The transcriptome is the total set of expressed RNA in a cell or population of cells at a specific time point. Mature messenger RNA (mRNA), which is the interim carrier of information between the genome and protein, is transcribed from a very small fraction (less than 2% to 3%) of cellular DNA (International Human Genome Sequencing Consortium, 2004). Multiple regulators decide the fate and character of the message passed on to form proteins through various mechanisms including alternative splicing and RNA editing (de Hoon et al., 2015; Abascal et al., 2020). The regulation of the whole machinery is extremely complex and involves long non-coding RNAs (lncRNAs), microRNAs (miRNAs), transfer RNAs (tRNAs), ribosomal RNAs (rRNAs), small nuclear RNAs (snRNAs), small nucleolar RNAs (snoRNAs), short interfering RNAs (siRNAs) and other transcripts. Furthermore, high throughput technologies have identified a plethora of novel RNA molecules but their involvement in various cellular activities is unclear (Pertea, 2012; Palazzo and Koonin, 2020).

RNA is the most dynamic cellular component regulating gene expression through complex processes including transcription, maturation and degradation (Cao and Grima, 2020). Transcription mostly occurs intermittently (on/off promoter switches) and the size and frequency of transcription bursts contribute to the molecular phenotype of a cell at a particular time point (Eling et al., 2019). Deterministic factors drive the mean expression of a gene without accounting for stochastic processes (Kærn et al., 2005). However, intrinsic molecular fluctuations (stochastic noise) have been linked to important processes such as cell fate, immune plasticity, ageing and cancer development. The combination of deterministic and stochastic components drives non-genetic heterogeneity which is modulated by gene-regulatory circuits and results in variability in transcript abundance across seemingly homogenous cell populations (Eling et al., 2019). Although there is an inverse correlation between mean gene expression and fluctuation, it has been recently shown that changes in transcriptional noise can initiate cell re-programming and development while mean gene expression remains stable (Desai et al., 2021). Collectively, therefore, RNA corresponds to a snapshot of the cellular state and has enormous potential for application to clinical diagnostics (Byron et al., 2016).

Technological advancements have enhanced our understanding of the transcriptome. Reverse transcriptase polymerase chain reaction (RT-PCR) is considered a gold standard for detecting qualitatively and quantitatively a limited number of transcripts (Dramé et al., 2020). Microarrays have revolutionised our approach to RNA measurement by using probes on a solid surface that hybridise with thousands of transcripts (Schena et al., 1995). They were recognised as a key tool in advancing personalised medicine (Shi et al., 2006), but despite 25+ years of development, their clinical utility remains limited (Piccart et al., 2021). Variation in sample preparation decreases reproducibility and background noise obscures the detection of low signal transcript expression. RNA sequencing (RNA-seq) technologies are more powerful tools as pre-defining RNA targets is not required and they have a greater dynamic range. RNA-seq allows the detection of the diversity in the transcriptome through the quantification of known and novel transcripts regardless of their abundance, including non-coding RNA, single nucleotide variants, fusion genes and splice variants (Byron et al., 2016). Moreover, RNA-seq at the level of single cell (scRNA-seq) allows the detection of previously unexplored processes such as transcriptional noise (Eling et al., 2019; Desai et al., 2021). Although RNA-seq outperforms microarrays in assessing complex gene expression profiles, prediction of clinical endpoints is not affected by the platform (Zhang et al., 2015), and data based on microarray experiments have driven a plethora of discoveries. A caveat of whole RNA sequencing is its relatively poor ability to identify and quantify low abundance transcripts. Probe-based assays targeting genes of interest, such as RNA CaptureSeq have been developed to fill this gap along with sophisticated bioinformatic algorithms aiming to increase detectability of unknown sequences (Grioni et al., 2019).

Gene expression profiling provides an enormous amount of high-resolution data from a single experiment. The size of the human transcriptome remains debatable with the majority of it referred to as “dark matter” because its function is unknown (Kapranov and St Laurent, 2012). RNA-seq exceeds the size of the human genome by generating up to six billion short reads and their assembly into the transcriptome is a challenging task (Pertea, 2012). Complex computational algorithms are deployed at multiple stages of data analysis and require bioinformatics expertise (Kukurba and Montgomery, 2015). The analytical pipelines attempt to identify a set of informative genes to guide the elucidation of novel molecular mechanisms, the development of prognostic and predictive biomarkers and the identification of druggable targets.

Over 100 genetic tests for 30 conditions in the field of oncology, haematology, genetic disorders and pharmacogenetics, have received FDA approval to date. Less than ten tests are based on RNA measurement and only four utilise gene expression profiles with more than two RNA targets (FDA, 2021).

Molecular diagnostics focusing on the genome suffer from a limited ability to reflect accurately the in vivo variability within a condition at a particular time-point and among patients. The hallmark of acute lymphoblastic leukaemia (ALL), for instance, is numerous genetic aberrations stratifying patients into prognostic and therapeutic groups (Pui et al., 2019). However, multiple mutations identified at the genome level may not be contributing to the disease. Transcriptome sequencing characterises clinically relevant genomic alterations and variants in real time with higher sensitivity compared to whole-genome sequencing and it has been crucial in the discovery of novel subtypes and therapy tailoring (Roberts and Mullighan, 2015). Gene expression studies identified the Philadelphia chromosome-like ALL subtype and the downstream involvement of kinases guiding the use of tyrosine kinase inhibitors (TKI) (Inaba et al., 2017).

Transcriptomics is explored as a complementary method to genomic testing for precision-based treatments in cancer patients (Lee et al., 2021; Tsimberidou et al., 2022). The Worldwide Innovative Network (WIN) study to select rational therapeutics based on the analysis of matched tumour and normal biopsies in subjects with advanced malignancies (WINTHER, NCT01856296) was the first large-scale prospective clinical trial that allowed a fraction of patients with no actionable DNA alterations to have RNA-guided treatments using a novel algorithm (Rodon et al., 2019). The Individualised Therapy For Relapsed Malignancies in Childhood (INFORM) registry collects real-world clinical and multi-omic data from routine biopsies to translate them to precision treatments and inform future clinical trials (van Tilburg et al., 2021). The first trial (NCT03838042) is ongoing and investigates the combination of Nivolumab and Entinostat in children and adolescents with refractory high-risk malignancies. Stratification of patients in accordance with their tumour genetic mutation and gene expression profiles will serve for the purposes of biomarker development and to minimise unnecessary risks in patients (van Tilburg et al., 2020).

Liquid biopsy of extracellular RNA (exRNA) has been embraced as a promising tool for screening and disease monitoring purposes and as an alternative to invasive methods of diagnosis such as tissue biopsy (Heitzer et al., 2019; Zhou et al., 2020; Wu et al., 2022). Although studies investigating exRNA in clinical application are scarce, recent developments in oncology are paving the way by enabling the distinction between tumour-specific RNA and total circulating extracellular transcriptome (Vermeirssen et al., 2022; Zong et al., 2023). Moreover, analysis of intracellular RNA of circulating tumour cells and peripheral blood mononuclear cells (PBMC) has identified prognostic pathways for response to treatment in patients with metastatic castration-resistant prostate cancer (Zhang et al., 2022).

In breast cancer, patient stratification based on expression of tumour markers (e.g., ER, PR and HER2 in breast cancer) has guided treatment strategies for over 30 years (Cardoso et al., 2016) laying the foundation for remarkable advances in molecular diagnostics (Buus et al., 2021). Early breast cancer (Supplementary Box S1) represents a successful paradigm of the applied knowledge accrued from transcriptomics in clinical practice. Only a small proportion of patients with oestrogen receptor positive (ER+) and lymph node negative (LN-) breast cancer benefit from adjuvant chemotherapy. Unfortunately, clinicopathological features poorly characterise ER+/LN- tumours and immunohistochemical techniques cannot be relied on to make treatment decisions (Fitzgibbons et al., 2000; Eifel et al., 2001). The standard practice has been to use a combination of hormonal and chemotherapy regimens, despite evidence suggesting that around 80% of patients were overtreated and unnecessarily exposed to chemotherapy and the potential toxicity (van 't Veer et al., 2002). Hence, identification of gene expression signatures able to predict risk of recurrence, and therefore stratify treatment, was a breakthrough in early breast cancer management (Schaafsma et al., 2021).

Commercially available assays, such as Oncotype DX (Genomic Health), MammaPrint (Agendia), EndoPredict (Myriad Genetics) and Prosigna (Nanostring Technologies) are endorsed by the UK National Institute for Health and Care Excellence (NICE) and international guidelines (Table 1). Expression levels of specific genes are measured in tumour samples with RT-PCR (Oncotype DX) or microarrays (MammaPrint) and a prognostic score is calculated with mathematical models in order to stratify patients into risk groups (Sparano et al., 2018; Piccart et al., 2021). EndoPredict produces a score based on both transcriptional and clinical (tumour size and nodal status) features. Prosigna classifies breast cancer into subtypes and calculates a score based on gene expression, subtype, clinical parameters (tumour size and nodal status) and proliferation pathways (Paik et al., 2006). Oncotype DX is based on a 21-gene signature which is independent of clinicopathological factors (Sparano et al., 2018). It is the only multi-gene assay which is validated to predict adjuvant chemotherapy benefit in addition to prognosis (Syed, 2020).

The development of Oncotype DX was a gradual process involving the use of data from large clinical studies and diligent address of issues (Supplementary Box S2). Due to the remarkable molecular diversity of breast tumours (Perou et al., 2000), numerous clinical and immunohistochemical biomarkers and their combinations had failed to guide treatment decisions (Hayes, 2000). Moreover, previous attempts to identify predictive and prognostic gene expression signatures were based on single studies which were neither standardised nor reproducible. The 21-gene signature in Oncotype DX was derived from a set of 250 genes which was selected from well-designed studies and public databases utilising microarrays (Paik et al., 2004). The 250 candidate genes were narrowed down to 21 through three independent clinical studies including almost 500 patients who received adjuvant hormonal treatment plus chemotherapy or hormonal treatment alone (Paik et al., 2003; Cobleigh et al., 2005). An algorithm was developed to produce a continuous variable, the Recurrence Score (RS) based on the expression of these genes, which is comprehensible by clinicians and stratifies patients into high and low risk groups for distant recurrence within 10 years of surgery (Paik et al., 2004). RS showed remarkable statistically significant prognostic ability and predictive ability and has been extensively validated in large prospective randomised clinical trials and real-world data from population-based registries (Paik et al., 2004; Nitz et al., 2017; Sparano et al., 2018; Syed, 2020). Further analyses of these studies have identified that pre-menopausal women would benefit from the addition of clinical factors (age, tumour size, and histologic grade) along with RS for shaping management strategies (Hunter and Longo, 2019; Sparano et al., 2019).

Oncotype DX and the Decipher Genomic Classifier (21 and 22 expressed genes, respectively) have been shown to be cost-effective approaches for guidance of treatment decisions (Lobo et al., 2017; Berdunov et al., 2022). This results from a combination of test accuracy in reducing unnecessary toxic treatments such as chemotherapy and radiation while not excluding patients from beneficial treatments (Lux et al., 2022).

Following on from the success of the oncotype Dx for early breast cancer, the Oncotype DX Genomic Prostate Score has been developed on the same principles and similar processes (Supplementary Box S3). It aims to prevent unnecessary surgery and radiation by stratifying patients into low-risk and aggressive disease. However, there are no large prospective studies to validate the prognostic performance of the assay for clinical outcomes (Eggener et al., 2019; Brooks et al., 2021). Attempts to identify a gene expression signature prognostic of prostate cancer are based on tissue samples derived from needle-core biopsies and the limited amount of tissue may be a constrain to characterise heterogeneity (Supplementary Box S3).

Disease heterogeneity is a major caveat in the design of diagnostic biomarkers. Inter-assay comparisons revealed discordance in prognostic performance of gene expression-based tests for stratification of patients with early breast cancer (Varga et al., 2019; Abdelhakam et al., 2021; Buus et al., 2021). These discrepancies may derive from the diversity in gene sets, methodology and algorithms and design of studies. Of note, there is only minor overlap of genes among predictive tests (Supplementary Table S1). Heterogeneity in gene composition reflects the variety of molecular mechanisms involved in disease progression and it may not necessarily influence prognostic ability. Currently, a prospective study is investigating the clinical validity of Curebest 95GC, a microarray-based measurement of the whole genome in tumour tissues (Naoi et al., 2021). The results are anticipated to shed light on the number of transcripts required for stratification of patients with early breast cancer. However, an increased number of genes may be a major obstacle in the development of a cost-effective marker and mechanistic studies could assist with reducing the number (Gliddon et al., 2018).

Heterogeneity at the tissue level is multi-layered and not confined to the oncogenic cells. Neoplastic cancer cells are nurtured by neighbouring stromal cells comprising the tumour microenvironment (TME). A diverse community of tumour infiltrating immune cells is a major component of the stromal microenvironment exerting both beneficial and detrimental effects (Hanahan and Coussens, 2012). Growing evidence shows that quantification of the proportion of leucocyte subsets can assist in prognosis and therapy choice (Gentles et al., 2015). Traditional methods such as immunohistochemistry and flow cytometry can identify a limited number of pre-defined cell populations but fail to discriminate unknown or closely related phenotypes.

By contrast, gene expression profiling coupled with computational algorithms can characterise cell composition of complex tissues (Finotello and Trajanoski, 2018; Xu et al., 2021). Many tools have been developed based on two main methods:

• in silico deconvolution (CYBERSORT, TIMER, EPIC, quanTIseq, DeconRNAseq, PERT, DSA, MMAD, ssKL); and

Deconvolution is based on a linear model of the expression of a gene in the different cell types. Digital dissection of the tumour into the relative fractions of cell types is estimated against a library of cell-specific expression signatures (reference signature or matrix). The matrix appears rigid considering the diversity of infiltrating immune cells (extend, type, activation status, interactions, and closely related cells) among tissues and cancer stages (Hanahan and Coussens, 2012; Newman et al., 2015). Refinement of the gene set enrichment method attempts to circumvent the issue by producing enrichment scores based on the expression levels of a set of cell-type-specific marker genes by analysing various data sources (Aran et al., 2017). However, performance is poorer in real mixtures compared to simulated mixtures and statistical significance is not reported for prediction of cell abundance (Newman et al., 2015; Aran et al., 2017). Deconvolution algorithms which simultaneously estimate relative cell fractions and produce a matrix of expression profiles have been developed (MMAD, DSA, ssKL, ssFrobenius, and deconf), but they are flawed by mathematical complexity and a limited ability to quantify a higher number of immune cells (Finotello and Trajanoski, 2018). Although several studies have tested these computational approaches in simulated samples and publicly available datasets showing good performance, evidence about their clinical validity is scarce (Desmedt et al., 2018; Newman et al., 2019; Waks et al., 2019; Craven et al., 2021).

It is unclear if gene signature enrichment and deconvolution approaches accurately portray the complexity of cellular heterogeneity in cancer samples and more work is warranted before testing in clinical settings. Definition of reference expression profiles is a fundamental caveat allowing for the identification of only a few dozens of cell types which may not reflect all heterogenic subsets in tumours. The effort should probably be on revealing hallmark phenotypes with prognostic and predictive capability in clinical settings to populate the reference matrix or marker gene-sets. For instance, the role of exhausted (increased PD-1 expression) CD8⁺ tumour infiltrating lymphocytes is well established in melanoma, renal and non-small cell lung cancer and it has guided the use of immune check point inhibitors (ICIs) (Sade-Feldman et al., 2018; Thommen et al., 2018; Young et al., 2018; McLane et al., 2019). Tumour-associated macrophages are another interesting group of cells because of their abundance in the tumour microenvironment. Unravelling of the complex subpopulations has shown that the classical categorisation to M1 and M2 polarised macrophages is an oversimplification of their crucial role in cancer regulation (Mantovani and Longo, 2018; Duan and Luo, 2021; Xiang et al., 2021).

Immune response to infection is initiated by a “genomic storm” of both pro-inflammatory and anti-inflammatory cytokines expressed concomitantly (Nakamori et al., 2020). In sepsis there is acute cellular reprogramming and failure to restore balance between immune activation and suppression can present with life-threatening organ dysfunction (Singer et al., 2016; van der Poll et al., 2017). Gene expression studies have revealed remarkable heterogeneity in sepsis due to host parameters (e.g., genomic variation and co-morbidities), source of infection and stage of illness. This may explain, at least partly, the reason for the failure of numerous promising therapeutic agents in clinical trials (Marshall, 2014; Davenport et al., 2016; Peters-Sengers et al., 2022). The definition of sepsis has also been revised several times, and each definition can dramatically alter the composition of cohorts that are included in studies (Johnson et al., 2018). This can also negatively impact model development, particularly where retrospective data collection is required or data are pooled across studies (Sauer et al., 2022).

To address individual variation in the response to sepsis, a multi-layered approach, including at the molecular level, for stratification in treatment subgroups is required. As a proof-of-concept, machine learning algorithms which classify patients based on routine clinical data have been shown to accurately predict clinical outcomes and sepsis onset (Komorowski et al., 2018; Seymour et al., 2019; Fleuren et al., 2020). Machine learning is a very powerful tool for harnessing large-scale data with the aim of identifying predictive biomarkers (Zhang et al., 2021). The use of diverse methods analysing transcriptomic data in various conditions has been previously reviewed (Vadapalli et al., 2022). Appreciation of common pitfalls and focus on interpretable findings has transformed these complex computational approaches into comprehensive tools (Sidak et al., 2022; Whalen et al., 2022). However, despite our increased understanding of sepsis pathogenesis with new technologies, translation of research knowledge to improvements in clinical practice has been exceedingly difficult.

Transcriptomic-based real-time subclassification of patients has been developed and validated in individual studies (Table 2). The Knight group investigated gene expression profiles in peripheral blood leukocytes of patients on Intensive Care Units (ICU) with faecal peritonitis and community acquired pneumonia (Davenport et al., 2016). They proposed two sepsis phenotypes associated with prognosis. Genes comprising the sepsis response signature (SRS) demonstrated significant overlap between the two sources of infection and with trauma patients, while gene expression and SRS membership changed temporally (Burnham et al., 2017). Single-cell multi-omics evaluation showed that an immature immunosuppressive population of neutrophils together with enrichment in the IL-1 pathway are the biological underpinnings of the SRS1 group who experienced increased early mortality (Kwok et al., 2022). In contrast, the immuno-competency of the SRS2 endotype was compromised by corticosteroids in a randomised clinical trial which showed an association between hydrocortisone use and higher mortality in the SRS2 group but not in the SRS1 group (Antcliffe et al., 2019). The SRS investigators upgraded their classifier to the SepstratifieR framework which can be applied to multiple infecting pathogens and data accruing from different platforms (e.g., RNA-seq and RT-qPCR). SepstratifieR utilises expression levels of signature genes, including an extended 19-gene set expected to be robust to technological variation, to align samples to a corresponding reference map and returns the SRS endotype and a severity score (SRSq). SRSq reflects immune deregulation and has the advantage of modelling patients as a continuum which is a better descriptor of molecular profiles compared to classes (Cano-Gamez et al., 2022).

The Molecular Diagnosis and Risk Stratification of Sepsis (MARS) project identified four endotypes (Mars1–4) in patients with sepsis admitted to ICU in the Netherlands (Scicluna et al., 2017). Biomarkers for each endotype were derived from a 140-gene expression signature (Table 2). The authors proposed that patients classified as Mars1 were the most clinically relevant group with consistently increased mortality. Comparisons with the SRS revealed an overlap between the low-risk groups SRS2 and Mars3, but not the expected enrichment of Mars1 patients within SRS1 (Scicluna et al., 2017; Cano-Gamez et al., 2022). One explanation could be the primarily leukocyte-based training data for SepstratifieR differing from the whole blood-derived RNA used for MARS signatures. The differences could also be attributed to variation in populations utilised for classifier development, technical procedures, bioinformatics analysis and study design (Table 3). At the gene level though, similarities in differential expression and active pathways were observed. Moreover, classification of MARS patients into SRS endotypes showed that SRS1 had a higher proportion of septic shock and elevated Sequential Organ Failure Assessment (SOFA) scores, but not increased mortality, reflecting the presence of unobserved variables preventing the severe sequelae of sepsis (Cano-Gamez et al., 2022).

Paediatric patients with septic shock were categorised into three groups based on a 100-gene set microarray-derived signature (Wong et al., 2009; Wong et al., 2011). Two (A and B) of the three subclasses were identified with the use of a different platform for mRNA quantification (NanoString nCounter) which has the potential for clinical application due to its decreased turnaround time and cost (Wong et al., 2015). The authors noted that subjects in subclass B and C demonstrated similar clinical phenotypes, whereas subclass A patients had poorer outcomes. The previously reported association between mortality and corticosteroid use among patients of a specific endotype was also observed, but this time within the subclass with increased mortality (subclass A, Wong et al., 2015). Interestingly, when the Mars signature was applied to the original paediatric population, only three of the four endotypes were stably recognised and there was no association between endotype categorisation and mortality (Scicluna et al., 2017). The search for prognostic biomarkers in paediatric septic shock has led to the development of the paediatric Sepsis Biomarker Risk Model (PERSEVERE), which is a predictive tool of mortality and disease severity (Jacobs et al., 2019; Wong et al., 2019). A panel of 117 gene probes possibly associated with outcome in children with septic shock was used to select 12 genes with a protein product which had a mechanistic role in immune responses to infection and was readily measured in serum. Classification and regression tree analysis reduced the number of proteins to five and selected age among various clinical parameters as the best combination of factors to predict 28-day mortality (Wong et al., 2012). PERSEVERE incorporates the protein products of those five genes and has been tested as a predictor of sepsis-related organ dysfunction in various cohorts (Wong et al., 2016; Yehya and Wong, 2018; Stanski et al., 2020; Al Gharaibeh et al., 2022; Atreya et al., 2022). Clinical utility is yet to be decided through large prospective validation studies. Utilization of proteins identified through gene expression exploratory studies may achieve better reproducibility among cohorts, but correlation of mRNA and protein product is affected by various biological and technical parameters and clinical translation is yet to be decided.

The aforementioned unsupervised clustering studies (Table 2) defined novel molecular subgroups in sepsis and produced data-driven classifiers with potential for clinical implementation. Although such approaches are the foundation of precision medicine, results are often non-reproducible because they accrue in a method-specific computational manner and/or from underpowered sample sizes. The availability of high-dimensional data from various studies in public databases and meta-clustering techniques have allowed the development of transcription-based models with improved representation of disease and population heterogeneity. A large pool of bacterial sepsis transcriptomic datasets (23 datasets; n = 1,300) identified three clusters which were descriptive of underlying molecular pathways, the Inflammopathic, the Adaptive and the Coagulopathic (Table 2). Comparisons with previously published signatures showed that the inflammopathic cluster tended to overlap with the paediatric septic shock subclass B and SRS1 and the Adaptive cluster was associated with SRS2 (Sweeney et al., 2018a). Identification of the same group of sepsis patients in independent studies with separate techniques supports the existence of molecular subtypes. The addition of a third cluster in a bigger study underscores the importance of utilising large public datasets. In a community-based approach, three independent teams built four separate models to predict mortality in sepsis using all available gene expression datasets (Sweeney et al., 2018b). Despite common data inputs, there was little overlap in predictive genes between groups due to differences in analytical approaches. Still, the model performances were broadly similar. Moreover, the combination of gene expression-based predictors with routine clinical parameters was shown to improve prognostic accuracy (Wong et al., 2014; Scicluna et al., 2017; Sweeney et al., 2018b).

The predictive performance of candidate biomarkers attempting to distinguish between the presence and absence of infection in critically ill patients has historically been suboptimal (Pierrakos and Vincent, 2010; Wacker et al., 2013). An informative biomarker consisting of a gene expression ratio has been proposed to assist in discriminating between community acquired pneumonia (CAP) and non-CAP patients, but its relatively low negative predictive value precludes it from being a stand-alone diagnostic test (Scicluna et al., 2015). Similarly, the FDA approved SeptiCyte LAB (Immunexpress, Seattle, WA), which provides a score based on the expression of four genes, is intended to be used in conjunction with clinical factors and clinical judgement to distinguish patients with sepsis from non-infective systemic inflammation within 24 h of ICU admission (McHugh et al., 2015). Different studies evaluating the discriminative power of this novel biomarker have produced conflicting results (McHugh et al., 2015; Zimmerman et al., 2017; Koster-Brouwer et al., 2018). Comparison of three scores aiming to distinguish between the presence or absence of infection in critically ill patients (FAIM3:PLAC8, SeptiCyte LAB and MetaScore or SMS) demonstrated similar performance with some superiority of the SMS (Table 2) when applied to a different cohort of patients (Sweeney and Khatri, 2017; Maslove et al., 2019). The absence of gold standard reference test dictated the use of strict criteria to define cases and controls for a supervised analytical approach for classifier development (Table 2). As a result, the discovery cohort cannot mirror the wide spectrum of heterogeneity which is inherent in sepsis patients. It is likely that leveraging of clinical and technical heterogeneity seen in larger publicly available datasets and extensive validation may help in ameliorating limitations regarding generalisability.

Transcriptomic and genomic samples are collected during most clinical trials in cancer and other diseases (NIH, 2022). Their aim is to increase our understanding of molecular mechanisms. Investigators are not obliged to submit transcriptional data deriving from interventional clinical trials to public databases unless they are presented in a publication. Hence, a plethora of interesting data may become available later or never. Clearly, it is important for investigators to deposit data from their studies in a standardised format into publicly available databases as such democratisation of data undoubtedly accelerates the pace of progress. We think that adequate progress from the translational to the clinical stage can be achieved with combination of data from different populations and to this purpose investigators should be assisted in processing their raw data early and prompted to deposit them in public databases.

Although 80% of the blood transcriptome shows differential expression in critical illness, immune responses demonstrate significant commonality leading to a remarkable overlap in expressed genes in all-cause inflammation, regardless of the presence of an infection or not (van der Poll et al., 2017). A multicohort analysis of publicly available datasets showed that there is a small proportion of distinct genes in patients with sepsis compared to patients with a non-infective critical condition in samples obtained within 48 h of admission (Sweeney et al., 2015). These findings highlight the common trajectory of the transcriptional storm that settles down during recovery underscoring the importance of time-course-based approaches (Sweeney and Wong, 2016). Gene expression signatures which predict infection have been identified in the blood of hospitalised patients up to 5 days prior to onset of symptoms and/or diagnosis (Johnson et al., 2007; Cobb et al., 2009; Sweeney et al., 2015; Yan et al., 2015; Lukaszewski et al., 2022). These findings highlight the molecular events which occur before disease symptomatology. If the immune response is not successful in clearing the pathogen(s) during this period, more robust measures are deployed leading to a transcriptional storm (Figure 1).

Tests based on gene expression thus describe “the moment in time” which has the potential for guiding targeted therapies and personalised management (van der Poll et al., 2017). However, there is no way to match the expressed molecular moment to the exact point of the disease (Figure 1) because the duration of each stage varies significantly. As an example, many groups put their efforts into identifying a classifier within 24 h of ICU admission. We may assume that this is located within the transcriptional storm space, but we cannot say whether it is in the beginning, middle, end of the curve or even within the pre-disease space. The point of symptom onset relative to the infection point potentially varies among individuals and so does presentation and admission time. Hence, despite the efforts of time-based approaches, sampling time can be defined only clinically and not objectively across the gene expression course, i.e. “one fits all” is unlikely to succeed. Challenge studies with controlled infection and longitudinal designs could shed more light on the importance of defining timing of sampling, but are complex to perform and expensive, and need to have a careful ethical framework.

There are few promising biomarkers currently in the pipeline. A combination of three non-overlapping signatures identified from a multi-cohort analysis (Sweeney et al., 2015; Sweeney et al., 2016; Sweeney et al., 2018b) has led to TriVerity (formerly known as InSepTM HostDxTMSepsis and Inflammatix) (Mayhew et al., 2020). This 29-gene expression-based test with a turnaround time less than 30 min is expected to identify the presence, type (bacterial or viral) and risk of mortality of infection (Mayhew et al., 2020; Bauer et al., 2021; Safarika et al., 2021; Brakenridge et al., 2022; Galtung et al., 2022). A Point-of-Care Test claiming to distinguish bacterial from viral infections in children is in its infancy (Pennisi et al., 2021). It is based on the expression of two genes (IFI44L and FAM89A) which emerged from a microarray-based study in almost 500 febrile children (Herberg et al., 2016; Kaforou et al., 2017). There is a repertoire of promising findings in children with infections such as tuberculosis, bacterial pneumonia, rhinovirus and respiratory syncytial virus (RSV) and the transfer of transcriptomics knowledge to routine clinical care may be seen in the near future (Mejias et al., 2021). Investigators have also adapted a mechanistic-orientated approach to select a set of genes with known correlation with sepsis outcome instead of a crude exploration of bulk RNA (Chen et al., 2022; Kreitmann et al., 2022), but further consideration of this is beyond the scope of this review.