Precision Medicine Approaches to Mental Health Care

Abstract

Developing a more comprehensive understanding of the physiological underpinnings of mental illness, precision medicine has the potential to revolutionize psychiatric care. With recent breakthroughs in next-generation multi-omics technologies and data analytics, it is becoming more feasible to leverage multimodal biomarkers, from genetic variants to neuroimaging biomarkers, to objectify diagnostics and treatment decisions in psychiatry and improve patient outcomes. Ongoing work in precision psychiatry will parallel progress in precision oncology and cardiology to develop an expanded suite of blood- and neuroimaging-based diagnostic tests, empower monitoring of treatment efficacy over time, and reduce patient exposure to ineffective treatments. The emerging model of precision psychiatry has the potential to mitigate some of psychiatry’s most pressing issues, including improving disease classification, lengthy treatment duration, and suboptimal treatment outcomes. This narrative-style review summarizes some of the emerging breakthroughs and recurring challenges in the application of precision medicine approaches to mental health care.

Introduction

Because of the elusive etiology of mental disorders, current diagnostic methods, detailed in the Diagnostic and Statistical Manual (5th edition) (1) and the International Classification of Diseases (10th edition) (2), rely on distinguishing mental disorders by their symptoms. Syndromic disease classification relies upon diagnostic criteria that cannot be reliably reduced to a single condition (3), confounding both accurate diagnoses as well as treatments. The imprecision of psychiatric diagnoses can be seen through the currently low interrater reliability rates across clinical conditions. Precision psychiatry and longitudinal physiological patient monitoring have the potential to objectify diagnoses and improve treatment outcomes across the psychiatric spectrum. As psychiatric diagnoses become more precise, the molecular and physiological underpinnings of these diseases can be more accurately defined, opening up avenues for more targeted and effective treatment options.

The limitations of the current subjective approach to patient care are easily understood in major depressive disorder (MDD). Diagnostic error is common, especially in primary care settings, with low interrater reliability and studies reporting misdiagnosis rates of up to 66% (4, 5). In terms of treatment accuracy, >40 medications exist for the treatment of MDD alone, with few clinical guidelines existing for drug selection, making successful antidepressant selection an odds-defying task (6). Downstream effects of a trial and error-based approach to antidepressant selection include poor treatment response, with <30% of patients responding to the first antidepressant trial, and suboptimal remission rates, with an estimated 67% of patients ultimately entering remission (7). Even when patients are diagnosed correctly and prescribed an appropriate antidepressant, reduction in MDD symptoms can take 4–6 wk or longer (6). For those who are misdiagnosed and subsequently treated unsuccessfully, the timeline to recovery is extended and accompanied by side effects from ineffective course(s) of treatment.

This review highlights the potential of precision medicine and multi-omic physiological profiling for objectifying psychiatric diagnoses and improving treatment effectiveness. Although there has been a broad range of advancements in precision psychiatry, including approaches that use behavioral and psychometric factors (8–12), this review focuses on approaches that leverage biomedical factors. Advancements in precision diagnoses include a pilot electroencephalography (EEG)-based algorithm capable of differentiating patients with MDD from healthy control subjects with 98% accuracy (13) and blood-based biomarker tests for depression with 91% accuracy (14). Psychiatric treatment selection is also being reimagined by precision medicine, as in the case of an algorithm capable of informing antidepressant selection using physiological indicators including genomic, epigenomic, and imaging biomarkers (15). In addition to advancements that address the diagnostic and treatment aspects of patient care independently, some researchers are intertwining the two to address mental health care more comprehensively. For example, functional MRI (fMRI) biotyping of neural circuits can increase the resolution of MDD diagnosis and enable more effective treatment selection, with 77% predictive accuracy of treatment response to antidepressants (16).

Although many of the advancements in precision psychiatry have yet to be clinically translated, they represent a transition to treatments that are based on the field’s improving understanding of the biological basis of mental disorders. A more complete understanding of the underlying etiology of mental illnesses will enable greater specificity during diagnosis and treatment selection, improving patients’ experience with psychiatric care. Individual-level precision treatment and at-home continuous monitoring of dynamic fluctuations in health and disease may be several years away, but broader approaches to matching biological phenotypes to psychiatric conditions and treatment options within psychiatry already have the potential to improve health outcomes for specific patient subgroups in the near future, allowing for more immediate access to precision psychiatry as the field develops a more nuanced understanding of the physiological basis of mental health and disease (FIGURE 1).

FIGURE 1.

Overview of precision mental health care This figure provides an overview of the future of precision psychiatry. With precision medicine tools, such as EEG and multi-omics technologies, patients can be diagnosed and treated with greater specificity and accuracy, by virtue of a better understanding of their disease state at the biological level. fMRI, functional MRI; SNRI, selective norepinephrine reuptake inhibitor; SSRI, selective serotonin reuptake inhibitor.

Recent Breakthroughs in Precision Psychiatry

By providing more objective metrics for once subjectively defined disorders, recent advancements in genetics, epigenetics, and biomarkers are aiding medical professionals in their quest to provide accurate diagnoses and effective treatment plans. A summary of the discussed advancements can be found in Table 1 and Table 2.

TABLE 1.

Precision psychiatry biomarker summary table

Disorder	Diagnostic		Treatment
	Applications	Sample	Applications	Sample
Major depressive disorder (MDD)	MDDScore test capable of classifying blood samples as healthy or MDD with 91–94% accuracy. Model considers concentrations of alpha1 antitrypsin, apolipoprotein CIII, brain-derived neurotrophic factor, cortisol, epidermal growth factor, myeloperoxidase, prolactin, resistin, and soluble tumor necrosis factor α (TNF-α) receptor type II, BMI, and sex (17).	68 MDD, 86 control	fMRI findings of changes in 6 neuronal circuits can be used to explain specific features of depression (i.e., rumination, anxious avoidance), possibly allowing for more descriptive diagnoses. These more specific diagnoses may allow for the identification of more effective treatment selection (18).	NA
	Support vector machine capable of classifying subjects as MDD or healthy control with 98% accuracy, 99.9% sensitivity, and 95% specificity with a 5-min EEG recording (13).	34 MDD, 30 control (Malaysian descent)	Elastic net regularized regression-based model using demographic data, baseline severity, depression subtypes and symptoms, stressful life events, and medication history was shown to accurately predict patient response to escitalopram and, less accurately, nortriptyline (19).	793 MDD (328 nortriptyline, 465 escitalopram)
	Using a 3-min EEG recording to glean information about interhemispheric asymmetry, cross-correlation, and mixed features of the 1-s, 2-s, and 3-s segments of the α-, β-, and θ-frequency bands, a convolutional neural network (CNN) was created and capable of differentiating MDD from control with 96% sensitivity, 94% sensitivity, and 94% accuracy (20).	16 MDD, 16 control	With left and right hemisphere EEG data, a CNN capable of differentiating between patients with MDD and healthy control subjects was created. The model that utilized left hemisphere data obtained a classification accuracy of 93.5%, whereas the model that employed the right hemisphere data had a 95.5% accuracy (21).	15 MDD, 15 control
	Using discrete wavelet transform, total variation filtering, and relative wavelet energy on resting-state EEG data, a feedforward artificial neural network (ANN) can distinguish depression from healthy control subjects with 98% sensitivity, 98% specificity, and 98% accuracy (22).	30 patients with depression, 30 control subjects	ARPNet, a neural network that utilizes linear regression, predicts antidepressant response using 62 MRI features, 27 genes, and 136 sites of DNA methylation with 84.6% accuracy (15).	121 MDD (Korean descent)
	Via the quantification of effective connectivity from EEG data, a CNN can differentiate between MDD and control with 99.24% accuracy (23).	34 MDD, 30 control	Via kernel partial least squares regression, a model capable of classifying MDD patients as responders or nonresponders to sertraline based on pretreatment EEG with 86.6% accuracy was created (6).	22 MDD
Bipolar disorder (BD)	Model using differentially abundant 2,3-diphospho-d-glyceric acid, N-acetyl aspartyl-glutamic acid, and monoethyl malonate levels can discriminate BD blood samples from SZ and healthy blood samples with 88.6% accuracy (24).	68 BD, 54 SZ, 60 control	Blood TNF-α levels found to be elevated in patients who respond poorly to lithium, relative to good responders, opening up the possibility for usage as a biomarker of treatment response (25).	60 BD using lithium (17 good responders, 23 partial responders, 20 poor responders)
Schizophrenia (SZ)	Differentially abundant isovaleryl carnitine, pantothenate, mannitol, glycine, and GABA can discriminate blood samples of SZ patients from those of BD patients and control subjects with 80.6% accuracy (24).	68 BD, 54 SZ, 60 control	EHF rs286913 (P = 6.99 × 10−8), SLC26A9 rs11240594 (P = 1.4 × 10−7), and IL1A rs11677416 (P = 6.67 × 10−7) were found to be associated with improvements in vigilance, processing speed, and working memory, respectively, all features of improving neurocognition, a broad marker of treatment response (26).	738 SZ from CATIE
	Machine learning algorithm capable of classifying subjects with paranoid schizophrenia with EEG data with 96.8% accuracy (27)	14 paranoid SZ, 14 control	Baseline transfer entropy (EEG metric representing information flow) found to predict electroconvulsive therapy response with 88.7% sensitivity, 81.8% specificity, and 85.2% accuracy (28).	40 SZ, 7 schizoaffective
	Support vector machine considering sensor and source-level EEG data captured during an auditory oddball event can distinguish between SZ and control with 88.24% accuracy. When only sensor or source-level EEG data were used, classification accuracy dropped to 80.88% and 85.29%, respectively (29).	34 SZ, 34 control	Risperidone, olanzapine, and quetiapine (antipsychotics) found to differentially alter blood plasma lipid levels in nonresponders and responders (30)	54 SZ (23 risperidone, 17 olanzapine, 14 quetiapine users)

This table summarizes sample genomic features related to the field of precision psychiatry, aggregated by disorder and diagnostic or treatment relevance. BMI, body mass index; CATIE, Clinical Antipsychotic Trials of Intervention Effectiveness; EEG, encephalography; fMRI, functional magnetic resonance imaging; GABA, γ-aminobutyric acid; MRI, magnetic resonance imaging.

Table 2.

Precision psychiatry genomics summary table

Disorder	Diagnostic Applications				Treatment Applications
	Findings	Sample	Findings	Sample	Findings	Sample	Findings	Sample
Major depressive disorder (MDD)	MA examining 393 polymorphisms in 102 genes determined APOE, GNB3, MTHFR, SLC6A3, and SLC6A4 to be statistically associated with MDD susceptibility (31).	MA; 183 papers	BDNF, FKBP5, CRHBP, and NR3C1 gene promoters differentially methylated in PBMCs in MDD patients with and without suicidal ideation (32)	14 MDD patients with and 10 without suicidal ideation, 20 control subjects	Corticotropin-releasing hormone binding protein rs28365143 associated with SSRI (escitalopram, sertraline) response. G allele associated with greater treatment response to tested SSRIs but not SNRI (venlafaxine), in comparison to A allele (33)	636 MDD patients from iSPOT-D, 344 MDD patients from PReDICT	Kernel partial least squares regression capable of predicting MDD patient response to sertraline with 86.6% accuracy based on an EEG recording (6)	22 MDD patients
	MA revealed that rs1360780 T allele (P = 0.003) and rs3800373 C allele (P = 0.002) of FKBP5 significantly associated with MDD risk (34).	MA; 7 papers; 12,491 MDD patients, 14,091 control	Hypermethylation in the coding region of the ZBTB20 and increased global variation in methylation in whole blood samples from MDD patients (35)	50 discordant MZ twins with MDD (Australian and British descent)	UST rs2500535 (P = 3.56 × 10−8) associated with nortriptyline response. IL11 rs1126757 (P = 2.83 × 10−6) associated with escitalopram response (36)	706 MDD patients (394 escitalopram, 312 nortriptyline users) of European ancestry	ACSS3 rs10492002 (P = 4.5 × 10−6) associated with sustained citalopram response (37)	247 MDD patients with unsustained, 869 sustained improvement from STAR*D, replicated with 586 from GENDEP
	GWAS yielded 44 genome-wide significant loci associated with MDD. Implicated genes involved in anatomical differences in brain regions (TCF4), antidepressant mechanisms (CACNA1E, CACNA2D1, DRD2, GRIK5, GRM5), neuroinflammation (LRFN5), and gene splicing (RBFOX1). 19 multifunctional pathways (i.e., synaptic genes, neuronal morphogenesis, neuron projection, genes involved in SZ, CNS neuron differentiation) were found to be associated with MDD (38).	MA; 135,458 MDD patients, 344,901 controls	2 genome-wide significant loci, SIRT1 rs12415800 (P = 2.53 × 10−10) and LHPP rs35936514 (P = 6.45 × 10−12), were reported to be associated with recurrent MDD. Findings replicated in an alternate, ethnically Chinese sample (39).	5,303 women with MDD, 5,337 control subjects (Chinese descent)	10 SNPs (P = 4.98 × 10−7) of SACM1L gene associated with sexually adverse bupropion (aminoketone) side effects (40): rs2742417, rs2251954, rs2742421, s2742423, rs1969624, rs2673057, rs2742431, rs2742435, rs2245705, rs2742390	1,439 MDD patients from STAR*D	EMID2 rs17135437 (P = 3.27 × 10−8) genome-wide significant associations with onset of visioauditory side effects of citalopram. RPT2 rs6764050 (P = 8.73 × 10−7) associated with general side effect burden due to citalopram usage (41).	1,762 MDD patients from STAR*D
	GWAS revealed 2 replicable and genome-wide significant loci (KSR2 intron variant rs7973260 and DCC intron variant rs62100776) associated with depressive symptoms (42).	MA; 180,866 patients from PGC, UKB, and GERA with depressive symptoms	Although GWAS did not reveal any genome-wide significant SNP associated with rumination, rumination was found to be associated with the KCTD12 and miR-383 binding genes at the gene and gene set level, respectively (43).	1,758 adults (European descent)	Compared with pregnant CYP2D6 extensive and ultrarapid metabolizers, pregnant poor and intermediate metabolizers are 4 more times likely to stop antidepressant usage (44).	246 pregnant women using antidepressants (204 EM, 8 UM, 14 PM, 20 IM)	GDA rs11143230 (P = 8.28 × 10−7) strongly associated with increasing suicidal ideation during antidepressant treatment (45)	244 MDD patients, 462 controls (European descent) from GENDEP
					All CYP2D6 ultrarapid metabolizers discontinued treatment within 4 wk, no poor metabolizers discontinued treatment within 12 wk (46).	100 MDD patients (55 fluoxetine, 45 amitriptyline users)	Only CYP2D6 ultrarapid metabolizer responded positively to 375 mg venlafaxine without any adverse side effects (47).	47 MDD (47 venlafaxine users)
	SIRT1 rs80309727 (P = 2.95 × 10−10) significantly associated with melancholia, a severe subtype of MDD (39)	4,509 women with melancholia, 5,337 control subjects (Chinese descent)	SIRT1 mRNA expression dampened in postmortem peripheral blood (decreased by 12.3%) and brain samples in patients with MDD (48)	1,824 MDD patients, 3,031 control subjects (Chinese descent)	CYP2C19 PM genotype more strongly associated with reduction in symptoms and remission compared with EM genotype, in the context of citalopram or escitalopram treatment (49)	2,558 MDD patients from GENDEP, STAR*, GenPod, PGRN-AMPS	Therapeutic adjuvants (low-dose paroxetine) can be prescribed to maintain therapeutic ranges of nortriptyline and nortriptyline's metabolite, 10-hydroxynotriptyline, via modulating CYP2D6 activity (50).	15 psychiatric patients (unspecified disorder) using nortriptyline
							PAPLN rs11628713 (P = 6.2 × 10−7) associated with treatment-emergent suicidal ideation during citalopram treatment (51)	90 MDD patients with TESI from STAR*D, 90 control subjects (European descent)
Bipolar disorder (BD)	GWAS yielded 31 genome-significant SNPs in 30 loci associated with BD risk. These loci were genes coding for ion channels, neurotransmitter transporters, synaptic components, lipid-related biomarkers, and more (52).	20,352 BD patients, 31,358 control subjects (American/European descent)	Duplications on chromosome 16p11.2 were more common in BD patients compared with healthy control subjects (P = 2.3 × 10−4). Duplications at 1q21.1 (P = 0.022) and deletions at 3p29 (P = 0.03) were deemed nominally significant before correction for multiple testing (53).	2,591 BD patients from UK, 6,882 SZ, 8,842 control	Lithium treatment response strongly associated with ACTC on chromosome 15q14 (P = 1.4 × 10−5) and D7S1816 on chromosome 7q11.2 (P = 1.1 × 10−4) (54)	106 BD patients	GWAS revealed 4 SNPs in the same region of chromosome 21 to be genome-wide significant and associated with lithium response. rs74795342 (P = 3.31 × 10−9) and rs75222709 (P = 3.50 × 10−9) are located in the intron of AL157359.3, which codes for a lncRNA. Also coding for a lncRNA, rs79663003 (P = 1.37 × 10−8) and rs78015114 (P = 1.31 × 10−8) are found between AL157359.3 and AL157359.4 (55).	2,935 BD patients using lithium (European and Taiwanese descent)
	Hypermethylation of COMT and PPIEL genes associated with BD (Zhang et al. 2018)	150 BD patients, 50 control subjects	Epigenome-wide association study revealed that CpG sites in FAM63B and IR were significantly hypomethylated in BD, whereas CpG sites in TBC1D22A were hypermethylated in whole blood samples belonging to patients with BD (56).	450 BD patients, 457 control subjects	Via genome-wide miRNA analysis, 71 miRNAs were found to have nominally significant associations with lithium treatment response, with miR-633 demonstrating the strongest association with continuous response (P = 9.80 × 10−4) (57).	2,563 BD patients from ConLiGen
Schizophrenia (SZ)	GWAS yielded 108 genome significant loci associated with SZ risk. Implicated genes are involved in glutamatergic neurotransmission, synaptic plasticity, antipsychotic drug mechanism, calcium transport, and more (58).	36,989 SZ patients, 113,075 control subjects	Via convergent functional genomics, RERE rs3765971 (P = 3.87 × 10−8), ARL6IP4 rs55742290 (P = 1.50 × 10−7), and CENPM rs7293091 (P = 5.09 × 10−7) found to be potential risk variants for SZ (59)	35,476 SZ patients, 46,839 control subjects (European descent)	Using 13 or 25 SNPs and 53 sociodemographic and clinical variables, a latent group effectiveness modeling algorithm capable of predicting SZ patient response to ziprasidone (antipsychotic) with 75% sensitivity, 74% specificity, and 74% accuracy was created. 19 other models were created to predict response to 4 other antipsychotics, all but 2 performing better than chance (60).	100 SZ patients from CATIE	5 genome-wide significant loci found to be associated with general antipsychotic response: rs72790443 MEGF10 (P = 1.40 × 10−9), rs1471786 SLC1A1 (P = 2.33 × 10−9), rs9291547 PCDH7 (3.24 × 10−9), rs12711680 CNTNAP5 (P = 2.12 × 10−8), and rs6444970 TNIK (P = 4.85 × 10−8). rs2239063 CACNA1C associated with olanzapine response (P = 1.10 × 10−8). rs16921385 SLC1A1 associated with risperidone response (P = 4.40 × 10−8). rs17022006 CNTN4 associated with aripiprazole response (P = 2.58 × 10−8). Genetic risk score created with these SNPs could classify patients as responders or nonresponders with 64.8% sensitivity and 68.7% specificity (61).	3,792 SZ patients (Chinese descent) using olanzapine, risperidone, quetiapine, aripiprazole, ziprasidone, haloperidol, or perphenazine from CAPOC
	Hypomethylation of PSMD5, AEN, FAM20B, LRRN4 genes and hypermethylation of ID2 gene, increased expression of histone methyl-transferases, aberrant expression of serum miRNA associated with SZ (62)	Review of 12 papers	923 differentially methylated CpGs found to be associated with SZ; 625 replicated in an alternate cohort. Implicated genes were reported to be involved in neuronal cell fate, T-cell development, synaptic plasticity, and neurogenesis (63).	689 SZ patients, 645 control subjects from CGES, PAAAERS, MMFSS	GWAS revealed that GRM7 rs2133450 (P = 4.33 × 10−8) is associated with risperidone response 2 wk after treatment initiation. In an independent sample, the rs2133450 CC genotype was associated with poorer risperidone response over 9 mo compared with the AC and AA genotypes (64).	106 SZ patients (Italian descent) from GEAR-RisGW

This table summarizes sample precision psychiatry biomarker innovations, aggregated by disorder and diagnostic or treatment relevance. CAPOC, Chinese Antipsychotic Pharmacogenetics Consortium; CATIE, Clinical Antipsychotic Trials of Intervention Effectiveness; CGES, Consortium on the Genetics of Endophenotypes in Schizophrenia; CNS, central nervous system; ConLiGen, International Consortium on Lithium Genetics; EM, extensive (normal) metabolizer; GEAR-RisGW, Genes and Early Antipsychotic Response-Risperidone Genome-Wide Trial; GENDEP, Genome-Based Therapeutic Drugs for Depression; GenPod, Genetic and Clinical Predictors of Treatment Response in Depression; GERA, Genetic Epidemiology Research on Aging; GWAS, genome-wide association study; IM, intermediate metabolizer; iSPOT-D, the International Study to Predict Optimized Treatment in Depression; lncRNA, long noncoding RNA; MA, meta-analysis; MMFSS, Multiplex Multi-generational Family Study of Schizophrenia; MZ, monozygotic; PAAAERS, Project among African-Americans to Explore Risks for Schizophrenia; PBMC, peripheral blood mononuclear cell; PGC, Psychiatric Genomics Consortium; PGRN-AMPS, Pharmacogenomic Research Network Antidepressant Medication Pharmacogenomic Study; PM, poor metabolizer; PReDICT, Predictors of Remission in Depression to Individual and Combined Treatments; SNP, single-nucleotide polymorphism; SNRI, selective norepinephrine reuptake inhibitor; SSRI, selective serotonin reuptake inhibitor; STAR*D, Sequenced Treatment Alternatives to Relieve Depression; TESI, treatment-emergent suicidal ideation; UKB, United Kingdom Biobank; UM, ultrarapid metabolizer.

Precise Diagnostic Methods

Genetics.

Twin studies and genome-wide association studies (GWASs) suggest that complex gene-environment interactions are responsible for the onset of many psychiatric disorders, with, for example, genetic factors accounting for roughly 40% of MDD inheritance (65–67). Seeing as genetic factors interact with environmental stimuli to influence the onset and progression of MDD, genetic features have the potential to be useful in a clinical setting as supporting evidence for psychiatric diagnoses, especially in the context of differential diagnoses. However, to date, genetic features have shown limited predictive value, highlighting the nonmendelian nature of psychiatric disorders and the importance of contextualizing genetic contributions within the framework of environmental exposures and the epigenome.

Despite large sample sizes, the majority of GWASs have struggled to find genome-wide significant and reproducible genetic features associated with psychiatric disorders, especially MDD, which is postulated to require a sample size five times larger than that of schizophrenia to achieve equal statistical power (38). This is likely because, at least in part, psychiatric disorders are not as neatly defined in nature as they are in the clinic: what clinicians refer to as MDD may constitute dozens of distinct disorders with unique etiologies, physiological underpinnings, and genetic characteristics. GWASs that have focused on a specific subtype of MDD, such as melancholia, have shown greater success (38, 68).

As the genomic and epigenomic basis of psychiatric disorders becomes more detailed, polygenic risk scoring frameworks, which have shown limited predictive value to date, with some exceptions (38, 61, 69), have the potential to improve in clinical utility. Of note, many polygenic risk scoring frameworks may be more strongly associated with ancestry than psychiatric disease, as in the case of schizophrenia (70). Specific genetic and epigenetic features associated with distinct psychiatric conditions are detailed in Table 2.

Blood-based biomarkers and point-of-care tests.

Despite only recently being explored in the realm of psychiatry, blood-based biomarkers have been historically relied upon in virtually all other medical specialties. With further validation and discovery, blood-based biomarkers are poised to redefine the psychiatric diagnostic paradigm, given their demonstrated ease of clinical translation.

Several blood-based diagnostic tests have been developed, including for schizophrenia (24, 71–73). Tasic et al. (24) demonstrated the feasibility of differentiating between the blood analyte signatures of patients with schizophrenia and bipolar disorder, as well as healthy patients. Biomarkers differentially expressed in schizophrenia, bipolar disorder, and healthy samples were revealed via ¹H nuclear magnetic resonance (¹H-NMR) spectroscopy data from 182 blood samples (n = 54 schizophrenia, n = 68 bipolar, and n= 60 healthy control) (24). 2,3-Diphospho-d-glyceric acid, N-acetyl aspartyl-glutamic acid, and monoethyl malonate were common to patients with bipolar disorder, whereas isovaleryl carnitine, pantothenate, mannitol, glycine, and GABA levels were reported to be higher in schizophrenic patients (24). Additionally, 6-hydroxydopamine, a neurotoxin that selectively destroys dopaminergic and noradrenergic neurons in the brain (74), was detected in both bipolar and schizophrenic blood samples but not in those of healthy control subjects (24). Using these differentially abundant physiological markers, the resultant model was capable of classifying blood samples as belonging to patients with schizophrenia with 80.6% accuracy, bipolar disorder with 88.6% accuracy, and healthy control subjects with 83.9% accuracy (24). Tests capable of differentiating between bipolar disorder and schizophrenia, such as that of Tasic et al., are of clinical value because of the challenging and common differential diagnosis.

A similar blood-based diagnostic test was developed by Papakostas et al. (14) for MDD. This test utilizes the serum levels of nine biomarkers: alpha1 antitrypsin, apolipoprotein CIII, brain-derived neurotrophic factor, cortisol, epidermal growth factor, myeloperoxidase, prolactin, resistin, and soluble tumor necrosis factor α receptor type II (14). A sensitivity of 91.7% and a specificity of 81.3% were reported in the pilot study (14). Similar results were achieved upon replication in an independent cohort (91.1% sensitivity, 81.3% specificity), suggesting that this test can differentiate between blood samples from patients with MDD and healthy control subjects (14). Whether similar outcomes would be achieved in alternative patient populations (such as ethnically diverse samples and varying age groups) is unknown. Using the previously identified biomarkers, sex, and body mass index (BMI), the group created MDDScore, an algorithmic laboratory aid used to diagnose MDD (17). This test is capable of differentiating between MDD and healthy control blood samples with 94% sensitivity, 92% specificity, and accuracy ranging between 91% and 94% (17). The addition of sex and BMI improved model accuracy, suggesting the importance of contextualizing biomarker abundance using these variables. Similar results were seen for patients who were both drug-free and on antidepressants (17). Given that patients seeking psychiatric help are frequently taking antidepressants, models that perform well under real-world conditions, like MDDScore, are of immense clinical value.

An individualized approach to precision medicine can involve comparing an individual to themselves over time (e.g., high mood states vs. low mood states) to identify markers that change with mood. Blood-based biomarkers have been explored on an individual level by examining within-subject changes in biomarker expression levels. Comparing a cohort of patients during low and high mood states, Le-Niculescu et al. (75) identified 26 candidate biomarkers useful for elucidating depression severity and future risk of mood disorder onset. These biomarkers are implicated in circadian, neurotrophic, and cell differentiation processes (75). The resultant panel is capable of producing an interpretative report, yielding a depression score (severity of depression) as well as future risk metrics for depression and bipolar disorder (75). Given the easy interpretation of patient depression status and risk, this kind of tool has potential to objectify mental health care and provides clinically actionable information that can be utilized by health care providers who are not trained in psychiatry but regularly treat mental illnesses, such as family practice physicians. Furthermore, risk-prediction models such as this one begin to address the need for a reliable way to identify individuals who are predisposed to mental illnesses and opens the possibility of targeting preventative care to those at an elevated risk.

Well-validated blood-based biomarkers have the potential to provide accurate diagnoses and are of exceptional clinical interest, as they are both noninvasive and widely implementable. Before implementation, the validity of the aforementioned blood-based biomarkers and algorithms as well as other emerging solutions must be tested on a larger, ethnically diverse sample to validate their accuracy and that models are not overfit to their initial study cohorts. Although some studies have demonstrated reproducibility of their findings in independent secondary cohorts, many have not yet, raising the possibility of model overfitting. Additionally, further investigation is warranted to develop tools capable of discriminating between current mental health diagnoses as well as helping to better define diagnostic categories altogether.

Electroencephalography-based diagnostics.

Another promising tool for precision psychiatry diagnostics is electroencephalography (EEG). With EEG, candidate diagnostics have been developed for MDD and schizophrenia, many of which have produced models with predicative accuracy > 90%. Given its noninvasive nature, low cost, and ease of use for longitudinal monitoring, EEG is positioned to enable the development of highly accessible and objective diagnostics in psychiatry.

Many EEG-based models have been developed with the aim of differentiating MDD patients from healthy control subjects (13, 20, 21, 23). Mumtaz et al. (13) created a model capable of classifying patients with and without MDD using a sample of 34 patients with MDD and 30 age-matched healthy control subjects, all of Malaysian descent. Depressive symptoms were quantified with the Beck Depression Index and the Hospital Anxiety and Depression Scale (HAM-D) (13). From 5-min EEG recordings captured with the standard 10–20 lead montage, several models (support vector machine, logistic regression, naive Bayesian) were tested (13). The support vector machine model showed the highest predictive power, with 98% accuracy, 99.9% sensitivity, and 95% specificity (13). Considering the high performance and balanced data set, these results are promising and reflective of EEG’s predictive power. However, the non-sex-matched nature of the study cohort as well as its small sample size are a significant limitation.

EEG-based algorithms to inform schizophrenia diagnoses have also shown promise (27, 29, 76–79). Buettner et al. (27) developed a machine learning model that utilizes a 1-min EEG recording as input, also using the standard 10–20 montage. The data were classified with the random forest classifier and validated with k-fold cross validation with 10 iterations (27). Of 124 samples, only 4 were misclassified, producing a balanced accuracy of 96.8% (27). This study had a small sample size of 28 participants (14 healthy and 14 paranoid schizophrenic patients), and participants were all over 18 yr of age, drug-free for 7 days, and without diagnosed comorbidities (27), and thus the applicability to patients who fall outside these categories is unknown. Despite these limitations, these results are encouraging and reflective of the benefits of developing precision medicine tools for specific phenotypes of broader disease such as paranoid schizophrenia.

EEG-based biomarkers and algorithms have shown great promise as potential diagnostic tools, with many studies reporting high diagnostic accuracy rates. Similar to blood-based assays, there is a potential for using data from EEG to restructure our current understanding of psychiatric disease classifications so they better reflect the underlying physiology of disease states. Despite the high accuracy of initial EEG studies, few studies have demonstrated the reproducibility of their results in secondary cohorts. This is especially notable for EEG-based biomarkers, given their extremely high reported accuracies and their, on average, smaller study cohorts. If EEG-based biomarkers and resulting algorithms are ever to be clinically translated, more validation work must become standard practice in the field.

Precise Treatment Methods

Given the limited mechanistic understanding of psychiatric disorders and the dozens of treatments available for each disorder, selecting the most beneficial treatment is a formidable challenge. Precision psychiatry is starting to address this issue by matching individual disease phenotypes to targeted courses of treatment. The remainder of this review focuses on applications of precision medicine to psychiatric treatment selection, specifically focusing on computational models, genomic factors, and the intersection of the two. Given that the majority of the advancements in the precision psychiatry treatment space are focused on MDD, this section focuses on treatment breakthroughs in MDD specifically.

Computational models for treatment selection.

Considering the complex nature of psychiatric disease presentation and treatment plan development, computational models are well suited for assisting physicians. Although many of the following innovations are in the preclinical laboratory phase, tools that can assist physicians in optimal treatment selection broadly hold promise for improving psychiatric patient outcomes (80, 81).

Predictive models that exclusively consider genetic factors have yet to perform with the accuracy necessary for clinical implementation. In a study conducted by Maciukiewicz et al. (82), 186 patients with MDD treated with duloxetine [a selective norepinephrine reuptake inhibitor (SNRI)] for up to 8 wk were categorized as responders [Montgomery–Asberg Depression Rating Scale (MADRS) score decrease > 50%] and remitters (MADRS score ≤ 10). Through genome-wide logistic regression, Maciukiewicz et al. (82) discovered 38 single-nucleotide polymorphisms (SNPs) associated with duloxetine response and 44 associated with remission (P < 5 × 10⁻⁵). Of these potential genetic markers of treatment response, those with the highest predictive power were rs2036270, located in the intron of retinoic acid receptor beta, and rs1138545, a missense variant of the tenascin C gene (82), which drives production of inflammatory cytokines and activates innate immunity (83). Based on the variants of interest identified, least absolute shrinkage and selection operator (LASSO) regression, classification-regression trees, and linear support vector machine models were created to independently predict treatment response and remission (82). Models designed to predict duloxetine response did not perform better than chance, and models for remission only performed somewhat better than chance (52% accuracy, 58% sensitivity, 46% specificity), with the best-performing models having sensitivity and specificity of 70% and 61%, respectively (82). The authors did not correct for imbalances in the data set (70% of the sample were duloxetine responders) (82), which may have led to inflated model performance. Additionally, poor model performance may be due to the multicausal nature of MDD and treatment response, reaffirming the accepted theory that MDD onset, presentation, and remission are multifactorial.

To address the challenge of antidepressant selection, Chang et al. (15) created the Antidepressant Response Prediction Network (ARPNet), a neural network tasked with predicting antidepressant response. This linear regression model considers neuroimaging biomarkers (62 MRI features), genetic variants (27 genes of interest associated with antidepressant response), and DNA methylation at 136 CpG sites from SLC6A4, BDNF, IL11, and MAOA genes (15). The model predicts the efficacy (change in HAM-D score) of 14 individual antidepressants and 91 combinations of 2 antidepressants (15). The model had an accuracy of 84.6% in a sample of 121 individuals of Korean descent with MDD (15). ARPNet’s antidepressant response prediction accuracy, as well as its ability to predict which antidepressants a patient will respond to and their degree of improvement, suggest that this tool has high clinical utility (15). Moreover, ARPNet is realistically implemented in a clinical setting because of its ability to predict the outcomes of coadministered antidepressants.

Also addressing the need for more empirical antidepressant selection, Iniesta et al. (19) used clinical and demographic variables to predict response to escitalopram [selective serotonin reuptake inhibitor (SSRI)] and nortriptyline [tricyclic antidepressant (TCA)], demonstrating the utility of these variables in predictive models. In a cohort of 793 MDD patients (328 nortriptyline, 465 escitalopram users), elastic net regularized regression (ENRR) was used to predict treatment outcomes (19). Tenfold cross validation repeated 100 times was conducted to ensure accuracy (19). Of the four models tested, the model that considered demographic data, baseline severity, depression subtypes, symptoms, dimensions, and stressful life events most accurately predicted response to escitalopram (19). Individuals treated with nortriptyline were more likely to experience serious side effects and discontinue treatment prematurely (19), leading to reduced statistical power and predictive accuracy. Remission was predicted with sensitivity and specificity of 66% (19). Models for combination therapy with escitalopram and nortriptyline were consistently less accurate than models accounting for a single drug (19). The need for a diagnostic aid like that of Iniesta et al. is evident, but it is clear that more work must be done before clinical implementation.

EEG-based models can also be used to predict treatment outcomes (6, 84–87). Khodayari-Rostamabad et al. (6) developed a kernel partial least squares regression-based model to predict MDD patient response to sertraline (SSRI). In a cohort of 22 patients, pretreatment EEGs were recorded with the standard 10–20 montage (6). Distinctive features were identified from spectral coherence between electrode and channel pairs, absolute and relative power spectral density, log ratio of left to right hemisphere powers, and anterior-to-posterior power ratios (6). Candidate features were filtered by selecting features with a large Kullback–Leibler distance between the responders and nonresponders (6). An 11-fold nested cross validation was conducted, yielding an average accuracy of 86.6% and specificity and sensitivity of 85.7% and 87.5%, respectively (6). Treatment response was quantified with the HAM-D scale (6). The accuracy of the model suggests that it could be of clinical utility, but replicating the experiment on a larger sample is necessary before implementation in a diverse clinical setting. Nonetheless, these results demonstrate the practical value of EEG for predicting treatment response to antidepressants and potentially other psychiatric drugs.

In addition, functional magnetic resonance imaging (fMRI) is improving the understanding of mental illness at the biological level, which, like genetics, has potential to refine treatment selection and improve outcomes for nonresponders. Williams (18) details the potential of using alterations in activity and connectivity in six neural circuits to subclassify depression and anxiety. Biological information can be gleaned from circuit activity, allowing insight into the specific etiology for a given patient, which can potentially assist providers in selecting treatments that specifically target dysfunctional neuronal activity and aid in the development of more targeted therapies (18). For example, overactivity, hyperconnectivity, and structural changes in the default mode circuit, a network composed of the anterior medial prefrontal cortex, posterior cingulate cortex, and angular gyrus, are associated with rumination (18, 88,89), a symptom characteristic of many psychiatric disorders. Dysfunction in the default mode circuit has been used to predict treatment response to SSRIs with 77% accuracy (16). Classification as a potential SSRI nonresponder may lead a physician to recommend an alternative treatment (16), potentially allowing patients to avoid ineffective courses of treatment and enter remission sooner.

Encouragingly, several precision psychiatry interventions are now being tested in clinical trial settings. One example is the StratCare Trial, which evaluated the efficacy of a machine learning algorithm in assisting treatment selection for depression, anxiety, and post-traumatic stress disorder in a cohort of 951 patients (90). The algorithm utilized clinical variables, such as patient-reported levels of depression, anxiety, and functional impairment, as well as demographic data such as employment status, race, and ethnicity (90). Using this data, the algorithm would then assist the clinician in conducting stratified care by recommending low- or high-intensity treatment (90). Patients assigned to the algorithm-assisted group were 7% more likely to experience a reduction in depressive symptoms and required less treatment, for an additional cost of $150 per patient (90). Highlights of the StratCare approach include demonstrated, though small, clinical benefit, multi-disease applicability, and economic feasibility. Another example of a treatment selection algorithm was a randomized controlled trial of 614 patients with an affective or anxiety disorder (91). Lutz et al. (91) tested the efficacy of an algorithmic treatment selection aid that provided not only pretreatment recommendations but also adaptive recommendations based on patient psychometric data during treatment. Both trials demonstrate the field of precision psychiatry’s push toward larger, higher-quality trials for clinical translation of precision medicine tools (80).

Methods for predicting psychiatric treatment outcomes aim to produce a much-needed, evidence-based framework for therapeutic selection, transforming the standard of care for psychiatric patients. The integration of treatment prediction models into psychiatric care could help patients avoid repeated failed treatments and the myriad accompanying side effects. Precise approaches to mental health treatment can allow for a quicker remission and decrease the overall disease burden of psychiatric disorders.

Genomic results.

Multiple genome-wide association studies have identified SNPs significantly associated with treatment response, resistance, and serious side effects to antidepressants, many of which are summarized in a review by Lin and Lane (92). Findings with corresponding P values < 5 × 10⁻⁸ are generally considered genome-wide significant.

O’Connell et al. (33) explored the relationship between 16 SNPs associated with the hypothalamic pituitary adrenal (HPA) axis, which is dysregulated in MDD patients (93), and treatment response for escitalopram (SSRI), sertraline (SSRI), and venlafaxine (SNRI) in a cohort of 636 patients with MDD from the International Study to Predict Optimized Treatment in Depression (iSPOT-D). Participants completed genomic testing and HAM-D metrics at baseline and 8 wk into therapy (33). Results were validated using data from a second study: the Predictors of Remission in Depression to Individual and Combined Treatments (PReDICT) Study (33). Corticotropin-releasing hormone binding protein rs28365143 was found to predict treatment response (HAM-D score reduction of ≥50%) and remission (HAM-D score ≤ 7) for the SSRIs included in the study in both cohorts (33). The G allele of rs28365143 was associated with greater treatment response to the tested SSRIs compared with the A allele (33). This SNP was not found to be associated with response to the only SNRI tested, venlafaxine (33). These results highlight the importance of contextualizing antidepressant selection in a framework that considers genomic status, as antidepressant response appears heritable and can be leveraged by psychiatrists as they are formulating courses of treatment. Uher et al. (36) also demonstrated the feasibility of linking genetic variants to antidepressant response. In a cohort of MDD patients taking escitalopram (n = 394) and nortriptyline (n = 312), UST rs2500535 (P = 3.56 × 10⁻⁸) was associated with nortriptyline response, whereas IL11 rs1126757 (P = 2.83 × 10⁻⁶) was associated with escitalopram response (36). Hunter et al. (37) conducted a GWAS in hopes of identifying SNPs associated with sustained (n = 869) or unsustained (n = 247) improvement in MDD patients. ACSS3 rs10492002 (P = 4.5 × 10⁻⁶) was strongly associated with sustained response to citalopram treatment (37).

GWAS can also be used to find associations between genetic variants and the onset of antidepressant side effects, as seen in work by Clark et al. (40). With genomic data from 1,439 patients with depression from STAR*D (40), 10 SNPs, accounting for ∼25 kb in SACM1L, were found to be associated with sexually adverse side effects to bupropion (aminoketone) (P = 4.98 × 10⁻⁷) (40). Similarly, Adkins et al. (41) conducted a study examining SNPs associated with side effects of citalopram, using STAR*D genomic data for 1,762 patients with depression. The two most statistically significant SNPs found were EMID2 rs17135437 (associated with the onset of visioauditory side effects) and RPT2 rs6764050 (P = 8.73 × 10⁻⁷, associated with general side effect burden) (41). Moreover, Bopp et al. (94) conducted a GWAS on MDD patients being treated with lithium and explored the relationship between SNPs and weight gain, a common adverse effect of lithium treatment, finding that the rs6979832 SNP of the leptin gene predicted weight gain in the cohort (P = 0.037).

SNPs associated with treatment-emergent suicidal ideation (TESI) and suicidal planning have been discovered independently by Laje et al. (51) and Perroud et al. (45). Using STAR*D data, Laje et al. (51) found that PAPLN rs11628713 SNP (P = 6.2 × 10⁻⁷) was associated with TESI. Perroud et al. (45) conducted a study with similar aims, examining a cohort of 244 participants who experienced an increase in suicidal ideation after taking escitalopram (n = 120) and nortriptyline (n = 124) and 462 control subjects (n = 274 escitalopram, 188 control). rs11143230, found 30 kb downstream of the gene encoding guanine deaminase (GDA), was associated with heightened risk for TESI (P = 8.28 × 10⁻⁷) (45). This result was replicated independently by Menke et al. (95) in the same cohort. These genomic markers demonstrate potential utility in clinical practice, particularly given that treatment-emergent and baseline suicidal ideation are of major public health concern.

Other genetic variants of importance with regard to treatment selection include those coding for enzymes involved in drug metabolism, such as cytochrome P-450 enzymes CYP2D6 and CYP2C19 (96). By understanding the effect of genetic variation on enzymatic activity relevant to antidepressant metabolism, there is a possibility for improving antidepressant response in nonresponders by identifying compounds that modulate the activity of these genes to allow patients to maintain therapeutic ranges of antidepressants. Patients can be grouped based on their CYP2D6 and CYP2C19 genotype: ultrarapid metabolizers (increased enzymatic activity), extensive metabolizers (normal enzymatic activity), intermediate metabolizers (decreased enzymatic activity), and poor metabolizers (no enzymatic activity) (97). Notably, the frequency of these alleles varies across ethnic groups (98). With the aim of determining whether cytochrome genotype can be used to inform patient care, Solomon et al. (96) conducted a systematic review of all related literature published between 2013 and 2018. Sixteen studies were included, with five studies indicating no association between cytochrome genotype and treatment-related findings, two studies reporting mixed results, and nine studies reporting results indicative of cytochrome genotype influencing treatment response, most of which were nonrandomized trials with limited sample sizes (96). Of those that found associations between enzymatic activity and treatment outcomes, the following results were reported. Peñas-Lledó et al. (46) observed, in a cohort of 100 patients with MDD being treated with fluoxetine (SSRI) or amitriptyline (TCA), that all patients with the CYP2D6 ultrarapid metabolizer genotype discontinued treatment within the first 4 wk, whereas no CYP2D6 poor metabolizer discontinued treatment within 12 wk. In another study, Rolla et al. (47) observed a cohort of 47 patients with MDD prescribed varying doses of venlafaxine (SNRI). Of these patients, the only patient belonging to the ultrarapid metabolizer subgroup responded positively to a high dosage (375 mg) of venlafaxine without any adverse drug events (47). Although definitive conclusions cannot be drawn from this study because of limited sample size, it does raise the possibility that ultrarapid metabolizers respond better to high dosages compared to patients with other CPY2D6 genotypes.

Bérard et al. (44) found that pregnant CYP2D6 poor and intermediate metabolizers were four times more likely to discontinue antidepressant usage during pregnancy, relative to extensive and ultrarapid metabolizers. Following a cohort of 78 Chinese patients being treated for panic disorder, He et al. (99) reported that CYP2C19 poor metabolizers were more likely to respond to escitalopram treatment compared with extensive metabolizers between the second and fourth weeks of treatment (P < 0.05). Via meta-analysis, Fabbri et al. (49) reported that, based on data collected from a cohort of 2,558 MDD patients taking citalopram or escitalopram and enrolled in GENDEP, STAR*D, GenPod, and PGRN-AMPS, the poor metabolizer genotype was more strongly associated with reduction in symptoms and with remission, relative to the extensive metabolizer genotype. Jessurun et al. (50) have demonstrated that CYP2D6 activity can be modulated to alter antidepressant metabolite levels. The authors followed 15 patients who were taking nortriptyline and had dangerous levels (>200 μg/L) of 10-hydroxynortriptyline, the antidepressant’s cardiotoxic metabolite produced by CYP2D6, and had been prescribed 5 mg daily paroxetine to lower metabolite concentration, phenoconverting an ultrametabolizer to a poor metabolizer (50). Usually, patients who have this reaction to nortriptyline discontinue therapy because they cannot sustain a therapeutic range of antidepressant and its metabolite is dangerous in high concentrations (50). In all patients 10-hydroxynortriptyline levels decreased, and in 80% of patients 10-hydroxynortriptyline and nortriptyline levels entered their preferred therapeutic ranges (50). The ability to modulate drug metabolism with the addition of low-risk drugs like paroxetine is very promising, but continued testing on larger cohorts must be conducted before clinical implementation.

In a study by Torrellas et al. (100), the use of patient pharmacogenetic profiles (CYP2D6, CYP2C9, CYP2C19, CYP3A4/5 genotypes) to inform antidepressant and dosage selection was shown to result in greater reduction in depressive symptoms (P = 0.002), quantified with the HAM-D scale. Additionally, Torrellas et al. (100) found that 70% of patients who were prescribed antidepressants with the traditional trial-and-error approach were not taking the medication best suited for their genetic profile.

Although GWASs by themselves are difficult to integrate into clinical care, their resultant SNPs can be used to create polygenic risk scoring frameworks and integrated into multimodal tools capable of predicting clinically significant outcomes, from treatment response to drug metabolism. Similarly to blood-based biomarkers and EEG, these results need to be validated. SNPs must be further reproduced before inclusion in a polygenic risk scoring framework, as they were discovered either in a cohort with small sample size or exclusively in one ethnic group. Nevertheless, these studies serve as the first steps toward creating next-generation psychiatric frameworks that can change the way antidepressants and other psychiatric treatment options are prescribed. Although this approach has been most exemplarily demonstrated in MDD, it can be taken for any psychiatric disorder.

Looking Forward

Despite its novelty, the field of precision psychiatry has produced encouraging results. Innovations in computational modeling, genomics, and biomarker discovery are revolutionizing a new vision for diagnosis and treatment selection, generating more objective approaches to psychiatric care. In addition, these advances are improving the scientific community’s understanding of the molecular basis of mental illness. However, despite promising progress in research and development, as of now the majority of the examples included in this review have not been translated into clinically active solutions and the benefits of precision psychiatry are still not experienced by the average patient.

As precision psychiatry becomes the standard of care, the way in which these innovations are developed must be considered to ensure equitable and efficacious outcomes. In particular, computational and machine learning-based approaches to prediction are highly subject to bias based on the composition of study populations. Small, ethnically homogeneous samples are not representative of the psychiatric patient population, yet they serve as the cohorts for many of the mentioned studies. More effort must be made to develop diverse patient samples, including a variety of ages and ethnic groups. An example of such an ideal is the STAR*D cohort, in which thousands of ethnically diverse participants enrolled in a multicenter study. Additionally, although the EEG studies had extraordinarily high accuracies, in the majority of cases they were not evaluated on independent cohorts, and it is unclear whether the models are overfitted to the study samples and thus may not be applicable to the general population. Future research can examine reproducibility of precision psychiatry findings, as they are, by virtue of precision medicine, highly individualized.

One area that has been highlighted as a possible avenue for innovation in fields with highly diverse patient populations, such as psychiatry, has been an integrated research pipeline that utilizes real-world patient data. Experts have highlighted a need to integrate precision medicine research and innovation directly into a development pipeline that involves training prediction models using observational “big data” arising from routine clinical care settings, the development of computerized decision tools, and the subsequent experimental testing of these tools in clinical settings (101, 102). This type of approach helps to overcome some of the primary challenges in psychiatry, namely the highly diverse patient populations, low sample sizes in research settings, and difficulty of conducting large-scale, randomized controlled clinical trials (101, 102). Rather, many advocate for using large prospective, observational trials to emulate randomized controlled trials to generate hypotheses, which then can be independently validated in smaller randomized controlled clinical trials (101, 102). This approach has the added benefit that the outcome of such trials is likely applicable to real-world clinical populations and, by virtue of being developed with a clinically focused pipeline, is feasibly implemented in standard clinical practice.

As we enter this new era of medicine, where algorithms and computational aids are contributing to diagnoses and treatment plans, it is important to recognize the ramifications and interpretability of different machine learning approaches based on how these tools are created. When structured approaches to algorithmic development are utilized, it is possible to create algorithms capable of classifying specific psychiatric conditions, for example, MDD. However, a major limitation is that both diseased and healthy populations are defined by the subjective clinical diagnostic parameters, which reflect the medical community’s working understanding of these conditions rather than true disease states. Unstructured learning approaches and exploratory studies, on the other hand, may not completely agree with current diagnostic categories but can provide insight into the underlying biological realities of mental illness and may lead to the discovery of more objective diagnostic categories of mental illness and/or novel therapeutic targets. Although no one approach is universally best, it will be important for researchers and clinicians to understand the appropriate use and context of different algorithmic tools.

Although all the breakthroughs described in this review are contributing to a new paradigm of objective psychiatric care, the next big shift will come when we stop comparing people to each other and start comparing each person to themselves. Precision medicine has provided many benefits when applied to groups, and, more recently, individualized multi-omic profiling has been able to predict disease earlier, leading to actionable insights well ahead of the standard of care (103). The application of individual continuous monitoring with consumer wearable devices, which have emerging promise in predicting viral disease and other health outcomes such as cardiovascular events (104) and diabetes (105), is an interesting future direction for precision psychiatry and can help untangle dynamic interactions and fluctuations between mental health and physiological states. Through improved understanding of mental health on a physiological level, precision medicine can help break down the barriers between arbitrarily distinct medical specialties and highlight the interplay between mental disorders and biological systems.

That being said, psychiatric patients are not the only group that would benefit from more precise mental health care, as mental illness is known to influence outcomes across medical specialties, including endocrinology, cardiology, oncology, gastroenterology, and pain management. In fact, the World Health Organization predicts that mental illness will be the leading cause of morbidity by 2030 (106). This is not surprising considering that mental illness not only impacts psychological and biological states but also illness behavior and lifestyle, such as diet and exercise, which are known drivers of health and disease. Mental illness is already a major contributor to morbidity and mortality and contributes to all the leading causes of death, including cardiovascular disease, diabetes, cancer, stroke, suicide, and accidents (107–109). Precision psychiatry will bring forth a renewed understanding of the molecular etiology of mental illness, allowing physicians to address both mental and physical health care needs for their patients in a more effective and comprehensive way, ensuring that no patient is left behind.

Conclusions

Precision medicine approaches to the diagnosis and treatment of mental illnesses can overcome the limitations of the outdated model of syndromic disease classification, ushering in a new era of psychiatry. With recent breakthroughs in next-generation multi-omics technologies and data analytics, it is finally possible to map complex, heterogeneous disease mechanisms that were previously elusive. Precision medicine has the potential to reestablish the framework for psychiatric diagnosis and treatment selection as one built upon a thorough understanding of the physiological basis of psychiatric disease, identifying subtypes within existing disease classifications as well as identifying new disease frameworks based on biological data. In the coming years, as precision psychiatry becomes the new standard of care, the shift to include objective markers of disease classification and progression within mental health care will lead to the discovery of more effective treatments and radically improve patient outcomes. In addition, objective markers of mental health may allow for a shift toward optimizing positive outcomes such as thriving, rather than simply treating disease, opening up new possibilities for preventative health and mental fitness within the field of psychiatry. That being said, the emerging innovations described in this review have yet to be translated into clinical practice. The extent to which the average patient experiences improvements in mental health care will be contingent upon the accessibility and clinical feasibility of research findings, in addition to funding agencies and research institutions prioritizing these avenues of investigation. In addition, before clinical translation of precision psychiatry research, rigorous testing in diverse patient populations is needed to ensure that emerging innovations are safely and effectively implemented in an equitable and inclusive fashion. With these considerations in mind, precision medicine will continue to enable targeted approaches to mental health and define a new paradigm of cutting-edge patient care.