Step 1: identification of transdiagnostic genetic biosignatures underlying treatment resistance in SCZ, BD and MDD

We will derive genetically based biosignatures (e.g., PRS) associated with transdiagnostic TR in SCZ, BD, and MDD. This objective will also address the clinical need to identify the common biological basis of TR across these disorders and highlight genetics-based predictors for early identification of TR. These constitute essential components of any future clinical decision support system by providing genetic biomarkers for the early detection of individuals at risk for TR, eventually contributing to a substantial reduction in the trial-and-error period of medication evaluation in each patient.

We will utilize standard, established genome-wide association study (GWAS) and polygenic methods as well as exploiting new approaches in large, harmonized data sets; such as pharmacogenomic allele calling from array data using PyPGx [5]. To this end, the Psych-STRATA consortium brings together the largest collection of treatment resistant individual level genetic data to date, amounting to more than 150,000 cases (90,000 from biobanks).

We will perform within disorder GWAS of TR and refine optimal definitions of TR according to single nucleotide polymorphisms (SNP)-based heritability estimates, genetic correlation of definitions within and across disorders as well as polygenic predictive ability into the independent deeply phenotyped cohorts in step 2. In addition, we will perform cross-disorder GWAS for TR across SCZ, BD, and MDD using standard and cutting-edge methods [6] and will ascertain whether SCZ, BD, and MDD TR risk loci are disorder-specific or act across disorders (TR pleiotropic loci).

We will gain insights into the shared biology of TR by applying statistical fine-mapping to the TR GWAS results as well as downstream analyses at a single cell, tissue, actgene, and pathway level. These models will be used to pinpoint genes, pathways, cell types, and tissues associated with TR using transcriptome and pathway wide association analysis [7]. Methodically we will identify specific genes and biological pathways relevant to TR across disorders through gene prioritization methods, which combine state-of-the-art statistical fine-mapping [8] and co-localization of expression quantitative loci (eQTLs) [8], gene-set enrichment analysis framework [9] as well as functional mapping of with gene-set analysis (GSA)-MiXeR [10] within and across disorders.

In this step, we will also perform PRS analysis [11, 12] to assess liability to TR across all cohorts for (1) within-disorder TR (case/case) for SCZ, BD, MDD and (2) combined TR (case/cases) across diagnoses. We will leverage emerging methods (e.g., polygenic risk score–continuous shrinkage (PRS-CS), linkage disequilibrium (LDPRED), PRS of CHD-associated biomarkers (BioPRS)) [13] to identify the most powerful approach for assigning a genetic based biosignature risk score for TR across disorders and additionally consider the potential role of pharmacogenomic markers in mediating or moderating polygenic effects [14, 15] along with polygenic overlap with metabolic dysfunction [16]. We expect that PRS calculations will explain rather small but nevertheless significant proportions of the cohort variance. We will perform transcription (TWAS) and pathway activity level (PALAS) genome wide association studies for TR liability within and across disorders and leverage the results to compute tissue specific and multi-tissue gene and pathway level risk scores for each individual [12]. Additionally, recently reported analysis on polygenic liability for antipsychotic dosage and polypharmacy will be implemented [17].

Finally, we will evaluate distinct biosignature scores (PRS, TWAS-PRS, etc.) with respect to their discriminatory power of TR vs non-TR patients individually as well as in combination. Figure 3 gives an overview about the step 1 analysis strategy.

Fig. 3

Overview of the analysis strategy of step 1

Step 2: identification of multi-modal biosignatures profiles associated with treatment outcome

In this step, we will first derive multi-omics biosignatures and biomarkers associated with treatment outcome and TR. We will conduct proteomic analyses on longitudinal cohorts of SCZ, BD, and MDD patients to assess treatment effects on the serum proteome. Using a multi-stage approach, we will examine over 5400 proteins in ~ 900 serum samples from TR/non-TR individuals during the discovery phase, leveraging the Olink Explorer HT panel [18]. This will elucidate proteomic changes linked to different treatment responses. Subsequently, we will establish a thoroughly selected candidate protein panel of up to 380 proteins for the validation phase. We plan to use the commercially available kits with 5000 proteins. This will consist of a high-throughput antibody-based suspension bead-array assay utilizing antibodies from the Human Protein Atlas, across approximately 6000 samples of TR/non-TR individuals with SCZ, BD, and MDD.

In the described datasets, we will aim to identify proteomic biomarkers associated with (1) response to specific treatments, depending also on sample size considerations, (2) response irrespective of treatment type, and (3) TR vs response for all patients and those with two measurement time points [19]. We will consider the use of both linear and non-linear methods for detecting possible associations and the use of strategies to limit multiple-testing burden, and validate the results of genetic analyses (Step 1). In detail, we will consider the following: (1) genes with evidence for association with TR based on TWAS analysis, (2) membership or interaction with TR associated pathways, (3) classification as druggable gene (e.g., Drugbank, https://go.drugbank.com/), and (4) proteomics-level pathway analysis. With the aim to integrate the different layers of data, we will also integrate genetic and proteomic data to identify protein quantitative trait loci (pQTL) associated with treatment outcomes.

These insights will then be combined with clinical and psychosocial variables to develop multi-modal predictors that integrate the different key components contributing to treatment outcomes. For this scope, we will evaluate different machine learning and integration strategies (e.g., multimodal neural networks) [20] to integrate *omics and clinical predictors (e.g., socio-demographics, severity and cause of illness, comorbidities, psychosocial function, clinical digital markers on sleep and activity).

Finally, we will leverage personalized disease models based on induced pluripotent stem cells (iPSCs) from TR/non-TR patients to validate the identified biomarkers/biosignatures [21], identify new biomarkers, and dissect the neurobiological mechanisms contributing to treatment outcomes. RNA-Seq and metabolome profiling of iPSC under baseline and exposure to first line drug treatments considered in Step 3 will be assessed. Finally, these results will be integrated with omics results of the previous steps, to pinpoint mutually validate the TR associated molecular changes/biosignatures.

Step 3: randomized controlled trials on efficacy and safety of early intensified pharmacological treatment

The main objective of this step will be to conduct an innovative RCT for MDD, SCZ, and BD patients (strategy in Fig. 4). The trial will be conducted as an open-label study, with the assessors of the primary outcome measures blinded to treatment. Patients, whose illness did not respond to a first-line therapy in the current disease episode, will be randomized (1:1 randomization) to early intensified pharmacological treatment (EIPT) or treatment as usual (TAU). The treatment phase lasts 6 weeks, but it is an intent-to-treat approach, which means that patients might stop treatment earlier or could be treated longer. Follow-up is carried out up to 6 weeks after the end of the treatment phase.

Fig. 4

Overview of the randomized controlled treatment (RCT) strategy

EIPT will consist of a pharmacological treatment that is currently recommended for patients with TR by treatment guidelines (i.e., patients who did not sufficiently respond to ≥ 2 treatments). In MDD, a switch to another oral antidepressant plus esketamine (EIPT) will be compared with a switch to a second line antidepressant (TAU). In SCZ, a switch to clozapine (EIPT) will be compared with a switch to a second-line antipsychotic (TAU). For BD, a switch to one of the following will be compared: 1. escitalopram, sertraline, or venlafaxine plus 2. two of the following: lithium, lamotrigine, valproate, or quetiapine (EIPT) versus quetiapine plus lithium or valproate, or lamotrigine (TAU).

Treatment outcomes will be measured over a period of six weeks. The primary endpoint is the efficacy so called change in symptom severity (according to standard scales); the secondary endpoints are cognition and multiple patient-centered outcomes such as quality of life, general functioning, and life satisfaction. Side effect profiles and long-term effects will also be assessed until week 12 after treatment. The total follow-up period will be three months, comprising five clinical study appointments; the baseline visit will follow within a week after the screening, and visits 3, 4 and 5 will be conducted at 2 weeks, 6 weeks, and 12 weeks post-baseline, respectively.

In addition to a comprehensive medical history, and numerous clinical self-assessment, and clinician assessment instruments, there are three special features of this RCT, as we will include a comprehensive assessment of cognition, digital markers, and biomarkers (blood and feces). Genomic and proteomic analyses will be performed for integration/validation of steps 1 and 2 (see the next paragraph). The digital biomarkers will be collected using electronic diaries, mobile sensor technology, and wearables, in some cases via smartphone.

The efficacy of the medication will be measured through the Positive and Negative Symptom Scale (PANSS, SCZ) and Montgomery Åsberg Depression Rating Scale (MADRS, MDD, and BD) and compared between both arms of the trial. The safety will be measured through General Assessment of Side Effects (GASE). A further aim of this step is to identify clinical and digital psychological predictors of treatment response/TR. Other key data of interest are cognitive phenotypes potentially associated with treatment response, and the digital mental health data will be analyzed to identify digital measures associated with treatment outcomes [22, 23].

Step 4: identification of multi-modal signatures associated with TAU and early intensified treatment response for evidence-based treatment recommendations in clinical practice

This step will integrate and extend the results of step 1–step 3. First, we will leverage the multi-modal data from the RCT in step 3 to validate identified genes, pathways, proteins, pharmacogenomic alleles, and genetic scores predictive of treatment outcome based on steps 1 and 2. More specifically, we will validate biosignature based biomarkers (genetic and proteomic) associated with TR liability and treatment outcome in SCZ, BD, and MDD. We will compute PRS and genetic/multi-omic based biosignatures defined in step 1 and 2, respectively, for all RCT patients at baseline. We will also perform multi-omic integration to validate biomarker candidates from joint step 1, step 2, and RCT cohorts across molecular layers and explore the predictive capacity of multi-omic features with respect to treatment outcome.

Second, we will conduct a further biomarker discovery effort based on the data collected in step 3 to identify transcriptomic or proteomic biomarkers based on their capacity to classify SCZ, BD, and MDD patients into responder/non-responder groups. We will target genes and serum proteins that are associated with 1. response to TAU, 2. response to EIPT, and 3. biomarkers predicting differential response to TAU vs EIPT in each disorder and transdiagnostically. In a complementary effort, we will connect the distinct molecular layers and also identify expression and protein quantitative trait loci (eQTLs and pQTLs) under baseline and post-treatment conditions [24] to pinpoint eQTLs/pQTLs associated with treatment response.

Finally, we will build on the validated biosignatures from step 1 and 2 as well as the latter results and establish a multi-modal machine-learning model to predict treatment outcome for TAU or EIPT. To this end, we will extend the multi-model machine-learning model developed in step 2 and integrate genetic, transcriptomic, and proteomic profiling of steps 1–3 with deep clinical phenotyping data to build a multi-layered treatment response model for TAU/EIPT. More specifically, we will combine the following data modalities and evaluate their contribution to overall prediction performance:

genetic based scores and/or pharmacogenomic alleles associated with TR found in step 1,

omics based biomarker scores for TR found in step 2,

omics based individual and aggregated biomarker scores of response to TAU vs EIPT,

clinical, psychosocial, and neurobiological measures (symptom rating, cognition, etc.), and

digital mental health features.

These data modalities will be integrated using different early (e.g., canonical correlation analysis, multi-modal autoencoders), intermediate (e.g., multi-modal neural networks) and late (e.g., stacking) integration strategies for multi-modal data. We will evaluate candidate features with respect to added value via nested-cross validation designs.

In addition, patient trajectory modeling and simulation will be implemented in this step, using e.g., the previously published Variational Autoencoder Modular Bayesian Networks (VAMBN) and/or MultiNODE algorithms [25, 26]. We will evaluate statistical dependencies between biomarkers, clinical outcomes, and digital assessments and generate virtual patients to simulate different treatment scenarios. Thus, this step will provide the data and analysis groundwork for the evidence-based treatment recommendation component of the mental-health board in clinical practice (step 5).

Step 5: development of a treatment decision-making mental health board

In this step, we will focus on three main objectives. The first one is the implementation of a software tool embedding AI/ML models developed in step 4 for treatment decision support. This software will help clinicians to obtain predictions of treatment outcome for individual patient using their specific characteristics (multi-omic, clinical, and digital data). Model predictions will be explained via methods such as Shapley Additive Explanations (SHAP). Moreover, we will embed interactive features, e.g., mapping of omics-derived features to molecular networks. For example, these networks can include known drug-target interactions, retrieved from public databases (e.g. Drug Bank), hence pointing to potential alternative treatments. Finally, we will evaluate this software tool regarding its technical correctness and its clinical utility.

The second objective is the creation of a decision-making mental health board, which aim will be to conceptualize and facilitate a shared decision-making (SDM) process between patients, carers, clinicians, and scientists. We will realize this in three phases. The first will be the development of a shared decision-making platform, which we will evaluate by running focus groups and surveys. We will also analyze and evaluate the potential impact and obstacles of implementing biosignature-assisted SDM processes. Another important goal is the implementation of the decision-making mental health board. It is anticipated that this board will consist of an interdisciplinary group of scientists, clinicians, primary care physicians, psychotherapists, patients, and carers. We will conduct a pilot study utilizing the decision-making mental health board. The outcomes of this pilot study will provide first experience of the mental health board in national and international clinical settings.

The last objective of this step is the creation of a framework for implementation of treatment guidelines. For this, we aim to develop definitive, prospective level evidence [27]. We will convene an expert working group (clinicians, researchers, and patient representatives) to review the existing evidence [28, 29] and submit a consensus report to the major international guidelines. We will also prepare a publication.

Step 6: patient empowerment, dissemination, and education

The last step of Psych-STRATA focuses on patient empowerment, dissemination, and education in different ways. First, we will generate opinions maps at the European level regarding the use of digital apps and continuous mental health assessments among relevant stakeholders. We will integrate app-based surveys in order to collect participant feedback from the RCT (step 3). Likewise, we will develop surveys on the same questions for other stakeholders: participants in other clinical trials, relatives, clinicians, researchers, policy makers, and any other relevant group of individuals.

Survey participants will be recruited via channels such as the GAMIAN-Europe’s website, touching base with local contacts of each patient group embedded in GAMIAN-Europe, presentations at local and international meetings, newsletters, and social media. Another important objective will be the comprehensive evaluation of the novel mental health board with shared decision-making (step 5), using a structured survey that will target relevant stakeholders.

The last objective is the dissemination of findings and education. The World Psychiatric Association (WPA) will be an important partner in the dissemination and education part (workshops, surveys, connecting with groups around the globe). Part of this is setting up and maintaining the official Psych-STRATA website (https://psych-strata.eu/), Psych-STRATA Twitter/X account (@psych_strata), and a LinkedIn account. We are currently creating and updating communication information material, developing project factsheets, posters, official presentations, and all types of dissemination materials including newsletter (https://psych-strata.eu/newsletter/). Moreover, we will organize workshops for healthcare providers and disseminate our expert consensus document. Lastly, press releases will be issued when significant results have been achieved to reach and inform specialized and/or generalist media on key activities and results.

Step 7: in silico drug repurposing for identifying potential therapeutics in SCZ, MDD, and BD and development of the Psych-STRATA software as a service (SaaS) platform

This step represents a transformative approach to in silico drug repurposing, integrating advanced computational techniques such as cheminformatics, text mining, network pharmacology, omics data analysis, and network-based drug repurposing methods. By adopting these methods, Psych-STRATA aims to introduce a rational approach to drug identification, improving the precision in identifying potential drug candidates for conditions like SCZ, MDD, and BD [30]. Unlike traditional trial-and-error methods, computational simulations and cheminformatics tools enable the prediction of drug efficacy, bioavailability, and potential side effects before clinical trials, thereby streamlining the drug development process [31]. Moreover, network-based techniques consider the complex interplay of biological pathways involved in mental disorders [32], which is crucial for conditions like SCZ, MDD, and BD that often share common biological pathways. By mapping these pathways and identifying key nodes or targets, Psych-STRATA can suggest existing drugs effective across multiple conditions, aligning with the emerging paradigm of poly-pharmacology [33]. These methodologies enable the exploration of latent linkages between different mental disorders, facilitating drug repurposing by identifying drugs effective for conditions not initially targeted.

Moreover, the Psych-STRATA SaaS platform represents a significant methodological advancement, offering a dedicated platform for seamless integration of ML/AI into healthcare providers’ daily decision-making processes [34]. This platform aims to provide clinicians with more personalized, actionable data, enabling real-time, data-driven decision-making in mental health care, which is often lacking in current systems. By offering patient-specific insights, particularly crucial for complex mental health disorders like SCZ, MDD, and BD, the platform enhances the quality of care while reducing the burden on healthcare systems. The integration of advanced ML/AI models into clinical practice streamlines decision-making, making diagnosis and treatment planning quicker and more efficient. Ultimately, by offering personalized data and enabling real-time decision-making, the Psych-STRATA platform has the potential to revolutionize mental health care provision, setting new standards in the field.