Sepsis, a global health priority, is defined as a severe host response to a systemic infection, leading to a life-threatening organ dysfunction (1), with an incidence of approximately 189 adult cases per 100,000 population/annually (2, 3). The global mortality rate was around 17% in 2017, with significantly higher rates observed in patients with septic shock (4). Sepsis survivors may face significant functional and cognitive long-term disability (5) with important health and socioeconomic consequences. The use of sepsis bundles could improve survival (6, 7) although early identification of sepsis is mandatory, which is challenging even by experienced clinicians (6).

Multiple studies have focused on early detection of sepsis using information from patient demographics, vital signs, laboratory results, biosensors, and/or genetics (8–11). Since the host immune response to microbial agents is influenced by genetic variation (8), recent genome-wide association studies (GWAS) showed a potential to identify genomic variants associated with sepsis susceptibility (12), sepsis-associated acute respiratory distress syndrome (13), and sepsis-associated mortality (4, 14–17) in adults.

The genes identified in the few GWAS on sepsis patients differ among studies (4, 12–17), which could be mainly attributed to the heterogeneity of the patient populations involved. In this context, post-operative sepsis represents up to one-third of all sepsis patients (18) and has mortality rates that can reach up to 50% (3, 19). Its management involves high economic costs due to the need for mechanical ventilation and prolonged hospital stays (19, 20). Early diagnosis is complicated by the similarity of its symptoms to normal post-operative inflammatory responses (21), which delays timely intervention. Additionally, antibiotic resistance and the need for additional surgical interventions increase the complexity of treatment, especially in immunocompromised patients or those with multiple risk factors, where available biomarkers lack sufficient specificity for an accurate diagnosis (20). Postoperative sepsis is highly relevant for genetic studies, because the time of the precipitating insult is known (surgery). Hence, a preoperative assessment of genetic susceptibility to sepsis could enable risk stratification, guiding intensified monitoring or preventive interventions to better protect high-risk surgical patients (12, 22).

The existing GWAS of sepsis are characterized by the use of a single-locus analysis to select significant single-nucleotide polymorphisms (SNPs) (4, 12–17). In this respect, conventional methods may fail to capture complex interactions of SNPs with intermediate specificity that may have a higher contribution to the heritability of phenotypes (23, 24). Conversely, explainable artificial intelligence (XAI) approaches, which provide an ability to explain the decisions taken by complex artificial intelligence (AI)-based algorithms (25), have shown their usefulness to prioritize disease-associated genes (26–29). Specifically, these approaches have been previously applied for accurate phenotype prediction and susceptibility genes identification from GWAS-derived SNPs in complex diseases, such as atrial fibrillation (26), attention deficit hyperactivity disorder (27), hypertension, or diabetes (28).

In this exploratory study, we have analyzed genotype data from a large and homogeneous cohort of post-operative patients with sepsis and population controls to accurately predict sepsis and prioritize novel genetic variants associated with sepsis susceptibility. We hypothesized that XAI-based analysis of GWAS data could lead to the identification of new post-surgical sepsis susceptibility genes. Accordingly, our main objective was to obtain an XAI model that allows us to identify SNPs contributing to accurate prediction of post-operative sepsis and analyze their biological and functional implications, thereby facilitating their interpretation.

To the best of our knowledge, this is the first XAI approach applied to GWAS-derived SNPs for patients with post-operative sepsis. We identified top-ranked SNPs with higher contribution to sepsis prediction involved in chromatin regulation, regulatory motifs, DNase I sensitivity, and significant eQTLs in different cell lines and tissues (including blood, fibroblasts, and immune response cells). Among the top20-ranked SNPs, we found SNPs associated with inflammation, blood cell count, and sepsis traits, and in sensitivity analysis of the sepsis cohort, we observed SNPs associated with hospital and ICU stay length, chronic liver failure, and hypertension. This exploratory study allowed the prioritization of three potential biomarkers with the highest contribution to post-surgical sepsis prediction (rs17653532, rs1575081785, and rs74707084) located in PRIM2, RBSN, and SYNPR genes, which are involved in integrative processes, such as gene expression regulation, DNA replication, and cell proliferation.

The SNP with the highest SHAP contribution was an intronic variant located in the PRIM2 gene involved in DNA replication (35), which is critical in sepsis for its role in apoptosis, oxidative stress, and metabolic changes. PRIM2 has been related to consecutive trauma-induced sepsis based on an expression profiling analysis (36). The second SNP with the highest SHAP contribution score is in the RBSN gene, which encodes a protein from the FYVE zinc finger family and is involved in vesicle trafficking. Zinc finger proteins may contribute to inflammation, immune cell function, and tissue repair by modulating gene expression and regulating key immune-related genes (37–39). In the context of sepsis, these proteins may also affect cellular dysfunction, apoptosis, and DNA repair mechanisms, thereby affecting cell survival and tissue integrity (39). Similarly, a gene co-expression network analysis identified a zinc finger family gene (ZNF721) in a gene cluster for septic shock patients (40). Finally, a third gene, SYNPR, which encodes Synaptoporin, a protein found in the central nervous system, is involved in synaptic vesicle trafficking and neurotransmitter release (41).

In addition, our analyses revealed two genes encoding phosphodiesterases (PDE10A and PDE4D), involved in cyclic nucleotide signaling (42, 43). Cyclic nucleotide signaling is involved in several cellular processes, including immune response and inflammation (44), and phosphodiesterase inhibitors have shown potential therapeutic effects in experimental models of sepsis and lung inflammation (45, 46). In fact, cyclic-nucleotide phosphodiesterase activity and the cAMP binding were the most enriched relevant molecular processes in our study. Furthermore, the most significantly enriched biological process was negative regulation of heart contraction, involving RNLS and PDE4D genes. Cardiac dysfunction is an important consequence of sepsis, caused by increased inflammation or suppression of fatty acid and glucose oxidation, or due to adenosine triphosphate (ATP) depletion (47, 48). Although we did not find any previously significant associated variants with sepsis in our study, some of the ranked SNPs (rs201088712, rs3015358, and rs114065456) were associated with white blood cell count, sepsis, and sepsis-associated death in the UK Biobank. Thus, we have identified new possible candidate genes associated with post-operative sepsis susceptibility, highlighting the role of genes related to gene expression, DNA replication, cyclic nucleotide signaling, cell proliferation, and cardiac dysfunction.

Comprehensive XAI approaches have identified clinical features (vital signs, laboratory values, or demographics, among others) from electronic health records (EHR) contributing to early detection of sepsis (49–51) but they have not been applied to sepsis-related genomic data until now. Several GWAS studies have identified SNPs and evaluated polygenic risk scores (PRS) associated with sepsis susceptibility and mortality (4, 12–17), offering insights that could inform early prevention and treatment strategies targeting sepsis-related complications. However, only Engoren et al. (12) described a polygenic risk score for sepsis susceptibility, achieving an AUC of 0.752. This sepsis prediction performance in Engoren et al. (12) could be attributed to: (i) including both sepsis-2 and sepsis-3 adult perioperative patients identified from EHR data; (ii) using peri-operative controls identified from a university EHR, who may present confounding comorbidities; (iii) applying a different sepsis prediction model than in our study (logistic regression vs. CNN), which may not capture complex SNP interactions contributing to sepsis predisposition. Regarding the methodology in existing GWAS for sepsis (4, 12–17), it is noteworthy that the SNPs most strongly associated with the phenotype (i.e., those with the lowest p-values) could not contribute to the most predictive genetic signature. Instead, the most predictive signature often consists of SNPs that provide complementary information (23). In this respect, our XAI methodology relying on a p-value <5 × 10⁻³ threshold prioritized those SNPs that have a higher influence for phenotype prediction. Among the top 20-ranked SNPs, we found both SNPs with high (i.e., p-value from GWAS <5 × 10⁻⁸) and intermediate specificity (i.e., p-value >5 × 10⁻⁸), being one of the SNPs with intermediate specificity also in the top 3 SNPs (rs17653532 (PRIM2), GWAS p-value = 1.18 × 10⁻³), thus confirming the limitations of using the standard threshold of 5 × 10⁻⁸ for statistical significance in GWAS studies (23). Importantly, our XAI approach prioritized three SNPs (rs17653532, rs1575081785, and rs74707084) with a high contribution to sepsis prediction (i.e., highest SHAP values) and good performance when validated in an independent test subset (AUC = 0.886), highlighting the potential role of these variants and related genes in sepsis risk in post-surgical patients. In clinical practice, the output probability threshold of the sepsis prediction models could be adjusted to identify patients at higher or lower risk of sepsis, prioritizing either sensitivity (to capture more high-risk patients) or specificity (to reduce false positives), depending on the intended clinical application and available resources (52). Our results agree with previous XAI approaches applied to analyze GWAS-derived SNPs in complex diseases, such as atrial fibrillation (26), attention deficit hyperactivity disorder (27), hypertension, or diabetes (28).

Although the in silico functional findings suggest potential implications for early and personalized prevention and treatment for post-operative sepsis, further biological validation of these exploratory findings through in vitro and in vivo analyses is required for clinical relevance and generalizability. XAI provides a powerful approach to prioritize SNPs for sepsis prediction, providing insights into cohort-wide and individual-specific genetic predispositions. We also highlighted the clinical and biological relevance of genes associated with the top 3 SNPs (PRIM2, RBSN, and SYNPR). Identifying these genetic variants in a preoperative blood test could help in the early detection of sepsis in surgical patients, may enable risk stratification, and allow prompt pharmacological treatment, thus reducing mortality and long-term disability. This concept is in line with the recent clinical study by Liesenfeld et al. (53), who clinically validated an AI-driven blood test using host mRNAs expression data to predict acute infection and sepsis. However, our findings remain exploratory and further external clinical validation is crucial to confirm their reproducibility and generalizability of our findings. Studies in ethnically diverse cohorts are also warranted, given that most existing sepsis GWAS have been conducted in populations of European ancestry and genetic associations often fail to replicate across ancestral group (54, 55).

We acknowledge some strengths and limitations of our study. The existing GWAS in sepsis patients include patients with sepsis of any etiology (4, 12–17), which leads to the identification of genes that differ among studies (4, 12–17). In contrast, we focus exclusively on cases based on patients with postsurgical sepsis, making our population highly homogeneous and with applicable results within this context. Nevertheless, patient heterogeneity persists due to differences in surgical procedures, infection sources, or host responses (56–58). Thus, future studies are warranted to investigate whether specific post-operative patient subgroups exhibit distinct genetic associations or clinical outcomes. Apart from this, the study samples were collected from two hospitals, and the sample size was not very large due to the challenging nature of data collection, and rare variants in or near identified regions may go undetected due to technological limitations. In this respect, the use of a test set derived from the same underlying cohort represents a potential limitation regarding external generalizability. However, the adoption of a hold-out validation scheme (with independent training, validation, and test subsets) provides a robust internal assessment of model performance in the absence of an available external or independent postsurgical sepsis cohort, that reduces the risk of overfitting. This validation approach has been shown to yield generalizable, reproducible and biologically meaningful results in recent AI-based genomic studies (26–28). Nevertheless, the validation in a large, independent, geographically, and ancestrally distinct dataset of post-operative patients with diverse ancestries would be a necessary future step to confirm our findings. This would also allow consideration of additional factors such as comorbidities, measurements at diagnosis, and sources of infection in sepsis prediction. Genome and whole exome sequencing analyses would also provide a better resolution to achieve this goal. Regarding the proposed XAI methodology, we used a CNN for sepsis prediction, originally designed for image analysis. Recent studies have shown that CNNs are suitable to analyze GWAS data (26, 27). Another limitation is the use of population controls who were not clinically evaluated for sepsis. As such, we cannot entirely exclude the possibility that some controls could have been cases exposed to relevant environment factors or survivors of a previous sepsis episode. However, the use of population controls provided access to a well-characterized and substantial sample size cohort, thereby increasing the statistical power of the study, and they are commonly utilized in large-scale GWAS studies of infectious diseases (31, 32, 59–61). Notably, the COVID-19 Host Genetics Initiative consortium studies demonstrated that genetic analyses using various comparisons of controls with COVID-19 severity (e.g., infection, hospitalization, critical illness) yield overlapping results, supporting the validity of using population controls in genetic research on infectious diseases (59). Similarly, several GWAS of infectious conditions have successfully relied on population-based biobanks as control sources when disease-specific non-affected cohorts were unavailable, highlighting the practical and methodological acceptance of this approach in complex traits (31, 32, 60, 61). Conversely, although this has been observed in other studies, and we adjusted for sex and age in the association analysis to reduce bias due to age imbalance between cases and controls, young patients may eventually undergo surgery and develop sepsis. Moreover, the genotype by environment interactions may show challenges in this kind of study due to the dynamic influence of the environment on gene expression. As a result, the effects of genetic variants could be masked or modified by environmental factors. Thus, future studies should include clinical validation in postsurgical or critically ill patients without sepsis as controls once reliable genetic data is available, to better delineate genetic variants specifically associated with sepsis susceptibility within the context of critical illness.

In conclusion, our XAI approach applied to GWAS-derived SNPs enabled the identification of significant risk loci associated with post-surgical sepsis that could be implemented in clinical practice for improving patient outcomes. We found variants with functional, regulatory and clinical implications, as well as genes related to gene expression regulation, DNA replication, cyclic nucleotide signaling and cell proliferation, and cardiac dysfunction, among other biological processes. We also identified three potential biomarkers with the highest contribution to sepsis prediction (rs17653532, rs1575081785, and rs74707084), located in PRIM2, RBSN and SYNPR genes, which could be determined in a preoperative blood test, allowing targeted and precise interventions to prevent and treat sepsis in patients undergoing surgery. Further investigations, including in vitro and in vivo analyses, as well as complementary studies in cohorts comprising sepsis and non-sepsis patients undergoing major surgery will be needed to optimally evaluate the genetic factors contributing to sepsis predisposition and to provide an external validation of our exploratory findings.