Several models have been proposed to represent human genomic information. An interesting approach for supporting genomic applications for health consists of a two-layer representation. In this approach, high-level concepts describing distinct aspects of the human genome at an abstract level are mapped to data representing actual physical measurements. This two-layer method allows users to formulate high-level queries on the concepts and map them onto real datasets. Additionally, the approach is extensible, allowing new conceptual views corresponding to specific genomic features to be mapped to the lower data layer without impacting previous mappings.
CEUR ceur-ws.org
The Human Genome, with its vast complexity, presents challenges in capturing, representing, and utilizing its extensive information; consequently, the landscape of genomic data sources is wide and diverse. Commonly used databases include The Cancer Genome Atlas, a landmark cancer genomics program now embedded within Genomic Data Comm1o]n; st[he 1000 Genomes Project2[], a catalog of common human genetic variation; G3T]E,ax r[esource database to study the relationship between genetic variation and gene expression in multiple reference tissues; and GEO4[], the most general and widely used among genomic repositories.
In the genomics domain, conceptual models have long been employed to efectively manage and represent extensive data, as well as to accurately depict the structure and functions of
CEUR Workshop Proceedings the genome. Starting in the late nineties, pioneers such as Okayama5]evteanlt.u[red into representing DNA genomic sequences in databases. In the 2000s, Paton6e]tinatl.r[oduced data models for transcription/translation processes, alongside genomic sequences and protein structures. Subsequent works leveraged conceptual models to articulate biological entities and interactions, leading to databases like GenMapper Ware7h]oaunsde B[ioMart8[].
This background research has later motivated conceptual modeling-based approaches focusing on either characterizing the genome’s structure conce9p]touralalpyp[lying it in data-driven contexts1[0, 11]. Bridging these perspectives emerged as a pertinent issue, more recently addressed in 1[2]. In their proposal, the authors describe a novel conceptual model that merges concepts-based and data-based perspectives for genomic information modeling. Specifically, they link aconcepts-layer delineating genome elements and their connectiondsattao-layer, detailing real-world datasets from genome sequencing. This dynamic linkage facilitates focused visualization, understanding of commonalities, and complex query expression across genomic data types, expanding the modular view-based approach to genomic data management.
This work focuses on the perspective of data users who frequently need to access and query genomic data resources. Genomic data practitioners typically perform similar types of queries repeatedly. Currently, there are systems that allow for basic data extraction using simple queries (conjunctive/disjunctive) over data. Examples of such systems in1]clfuodresi[ngle consortia databases an13d][and [14] for integrated databases. However, while basic queries are supported, more sophisticated approaches tailored for more advanced data analysis purposes are still lacking.
We propose apattern-driven approach to bridge the existing gap, facilitating more complex and specific data analysis tasks. This approach largely leverages the conceptual linking provided by the two-layer conceptual model describe1d2]in,s[erving as a foundation for generating these query patterns over paramount genomic data sources. The efectiveness of this approach is demonstrated through the instantiation of query patterns that yield significant results in contemporary clinical and genetic research. These patterns include the extraction of datasets for genetic case-control stud1i5e,s16[, 17], integrative multi-omics analys1e8s, 1[9, 20, 21], and family trio analyse2s2[, 23]. Our proposal aims to ofer a flexible and expandable representation of concepts, data, and their typical interconnections, providing simple query templates for guiding concept exploration, inspiring the identification of novel correspondences among data, and enhancing the findability of interoperable data instances.
In the remainder of the manuscript, Sec2tipornovides notions on the two-layer conceptual model; Section3 describes our core contribution, i.e., the data analysis patterns and several example instantiations; Sect4iodniscusses the implications and limitations of the approach; and Section5 concludes.
Our work takes inspiration from the holistic view presen1t2e]d,winh[ich bridges a model of the genomic concepts and a model of the genomic datasets to facilitate genome data management through robust conceptual modeling support. More specifically, we consider a two-layer conceptual model: 1) the “concepts-layer” encapsulates human genome mechanism knowledge; and 2) the “data-layer” portrays genomic data types and experiments through structured information formats. The abstract idea can be appreciated in F1i:gguerneomic information can be viewed as a dual system approached in opposite directions: connecting data to pre-existing abstract concepts (top-down) or building concepts based on available data (bottom-up).
A top-down approach initially models biological entities and then verifies data sources; this direction allows us to reveal issues with data structure definition and quality. Conversely, a bottom-up strategy starts from available data and subsequently constructs models to systematize and organize it, aiming to create user-friendly systems for domain experts.
Concepts layer
Gene expression view
Gene
DNA methylation view
CpGisland Data layer microRNA expression view miRNA
ExpTeyrpimeent 1
1* 1..* Donor 1 * Biosample 1 * Sample 1
1* Project
ExpressionLevel *1 Dataset
1* Schema
1* 1..* SRaemgipolne
DNA variation view
Variant Reading
The data-layer (depicted in blue in in Fig1u),rceenters on thSeample concept, representing a typical genomic data file, which contains a seStaomfpleRegions, i.e., rows in the file, which represent an interval of the genome on a specific chromosome strand, with start and end coordinates. Multiple samples are collected wDiatthaisnets, which are homogeneous in theSchema (i.e., their sample regions have the same columns and semantics) and in the ExperimentType (a description of the experimental assay run to produce the data). The experiment has been performed on biological material, which is describedBiboyStahmeple class, which belongs in turn tDooanor (an actual living patient tissue or an immortalized cell line or single cells that have undergone a sequencing process). Samples are grouped within Projects (informing on the management metadata information).
The concepts-layer has diferent modules (voierws, depicted as light blue rectangles in Figure1) describing aspects of the human genome, such as DNA variation, gene or microRNA (miRNA) expression quantification, DNA methylation levels, or any other genomic data type. To each experiment type in the data-layer, we associate aggenivomenic data view (see light blue arrows). Each view includes classes representing concepts that are measurable through genomic sequencing technologies (e.g., the expression levels of genes or the reading of a DNA variation). In Figur1e,these concepts are drawn in red; given concepts could be common to diferent views (e.g., the “expression level”).
The concepts-layer and the data-layer are linked through relationships between concepts (such as a variation in the DNA) and instances of data-layer classes (i.e., a specific data record). For example, aSampleRegion from a DNA-Seq experiment can be represented by its corresponding concept, aVariant spanning positions 43,044,295 to 43,170,245 on the negative strand of chromosome 17.
New links between the concepts and data-layers can be established when specific data types are selected (in the data-layer), thereby triggering the selection of specific views (of the concepts-layer). Through a classical Ontology-Based Data Access ap2p4r]o,aitc his[possible to allow access to datasets of a specific genomic data type by specifying a query on the view of interrelated concepts.
While the data-layer is static because new genomic data types are simply another instance of the related entities, the concepts-layer is flexible and can grow according to the specific needs of a use case. Figur2eillustrates a portion of the concepts-layer model containing classes associated with DNA variations, familial relationships, gene expression, miRNA expression, and DNA methylation. located_in
1..* ChromosomeElement located_in ElementPosition --ndaesmcerip:tsiotrnin:gstring 1 *
In this model, thIendividual is the primary class, representing a person. Individuals can be classified as aHealthyIndividual or anUnhealthyIndividual based on their diagnosis of a specific Disease. It is possible to establish familial links among individuals (FamiliarRelationship); individuals aggregateGirnoupOfIndividuals, such asFamily. These aggregations are modeled with the purpose, for instance, of exploring the interaction between DNA variations and diseases within families; this aspect is crucial for determining patterns of inheritance and the pathogenicity of variants. Individuals are composedLocfamtaionnys, such asTissues.
Diferent Measurements are performed on individuals. Two types of measurements are captured in the model, as relevant to our portioRne. aTdhineg describes the appearance of a DNA variation (oVrariant) in an individual. Variants are distinguished by a name and description, a type (substitution, insertion, or deletion), and a set of alleles (i.e., reference, alternative, and ancestral). Each variant can have multiple pVoasriitainontPso(sition), each determined according to a specific reference system, also knowAnssaesmbly. Variants are crucial in understanding the genetic basis of many diseases.
The second type of measurement is tEhxepressionLevel, which is always related to a genetic component. Unlike the previously mentioned type of measurement (i.e., readings), the expression level is specific for a giveTnissue, with significant diferences among tissues. Three are the diferent expression levels considered in the excerpt: • GeneExpression: a biological process that ensures coGrreencest(which areTranscriptableElements) are expressed at the right time and in appropriate amounts, enabling cells to perform their functions correctly. Gene expression measurement can help identify diferentially expressed genes between normal and cancerous tissues. • miRNAExpression: a biological process associated with biological components that regulate gene expression. miRNA expression measurements capture the lemvieRlNsAof, a kind of non-coding RNAnc(RNA), corresponding toMaatureTranscript (as opposed to genes, which are a kindPorfimaryTranscript). Measurement of miRNA levels allows for a better understanding of cancer development and progression, providing insight into the regulatory mechanisms underlying cancer. • DNAMethylation: a biological process altering gene expression, happening in correspondence withCpGIslands, particular featured regionCshionmosomes. Measuring DNA methylation is crucial to understanding how environmental factors afect -for instance- cancer development and progression.
As suggested in1[2], the two-layer representation –drawing direct linking between concepts
and data in the genomic domain– allows:
(a) real data inspection improving its conceptual representation (e.g., by identifying cases
where many diferent variant positions exist from chromosome elements or variants);
(b) use of abstract knowledge (i.e., concepts-layer) as an extractor of existing datasets (i.e.,
data-layer), for instance, by leveraging the explicit conceptual relation between positions
and elements (including genes and transcripts); and
(c) formulation of inter-data type queries over data (i.e., coming from Mdifearseunrtement
types), by controlling the datasets’ concepts that regard diferent genomic data types.
These aspects could be translated into simple view-driven queries, where concepts are selected
in the upper layer and are translated into queries over the data. Here, we propose to make a
step forward with respect to (a)–(c): we use conceptual linking as a glue for generating classical
genomic analysis patterns that are typically used in research practice. In this section, we describe
the most relevant ones, selected according to the enduring experience of the authors in the
ifeld, developed during several interdisciplinary collaborations with clinicians, biologists, and
geneticists. In this work, we focus on:
(
Finally, we show how patterns can be combined in complex patterns, e.g., joining the approach
described in (
Case-control studies constitute a commonplace method in clinical research aimed at comparing diverse genomic datasets from ill individcuaasles)(and unafected individualsco(ntrols) to delineate genetic elements that contribute to increased susceptibility or severity to disease. Publicly accessible data repositories such as The Cancer Genome Atlas (1T]C)fGoAr,c[ases and The GTEx Consortium3[] for controls are fundamental for increasing sample sizes and identify cases and controls in scenarios where they were previously unavailable, thereby enhancing the eficiency and robustness of genomic studies. Two types of case-control analyses are typically produced: 1. At thepopulation-level, examining cases and controls using data that is specific to a particular tissue, with the purpose of investigating how diseases or phenotypic traits afect that tissue. 2. At thepatient-level, analyzing both healthy and diseased tissue samples extracted from the same patient to determine the impact of cancer processes on a given tissue. Population-level case-control. The advantages of population-level case-control analyses focusing on the same tissue type have been extensively described in the literature. In the concrete domain of cancer genomics, such studies typically involve comparing healthy tissues (derived from patients without cancer) of the same tissue type as those giving rise to cancer (obtained from patients with a specific cancer subtype).
For instance, Araent al. [15] combined data from the TCGA and GTEx projects to analyze gene expression disparities between healthy and across eight tissues and corresponding tumor types. This approach facilitated the comprehension of tumor development and the discovery of novel biomarkers, critical for efective prevention and therapeutic stratagem selection.
In Figure3 we present a simplified version of the scenario outline1d5i]n,f[eaturing two distinct patients: one aflicted (extracted from TCGA, with ID = “TCGA-A2-A04N”) and one healthy (extracted from GTEx, with ID = “074b0792-df3c-4b59-9f50-793bc14bcb81”) individual.
Note that the aflicted patient has a non-healthy biosaimsp_lheea(lthy = false) associated with theDuctal and Lobular Neoplasm disease.
The process of selecting cases and controls with specific conditions and from the same tissue type within such datasets is nontrivial and necessitates sophisticated instance modeling. For example, identifying patients who are “male, white, and 79 years old” within1T]CisGnAot[ feasible. Consequently, pairing cases and controls entail not only ontological mediation (via the concepts-layer) but also an understanding of the data sources. Our proposed approach streamlines this process by enabling a consistent representation of biological concepts and a technologically-independent data representation. The expression levels are observed on a specific gene (TP53), which is fixed at the conceptual level and therefore searched in the SampleRegions of GTEx and TCGASamples to extract appropriate values.
Patient level case-control. Numerous studies have highlighted the clinical advantages of performing case-control analyses at the patient level. For individual patients, the analysis involves comparing samples from adjacent tumor tissue (considecroendtroals) and tumor tissue (case). Collectively, these two types of samples are referrepdatiroedassamples.
In [16], Kim et al. analyzed paired samples from patients wCiotlhon Adenocarcinoma, showing that this type of analysis significantly impacts the prediction of cancer recurrence. Oh and Lee [17], instead, examined the diferences in gene expression between paired samplLeusnign Adenocarcinoma andBreast Invasive Carcinoma, among others. Using machine learning models, they concluded that such analyses can aid in predicting the prognosis of certain cancers, thus facilitating appropriate clinical treatments. Both studies employed the TCGA public resource to obtain patient data. In Fig4uwree show the case of a patient possibly included in this patient-level case-control anal1y7s]i.sT[his patient, diagnosed with Lung Adenocarcinoma (id = “TCGA-44-6146”), holds paired samples available in TCGA.
TCGA is a repository that provides information on files resulting from specific genomic analyses of healthy and tumor tissues from cancer patients. The analytical nature of TCGA makes the search for paired samples challenging, requiring advanced data processing and search functionality to identify analyses from the same patient (e.g., files associated with the same Donor). Currently, this is not allowed even by the updated TCGA major entr1]y. point [ Note that, in the concepts-data framework, the data-layer enables easy identification of the patient (oDronor) from whom each sample originates, while the concepts-layer facilitates the identification of both samples as belonging to the same individual.
Multi-omics approaches are innovative frameworks that integrate multiple omics datasets to enhance understanding of genetic dise2a6s]e. P[articular attention is given to multi-omics studies to study cancer’s molecular and clinical features. Here, areas of research include segmentation into subtypes, improvement of survival predictions and therapeutic outcomes, and uncovering key pathophysiological processes across diferent molecular layers. Again, we refer to the TCGA data source, while other cancer genomics could also be considered (see the ICGC [27]). For multi-omics analysis, two types of inquiries typically hold significant interest: 1. At thepopulation-level, analyzing a specific disease or phenotypic trait, by using genomic samples that refer to diferent genomic data types (i.e., features in the genome). 2. At thepatient-level, analyzing a specific disease or phenotypic trait, considering specific patients whose samples have been analyzed according to multiple genomic data tests (i.e., for whom multiple data types are available).
Population-level multi-omics. Associating variation signatures or gene expression/methylation/miRNA profiles with diagnostic/prognostic values is of high importance in cancer research. Pinoliet al. [18] examine the rich presence of variants, abnormal methylation levels, as well as copy number alteration events, in the proximity of specific topological structures for 26 cancer types. Mehrgou and Teimourian19[] utilize gene expression, methylation, and miRNA datasets from both TCGA and GEO sources to derive insightCsoolonrectal cancer, with applications in diagnosis, prognosis, and targeted therapy.
Patient-level multi-omics. Focusing on specific tissues, we aim to find patients whose biological samples have been analyzed using diferent genomic experiments (i.e., for whom multiple data types are available). Grouping data by the same patient enables building richer disease models. We call this pattoenren-to-one linking and to the multiple samples derived from the same patient lainsked multi-omics samples (connecting mutations, expression, and epigenomic signals such as methylation levels).
Figure5 illustrates a patient (ID = “TCGA-33-4589”) with lung adenocarcinoma for whom data on variants, methylation levels, miRNA, and gene expression are available in TCGA. The comprehensive data of this patient facilitates multiple analyses with significant clinical applications. For instance, 2in0][, miRNA and gene expression data were analyzed in patients with lung adenocarcinoma to classify patients based on survival. This has crucial implications for cancer prognosis, enabling the identification of patients who may require more intensive monitoring due to a poor prognosis. Similar studies have been conducted for survival prediction in breast cancer, utilizing gene and miRNA expression, DNA methylation, and CN2V1]d. ata [
Rare disorders are conditions with a low frequency in the population and often have a genetic component. Despite the significance of genetics in these disorders, most patients remain undiagnosed after standard genetic test8in].gF[amily trio analysis involves comparing the genetic information of the patient with that of their parents. Consequently, it is possible to identifdye novo variations, i.e., DNA variations unique to the patient and not inherited from either parent. This kind of analysis has been shown to positively impact rare disease contexts, by improving diagnostic and serving as a powerful tool in identifying disorder-causing variations23[].
Figure6 illustrates information about a family trio repo2r2t]e.dIninth[is study, the authors examine family trios in the contexAtmoyfotrophic Lateral Sclerosis (ALS) to identify risk factor variants associated with this devastating disease. They identifieddseenvoevroavlariations, such as the v1 instance of the Variant class in F6i.gTuhries variant is considerdeednovo because it was identified only in the afected individual (son instance oUfnthealthyIndividual class) and not in either parent (father and mother instances in the concepts-layer). The identified variants helped the authors improve the understanding of the genetic role in ALS.
In this particular pattern, the data-layer represents fundamental information about the experiment, such as the sequencing technology, which cannot be represented in the conceptslayer. On the other hand, the concepts-layer allows us to infer that the variant shown in Figure6 is de novo, as it captures the familial relationships between individuals. The ontological connection between both models provides a holistic representation of all the relevant information needed for family trio analysis.
An important genomic data source, the 1000 Genomes Project, collected a huge dataset intending to identify all the genetic variants with frequencies of at least 1% in several worldwide populations. The last release of the project covered 26 populations and observed single nucleotide variants (SNVs) and insertions/deletions (indels) from diferent 602 parent/child trios produced within the projec28t][. The pattern described in Figu6rcean be reproduced on 1000 Genomes data to perform a database-wide analysfiasmoinly-trio samples.
Above, we have demonstrated traditional data analysis patterns in genomics. However, more complex analysis patterns have gained attention.
One example of a complex pattern involves family-trio case-control analysis, which corresponds to performing case-control analyses (see Se3c.1t)ioonn family trios (see Section 3.3). Specifically, it compares the genetic information of families with an afected individual to families with no afected individuals.2I9n],[this strategy was employed to identify genes and mutation types that are highly associated with Schizophrenia7. iFlilguusrterates that scenario, by using only two families (for simplification purposes). Here, it can be observed that the v2Variant appears in the ofspring of both families; this insight can be used to rule out the association of this variation with schizophrenia (as sU1nihseaanlthyIndividual, whereas s2 is aHealthyIndividual). Conversely, the vV1ariant appears only in the family member s1, who is afected by schizophrenia; however, it does not appear in any of her/his parents, which ofers strong evidence of the potential relationship between variant v1 and schizophrenia.
Another example of a complex pattern is multi-omics case-control analyses. Here, experts compare diferent data types from patients with a certain characteristic (cases) to those without it (controls) to determine if there is a clinically relevant relationship between any omic feature and the characteristic under study. For instan3c0e],, tinhe[ authors used multi-omic data to predict the risk of developing asthma, an3d1i]n,t[hey employed this type of analysis to predict the development of preeclampsia.
The two-layer framework described in Sec2tiaolnlows us to incorporate new concepts and relationships according to a data-agnostic approach. Indeed, as acquired knowledge in genomics is constantly evolving, new concepts will be added and changes will recur in the conceptslayer. The model will not remain the same as the one presented in2F,iwguhriceh -in turnextends previous work12[]. Conversely, the data-layer is typically not impacted by genomic concepts’ changes as long as all genomic data can be represenStaemdplaess containing SampleRegions. Even when experts’ understanding of genomic-related knowledge mutates, possibly impacting the interpretation of data analysis results, data keeps the same model; this favors the maintainability of potential data mappings, processing pipelines, and bio-tools that leverage this representation.
Here, we show that a strong connection between data and concepts compensates for the limitations of approaches that consider the layers separately, allowing a holistic representation of the genomic domain. The identification of interesting patterns of analysis and the consequent reasoning can only be explained by using an interactive two-layer representation. Our rationale is to use the conceptual linking as a glue for generating classical patterns of case-controls, multiomics, or family-trios by having the conceptuals-layer model in the middle and instantiating the data-layer model as many times as needed. At the patient-level, case-controls typically have two-instance-replication (see Fig4u).rIenstead, multi-omics have many-instance-replication; we showed four, in the example of Figu5,rreeplicating the data model forgfeonuormic data types thereby creating one copy for each sample of a same patient. In this way, we let classical genomic data analysis patterns emerge, where we “pivot” upon ontological knowledge (conceptslayer) as the mediator across several instantiations of the data-layer, playing clearly identified roles. We demostrated the capability to perform queries with high complexity, which facilitates the extraction of relevant data from highly-heterogeneus disorganized repositories and the advancement of data exploitation in the domain.
In this paper, we explain five basic patterns and then hint at how they can be composed to form more complex patterns. This framework will allow easy extension to novel query patterns that will go along with the rapidly-evolving state of genomic knowledge In current practice, domain experts typically navigate genomic data source interfaces and download data without a clear formalization of the underlying concepts and their semantic relationships. In this work, we describe a conceptual modeling-based framework that enables a unified querying strategy. Building on this, we envision a next-generation genomic data query builder that, starting from high-level concepts, allows users to execute abstract queries. This approach will relieve practitioners from the complexities of data formats and heterogeneity, enabling them to seamlessly formulate data extractions that align more closely with the classical problem formulations they are familiar with. The patterns presented in this work demonstrate initial prototypes of modular queries that can be implemented in such a system. Prospectively, this data query builder will be the main component of a visual model-driven query system for practitioners, where the conceptual model in the concepts layer is used to identify data instances in the underlying data layer.
Acknowledgements. This work is supported in part by CIPROM/2021/023), PDC2021-121243I00), PID2021-123824OB-I00, MICIN/AEI/10.13039/501100011033, and ACIF/2021/117 grants. [12] A. Bernasconi, et al., PoliViews: A comprehensive and modular approach to the conceptual modeling of genomic data, Data Knowl Eng 147 (2023) 102201. [13] F. Albrecht, et al., DeepBlue epigenomic data server: programmatic data retrieval and analysis of epigenome region sets, Nucleic Acids Res 44 (2016) W581–W586. [14] A. Canakoglu, et al., GenoSurf: metadata driven semantic search system for integrated genomic datasets, Database-Oxford 2019 (2019). [15] D. Aran, et al., Comprehensive analysis of normal adjacent to tumor transcriptomes, Nat
Commun 8 (2017). [16] J. Kim, et al., Transcriptomes of the tumor-adjacent normal tissues are more informative than tumors in predicting recurrence in colorectal cancer patients, J Transl Med 21 (2023) 209. [17] E. Oh, H. Lee, Transcriptomic data in tumor-adjacent normal tissues harbor prognostic information on multiple cancer types, Cancer Med 12 (2023) 11960–11970. [18] P. Pinoli, et al., Pan-cancer analysis of somatic mutations and epigenetic alterations in insulated neighbourhood boundaries, PloS one 15 (2020) e0227180. [19] A. Mehrgou, S. Teimourian, Update of gene expression/methylation and mirna profiling in colorectal cancer; application in diagnosis, prognosis, and targeted therapy, Plos one 17 (2022) e0265527. [20] K. Asada, et al., Uncovering prognosis-related genes and pathways by multi-omics analysis in lung cancer, Biomolecules 10 (2020) 524. [21] L. Tong, et al., Deep learning based feature-level integration of multi-omics data for breast cancer patients survival analysis, BMC Med Inform Decis 20 (2020) 225. [22] A. Chesi, et al., Exome sequencing to identify de novo mutations in sporadic als trios, Nat
Neurosci 16 (2013). [23] M. Mousa, et al., Whole-exome sequencing in family trios reveals de novo mutations associated with type 1 diabetes mellitus, Biology 12 (2023) 413. [24] D. Calvanese, et al., Ontology-based database access, in: SEBD, 2007, pp. 324–331. [25] G. Booch, et al., The Unified Modeling Language User Guide, Addison-Wesley, Reading,
MA, 1999. [26] S. Graw, et al., Multi-omics data integration considerations and study design for biological systems and disease, Mol Omics 17 (2020). [27] J. Zhang, et al., The international cancer genome consortium data portal, Nat Biotechnol 37 (2019) 367–369. [28] M. Byrska-Bishop, et al., High-coverage whole-genome sequencing of the expanded 1000 genomes project cohort including 602 trios, Cell 185 (2022) 3426–3440. [29] B. Xu, et al., De novo gene mutations highlight patterns of genetic and neural complexity in schizophrenia, Nat Genet 44 (2012) 1365–1369. [30] X.-W. Wang, et al., Benchmarking omics-based prediction of asthma development in children, Respiratory Research 24 (2023) 63. [31] A. Rahnavard, et al., Molecular epidemiology of pregnancy using omics data: advances, success stories, and challenges, Journal of Translational Medicine 22 (2024).