnoses, and possibly endophenotypes. Most current cellular models target individual candidate genes with a small number of cell lines, which can only assess the functionality of a few SNPs or genes. Improvements in scalability are on the horizon. A large-scale massively parallel reporter assay (MPRA) in iPSC-derived neurons was reported to study 240 GWAS loci simultaneously [144]. Hundreds of iPSC lines from patients and controls may soon be studied for genetic effects on cell proliferation, differentiation, electrophysiology, pharmacologic response, and calcium imaging to produce corresponding endophenotypes. iPSCs in mental disorders may become an area with heavy research investment in the coming years.

3. Brain imaging and EEG-based measures are the major 1.0 type endophenotypes in psychiatry today, as represented by the ENIGMA [145] and B-SNIP [146]. Various structural, connectivity, and activity metrics of specific brain regions have been implicated in distinct disorders [147–149]. Current methods can only be used for costly deep phenotyping. New portable instruments and mobile devices may increase the throughput and reduce costs in the future. Data quality and standardization across instruments and sites have been challenging, but early tests have delivered promising results [150, 151].

4. Treatment response is a category of Endophenotype 2.0, and individual variation in treatment response is well recognized. Several GWASs of treatment response have been published with dozens of loci that can survive multiple-testing corrections [152–154]. Various scales for symptoms [155–157] are commonly used in clinical trials to assess treatment effects. Although symptomatic improvement is the desired result, physiologic measurements and other changes in endophenotypes are related to the biological processes underlying treatment and have led to more effective predictors [8, 146].

5. Inflammation-related phenotypes are another Endophenotype 2.0 category related to the risk environment. They can reflect individual variation in response to inflammatory environmental insults, which include infections, allergens, vaccines, psychological stress, or wounds. Immune changes have been detected in peripheral blood and postmortem brains of patients with major psychiatric disorders [158–160]. For example, IL6, TNFα, VEGF, and CRP were found to be significantly higher in the plasma of psychosis probands than in controls [161]. Inflammatory factors were associated with blood-brain barrier pathology [162], neuroimaging, cognitive impairment, and symptom measures [161, 163]. Some immune changes are common across disorders, while others are disorder-specific [158, 164]. A critical question is how much individual variation in inflammation responses is regulated by genetics and further contributes to psychiatric risk. Using immunophenotyping flow cytometry in a genetic study of innate and adaptive immune traits with 497 adult female twins, 76% of 23,394 immune phenotypes showed a predominantly heritable influence [165]. Leukocyte subsets, CD39 on regulatory T cells, and CD32 on myeloid dendritic cell (DC) populations are examples of highly heritable traits detected in this study. Some of the inflammation-related factors were stable over time in healthy people [166–168]. 3′ UTR Alternative polyadenylation (APA) QTL (3′aQTL) was mapped in 18 human immune-related blood cell types and eight immune stimulation conditions [169]. Neural cell types have not yet been studied for genetic regulation of immune responses.

6. Psychological stress-related traits are another group of risk environment-related endophenotypes. Stress contributes to the risk of depression, anxiety, panic disorder, and post-traumatic stress disorder (PTSD) [170, 171]. SCZ patients report greater exposure to stressful life events than controls [172, 173]. Recent stressful life events were associated with the onset and progression of psychosis in some patients [174]. Similarly, psychological stress was linked to the onset, progress, and severity of BD, with particularly strong evidence for early-life stress [175]. The stress-regulated hypothalamic-pituitary-adrenal (HPA) axis system is associated with MDD through the classic dexamethasone suppression test [176], although the association is much more complex than initially proposed [177, 178]. A candidate gene study found that variation in the glucocorticoid receptor was associated with cortisol levels [179]. Defining and quantifying stress response for genetic research presents its complexities and challenges.

   Questionnaire-based methods like the Childhood Trauma Questionnaire (CTQ) [180–183] and structured (or semi-structured) interviews [184–188] are commonly used to collect stress-related data. It should be noted that data could be biased by the people who administrate data collection. Objective measures that are free of influence from the data collector are desirable. Cortisol levels in the blood, saliva, or hair, EEG measures, blood pressure, heart rate, and skin conductance would be objective measures of stress response at the time of measurement. Chronic stress is more difficult to measure objectively than acute stress. There is ample literature studying the mental health and even epigenetic changes associated with psychological stressors. One study showed that the heritability of response to repeated stress could be high (h² > 0.97) [189] in a twin study of the Trier Social Stress Test (TSST), which is a controlled acute stress experiment. Hair cortisol concentration was found to be highly heritable (h2 = 0.72) in a twin study [190]. Genetic studies of stress responses were reviewed in 2006 [191]. Progress has been made since. Meta-analysis of genome-wide interaction study (GWIS) of stress-sensitivity by modeling the interaction between SNPs and MDD status on neuroticism score in data of UK Biobank and Generation Scotland: Scottish Family Health Study identified one significant associated gene ZNF366 in gene-based tests [192]. ZNF366 was reported to regulate glucocorticoid receptor function [193]. Moreover, 3,820 glucocorticoid receptor-response-modulating cis-eQTLs were reported in the peripheral blood of 160 male individuals [194]. Out of these SNPs, MPRA validated 547 in glucocorticoid-responsive regulatory elements, with 233 showing allele-dependent activity [195].

7. Responses to environmental changes are essential for human survival and contribute to psychiatric disorders. These responses include responses to in-utero events in SCZ (infection and severe food restriction). The possibility that there are genetic components to these obstetric event responses was raised by Mittal et al. in 2008 [55], but to our knowledge, no molecular investigation has been conducted on this hypothesis. Chronotype, defined as an individual’s natural sleep timing preference, is a potential endophenotype. In addition to clock genes regulating circadian rhythms, GWAS of chronotype identified 351 significant loci. Moreover, schizophrenia, depressive symptoms, MDD, and intelligence were all negatively correlated with the morning chronotype [109]. However, genetic variants influencing sensitivity to varying durations of light exposure have not yet been investigated.

Two complementary approaches to the collection of endophenotype data

Deep and shallow phenotyping are two alternative approaches for collecting endophenotype data. They each have their own pros and cons as shown in Table 3.

Table 3.

Deep and shallow phenotyping.

	Deep	Shallow
Design	Precise and comprehensive collection of data on multiple phenotypes in one cohort.	Minimal phenotyping, large samples.
Methods	Advanced instruments and standardization.	Minimal standardization. May rely on self-reports or reported diagnoses, mobile devices, medical records, etc.
Examples	Brain imaging, eye tracking, EEG, molecular and cellular traits, cognition	Diagnosis, self-reported behavioral and online questionnaire, digital phenotypes
Procedural costs in time and money	Expensive and lengthy phenotyping of each participant.	High throughput, low cost.
Pros	Quality (accuracy and precision) of measures can be better monitored	Can generate large datasets in multiple populations.
Cons	Small sample size: generalizability may be a concern.	Concerns about validity/quality of phenotypes.
Hybrid approaches	Electronic Health Records contain both standardized measurements and descriptive information.

Deep phenotyping can be defined as a precise and comprehensive collection of data on multiple phenotypes in one cohort. Such phenotyping will be time-consuming and require advanced instruments and standardization. Several major projects in psychiatry, including B-SNIP [146], ENIGMA [145], and ABCD [196], are collecting deep phenotyping data, including brain imaging, eye tracking, EEG, and cognitive function.

Shallow phenotyping collects phenotype data in large numbers of subjects but with less content related to illness and less emphasis on standardization. When it was realized that tens of thousands of subjects would be needed to establish associations of common SNPs with polygenic diseases, a minimal phenotyping strategy was adopted. Self-reported or medical record diagnosis, survey at a distance (online), or passive data recorded by a mobile device was used, but concerns remain on the validity or accuracy of the phenotypes [197]. In a study of “chronotype”, which aimed at studying persons with a variant sleep cycle, validation of sleep episode timing was obtained from activity meters in a subset of the UK Biobank samples. However, the relationship of chronotype to specific sleep laboratory measurements (such as sleep phase or REM onset) could not be determined [109].

Shallow phenotyping strategies without any validation of the measurements have been used for most genetic studies of psychiatric disorders in the past decade, with success in detecting disease-associated SNPs, but without demonstrated successes from large-scale importation of other medical record data. 23andMe and the UK Biobank collect phenotype data mostly through shallow phenotyping approaches without external validation. Correlation between endophenotypes can help validate some endophenotypes collected from shallow phenotyping approaches and even help to impute missing data. The shallow phenotyping likely boosted sample size with the cost of adding noises. The success of genetic mapping depends on the fast-growing sample size over the noises.

Electronic Health Records (EHR) represent a hybrid approach that encompasses both deep and shallow phenotypes. Typically, EHRs include rigorously standardized lab tests and physiological results, complemented by less validated data like natural language descriptions. However, the data quality in EHRs can vary significantly across different fields.

To facilitate the development of endophenotype 2.0, we created An Endophenotype Quality Index (EQI) with 12 scoring criteria (Supplementary Material). This index differs from Glahn et al.’s Endophenotype Ranking Value [198], which emphasizes the heritability of endophenotypes and illness and their correlations. Our EQI criteria add several practical measures, including stability, throughput, and costs. We give extra weight to objective and quantitative measures. We also consider generalizability related to age, population, sex, and clinical state. A good endophenotype should be suitable for affordable, high-throughput, large-sample studies with a clearly defined target population.

Uses of the genetics of endophenotypes

Genetic analysis of endophenotypes may assist functional annotation of GWAS signals in disorders and illustrate the pathways, biological processes, and mechanisms involved. We expect the integrative analysis of the genetics of endophenotypes with disease diagnosis will facilitate the discovery of novel risk genes and fine-mapping of causal variants and may also improve the accuracy of disease risk prediction.

Colocalization can be used for fine-mapping disease-associated SNPs. When a disease shares association signals with an endophenotype, the posterior probability of colocalization at a single causal variant can be estimated by a colocalization method like HyPrColoc [199] taking Linkage Disequilibrium (LD) into account.

An endophenotype can also serve as a functional annotation of an SNP when the SNP and its LD region are associated with a disorder. However, this assumption could be wrong because of horizontal pleiotropy. A causal analysis, such as MR, better supports this type of connection.

Risk prediction is an important goal of genetic study, and endophenotypes can play a role

PRS is the major measure for risk prediction today, although there are large overlaps between case and control groups. In a recent evaluation of risk prediction methods for PRS in SCZ [200], the best PRS predictor accounted for 9.9% of risk variance on a liability scale in populations of European descent (EUR). EUR-PRS-based prediction accuracy is much lower when applied to non-EUR because of population genetic structure. Another study on PRS prediction of psychosis in high-risk populations came to similar findings [201]. The current use of PRS in predicting psychiatric outcomes has been critically reviewed [202, 203]. Methods are actively developed to incorporate endophenotypes to improve PRS prediction on diagnosis [204].

Coronary Artery Disease (CAD) offers an interesting example of endophenotype-related risk prediction for SCZ and related psychosis disorders. CAD is a polygenic disorder with a set of risk factors first quantified in the Framingham Heart Study [205]. The risk factors include the components of serum cholesterol, blood pressure, cigarette smoking, sex, and age. Serum triglycerides, previously thought to be a CAD risk factor, failed the MR test [206]. The c-statistic, which is equivalent to AUC, is used to quantify risk prediction in CAD [207]. In a 1998 CAD risk study, the c-statistic was 0.74 for women and 0.77 for men. Using a PRS for CAD in 2020, Damask et al. found that the use of both high baseline LDL-C and high PRS yielded improved risk estimations in the placebo control group as compared with the use of either factor alone [208].

Psychosis disorders and other polygenic disorders may also benefit from the study of endophenotypes for treatment and prevention, as well as for investigation of the biology of illness [8]. Multivariate analysis of commonly used endophenotypes, including fitted regression models and machine learning [209] in the same individuals enables endophenotype intercorrelations to be used in risk prediction [210]. Multivariate regression can improve prediction results over models restricted to genotypes on the risk of psychosis and other disorders [210, 211]. When there are large numbers of phenotypes, analysis faces challenges of multiple-testing burdens and correlated phenotypes [212, 213]. Significance thresholds must be adjusted for correlations among the phenotypes, sometimes requiring simulation or permutation in addition to theoretical distribution analyses [214].

Multi-dimensional endophenotype data for subtyping diseases and resolving disease mechanisms

Multi-dimensional data of endophenotypes may be used to create subtypes of disease if one disorder can be partitioned into subgroups with specific risk factors and genetics and subsequently with optimal treatment choices. The B-SNIP project has developed biotypes of psychosis as physiologically coherent subgroups [162], as an important pilot effort that may reshape future diagnosis systems. High and low-inflammation subgroups of SCZ and BD were also proposed [215, 216]. Fast-evolving machine-learning methods will particularly benefit from structured multi-dimensional data in developing new disease classification systems.

With large and multi-dimensional data of endophenotypes and diseases from diverse sources, path analyses can be performed as imagined in Fig. 1. Path analysis utilizes multiple regression to evaluate causal relationships among parts of a system. Every path, or edge in a networked system, can be assessed for effect size and probability of mediation effect. The networked paths link genetic variants to many endophenotypes, ultimately to a disorder. These networks may be investigated as hypothesized biological processes underlying the disorder.

In summary, to address the limitations of the current endophenotype formulations dating back over a half-century, we propose Endophenotype 2.0 with an updated definition and criteria. Building larger cohorts of patients and controls with deep and shallow phenotyping of endophenotypes and studying their intertwined genetics and genomics may be a fruitful scientific endeavor for the coming years. Multi-dimensional genetic/genomic studies of endophenotypes will hopefully provide biological insights, establish mechanisms for psychiatric disorders, uncover more risk genes, and inform translational use of genetic findings. Improving existing endophenotypes and developing new ones that meet quality criteria are current challenges for the field. In the end, endophenotypes at different levels can be expected to work together to solve the mechanistic steps between genetics and disease.

Author contributions

CL conceived the study and drafted the manuscript. ESG contributed additional ideas and assisted in revising and refining the manuscript.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-024-03195-1.