Submitted:
10 October 2026
Posted:
13 October 2026
You are already at the latest version
Abstract
Background: The chromosome 9 CTNNAL1–ELP1 locus is associated with pediatric sepsis susceptibility, but the regulatory gene and developmental context remain unresolved. Methods: We integrated pediatric sepsis fine-mapping (510 cases, 994 controls) with eQTL data from resting and stimulated adult monocytes, neonatal myeloid regulatory data, ELP1 exact-junction splicing, colocalization, and cis-Mendelian randomization using exposure side instruments selected and frozen before outcome extraction. Twenty-eight prespecified gene–state exposures across five regional genes were evaluated. Results: rs28361152 dominated the 11-variant disease credible set (posterior inclusion probability=0.605). ABITRAM showed strong colocalization after 24-hour lipopolysaccharide stimulation (PP.H4=0.904) and supportive sharing after interferon-γ stimulation (PP.H4=0.716), whereas CTNNAL1 regulatory signals were disease-distinct. Higher genetically predicted TMEM245 expression after 24-hour interferon-γ stimulation was associated with lower sepsis susceptibility (odds ratio=0.941, 95% confidence interval 0.908–0.975; P=8.11×10⁻⁴; false-discovery-rate q=0.011). Sensitivity analyses were directionally concordant; however, mediation remained unresolved because colocalization was not evaluable. Neonatal models yielded no purity-qualified credible sets, and rs28361152 was absent from all 11 ELP1 exact-junction credible sets. Conclusion: ABITRAM was the leading disease-linked regulatory candidate, whereas TMEM245 provided secondary Mendelian-randomization support without demonstrated locus mediation. These findings refine the locus into developmentally informed regulatory candidates and therapeutic hypotheses requiring functional validation.

Keywords:
sepsis
; child
; genetic predisposition to disease
; gene expression regulation
; quantitative trait loci
; mendelian randomization analysis
; monocytes
; colocalization
1. Introduction
Pediatric sepsis is a life-threatening consequence of infection in which outcome reflects not only the invading pathogen but also the magnitude and regulation of the host response [1]. Even after apparently comparable infectious exposures, children can follow markedly different trajectories, suggesting that susceptibility is partly encoded within the host. Early pediatric studies linked inflammatory-pathway variants to the severity of invasive infection, while family-based experiments demonstrated heritable differences in endotoxin-induced cytokine production [2,3]. Subsequent candidate-gene studies extended this concept to innate immune recognition, cytokine signaling, and coagulation; nevertheless, modest sample sizes, heterogeneous phenotypes, inconsistent replication, and biologically prespecified candidate selection limited causal interpretation [4].
Genome-scale studies subsequently shifted the field from isolated polymorphisms toward systems-level host biology. Pediatric septic shock is accompanied by broad transcriptional disruption involving inflammatory, adaptive immune, metabolic, and zinc-homeostatic pathways, while longitudinal profiling shows that these programs evolve during critical illness rather than remaining fixed [5,6]. Across the SIRS-sepsis-septic shock spectrum, increasing severity is accompanied by persistent innate activation and repression of adaptive immune and antigen-presentation programs [7], and unsupervised analyses have identified reproducible molecular subclasses within clinically similar pediatric septic shock [8]. Pathogen-specific studies similarly show enhanced monocyte/innate responses with relative suppression of adaptive immune programs in invasive Staphylococcus aureus infection [9]. Importantly, these host-response patterns are both developmental and temporal: neonatal septic shock differs transcriptionally from disease in older children, multi-cohort analyses demonstrate marked age-related variation, and serial sampling reveals substantial within-patient instability over hours [10,11,12]. Collectively, these studies define pediatric sepsis as a dynamic molecular syndrome; however, they largely describe the consequences of established illness rather than the inherited regulatory architecture that predisposes a child to invasive bacterial sepsis.
A pivotal advance came from the Swiss Pediatric Sepsis Study, which identified a genome-wide significant susceptibility locus on chromosome 9 encompassing the CTNNAL1-ELP1 region in 510 children with culture-proven bacterial sepsis and 994 controls [13]. The lead association, rs28361152, lies within a regulatory neighborhood containing CTNNAL1, ELP1, ABITRAM (formerly FAM206A), TMEM245, and FRRS1L. Yet a GWAS locus is not an effector gene, and the lead variant need not be causal. Because linkage disequilibrium can cause associated variants to tag nearby causal alleles, regional association alone cannot identify the causal variant or molecular phenotype [14]. The central post-GWAS question is therefore not which gene is nearest to rs28361152, but which regional regulatory signal genuinely shares the pediatric-sepsis susceptibility architecture.
That distinction is especially important in infection because regulatory effects are often conditional on immune activation. In primary monocytes, LPS and IFN-γ exposure can reveal cis-eQTLs that are weak or absent at baseline, with further differences between early and sustained stimulation [15]. Likewise, common variants can modify pathogen- and interferon-responsive transcription in dendritic cells, including response-QTL effects at activated transcription-factor binding sites [16]. Mechanistically, genetically determined chromatin accessibility and enhancer priming can precede stimulation and shape subsequent transcriptional responses [17]. Thus, the functional consequence of a susceptibility allele may depend on cell state, stimulus, and exposure duration rather than constitutive expression alone. Development adds a further layer: neonatal immune-cell eQTL mapping demonstrates extensive cell- and stimulation-specific regulation, including signals that differ from those detected in adults [18]. Accordingly, adult monocyte data may be highly informative for pediatric sepsis, but they cannot be assumed to reproduce neonatal myeloid regulation.
Regulatory interpretation is also complicated by the fact that gene expression and RNA splicing are distinct molecular phenotypes. GTEx established widespread genetic control of both expression and splicing across human tissues, with substantial tissue specificity and allelic heterogeneity [19], while the eQTL Catalogue extended harmonized expression- and splicing-QTL mapping across diverse cell types and experimental conditions [20]. Nevertheless, a strong regional eQTL or splice QTL does not itself establish mediation of a GWAS association. Bayesian colocalization was developed to distinguish shared causal architecture from coincident but genetically distinct associations within the same locus [21], whereas fine-mapping represents uncertainty through posterior inclusion probabilities and credible sets rather than equating the most significant SNP with causality [22,23]. Together, these approaches provide a principled route from regional association to candidate molecular mechanism.
Mendelian randomization offers a complementary layer of inference. Cis-acting variants associated with molecular traits can serve as genetic proxies to test whether genetically predicted perturbation of a candidate target is associated with disease susceptibility [24]. However, cis-MR remains vulnerable to weak or invalid instruments, linkage disequilibrium, horizontal pleiotropy, tissue specificity, lifelong genetic effects, and imperfect correspondence between genetic and pharmacological perturbation [24,25]. Crucially, an MR association does not prove that a molecular trait mediates a local GWAS signal unless the regional genetic architecture is also compatible. For complex immune loci, MR is therefore most persuasive when interpreted alongside fine-mapping and colocalization rather than in isolation [25].
Against this background, the chromosome 9 pediatric-sepsis locus remains mechanistically unresolved. The susceptibility association is established, but the effector gene, causal variant set, molecular phenotype, inflammatory state, and developmental context are uncertain [13,14,15,16,17,18]. It is unknown whether the signal acts through constitutive expression or emerges during innate immune activation; whether early and sustained inflammatory states reveal the same regulatory architecture; whether adult monocyte effects are reproduced in neonatal myeloid cells; whether expression and splicing implicate the same variant set; or whether apparently compelling regional regulatory associations genuinely share the pediatric-sepsis signal [19,20,21,22,23]. We therefore integrated disease fine-mapping with resting and stimulated monocyte eQTLs, neonatal myeloid regulatory data, exact-junction splicing evidence, signal-level colocalization, and prospectively specified cis-Mendelian randomization across five prespecified regional genes. Rather than assigning causality by genomic proximity, our objective was to distinguish disease-linked regulatory architecture from neighboring but genetically independent biology and thereby generate developmentally informed, experimentally testable molecular hypotheses [21,24,25].
2. Materials and Methods
2.1. Study Design and Prespecified Analytical Framework
We conducted a summary-level human genetic study to identify the regulatory gene or genes most likely to account for the chromosome 9 pediatric sepsis susceptibility signal centered on rs28361152. The analysis followed a prespecified, staged framework. First, we fine mapped the pediatric sepsis association and aligned regional variants to an external linkage disequilibrium (LD) reference. We then characterized cis-regulatory effects across adult monocyte activation states, neonatal myeloid cells, and splicing phenotypes. Next, we compared disease and molecular signals using LD-aware fine-mapping and colocalization. Finally, only molecular exposures and genetic instruments that satisfied exposure-side quality criteria were taken forward for Mendelian randomization (MR). Instrument selection was completed and frozen before pediatric sepsis outcome associations were accessed for MR estimation; accordingly, no instrument was selected, replaced, or optimized according to its association with the outcome.
The framework treated developmental stage and inflammatory stimulation as distinct biological contexts. Neonatal and adult datasets were therefore analyzed separately rather than pooled into a single exposure. This design allowed us to distinguish a shared disease-regulatory signal from a nearby but genetically independent regulatory association and to determine whether the evidence depended on age or immune activation state.
2.2. Pediatric Sepsis Locus and Candidate Genes
The disease association was defined using summary statistics from the Swiss Pediatric Sepsis Study genome-wide association analysis of culture-proven bacterial sepsis. The analytic dataset comprised 510 cases and 994 population-matched controls of European ancestry (N = 1,504). The lead variant, rs28361152, lies at chr9:111,727,529 on GRCh37 and chr9:108,965,249 on GRCh38. The primary regulatory interval extended 1 Mb on either side of the GRCh38 sentinel (chr9:107,965,249-109,965,249). Build-specific coordinates were retained in the analysis manifest throughout to prevent inadvertent mixing of genome builds.
Five genes carried forward from the original locus architecture and represented within the prespecified 2-Mb regional interval were fixed before downstream analysis: ELP1 (ENSG00000070061), CTNNAL1 (ENSG00000119326), FRRS1L (ENSG00000260230), ABITRAM (ENSG00000119328; formerly FAM206A), and TMEM245 (ENSG00000106771). The candidate set was locked before interrogation of pediatric-sepsis outcome associations; no gene was added or removed after outcome examination. This rule was intended to prevent post hoc nearest-gene selection. Physical proximity to rs28361152 was treated as locus context and, by itself, was not considered evidence of mediation.
2.3. Regulatory Datasets and Immune Cell Contexts
Adult innate-immune regulation was evaluated using Fairfax monocyte expression quantitative trait locus (eQTL) data indexed in the eQTL Catalogue. Four states were analyzed independently: resting monocytes (n = 420; QTD000081), interferon-γ (IFN-γ) stimulation for 24 hours (n = 370; QTD000082), lipopolysaccharide (LPS) stimulation for 2 hours (n = 256; QTD000083), and LPS stimulation for 24 hours (n = 325; QTD000084). Regional records were extracted across the prespecified 2-Mb interval and assigned to the five Ensembl gene identifiers. When a gene was represented by more than one probe, each probe-state combination was retained as a separate molecular exposure because premature probe aggregation could conceal distinct regulatory effects.
Developmental context was assessed using neonatal myeloid cis-eQTL data from E-MTAB-8977. Cord-blood-derived myeloid cells were studied at rest (n = 116) and after LPS stimulation (n = 125). Fine-mapping was performed for ABITRAM, CTNNAL1, and ELP1, for which analyzable neonatal regional data were available. Resting and stimulated models were fitted separately. Furthermore, neonatal estimates were not meta-analyzed with adult monocyte estimates because the developmental stage, cellular preparation, and stimulation protocols differed.
Because the pediatric sepsis locus had also been linked to ELP1 splicing, expression and splicing were treated as separate molecular phenotypes. ELP1 splicing was evaluated using GTEx version 10 SuSiE splicing-QTL summaries for the exact regional junction spanning GRCh38 positions 108,906,485-108,908,305. Thirteen exact-junction records across 11 tissues were reviewed. Where the archive provided posterior summaries without complete association statistics, disease-splicing concordance was assessed from credible-set membership and LD structure rather than formal H3/H4 colocalization. Adult macrophage data from Alasoo and colleagues were examined as supporting regulatory evidence when compatible variant-level information was available. (Supplementary Table S1).
2.4. Developmental- and Stimulus-Specific Comparisons
For each gene and cellular state, we recorded coverage of rs28361152 and neighboring variants, allele-oriented effect estimates, the strongest regional association, posterior inclusion probabilities (PIPs), and credible-set composition. Cross-state concordance required compatible allele orientation and regulatory direction together with a similar regional association pattern. Conversely, failure of the sentinel to regulate a gene was not interpreted as absence of regulation when another state-specific signal was present.
Neonatal fine-mapping was considered informative only when the fitted model converged and produced a localized 95% credible set that satisfied the prespecified within-set purity criterion. Converged models yielding no non-null credible set, very broad raw credible sets, or inadequate within-set LD were classified as unresolved and were excluded from colocalization and causal interpretation. Thus, an unresolved neonatal model was not treated as evidence that neonatal regulation was absent.
2.5. Genome Build Harmonization and Variant Quality Control
Disease, eQTL, splicing-QTL, and LD records were normalized to chromosome-position-reference-alternate identifiers while retaining rsIDs when available. GRCh38 regulatory variants were cross-referenced to GRCh37 using position, allele, and rsID information. The sentinel was represented as 9:111727529:T:C on GRCh37 and chr9:108965249_T_C on GRCh38. Variants with non-unique mappings, irreconcilable alleles, or unresolved strand orientation were excluded. Insertions and deletions were retained only when normalization and allele matching were unambiguous.
Because the analyses were conducted from summary statistics rather than discovery-cohort individual-level genotypes, LD was estimated externally from 503 participants in the 1000 Genomes Project Phase 3 European reference panel (CEU, FIN, GBR, IBS, and TSI), selected to match the European ancestry of the pediatric-sepsis GWAS. Variants were required to be polymorphic in this panel, and variants with European minor-allele count below 10 were excluded from primary LD-dependent analyses. Allele frequencies provided an additional orientation check. For every fine-mapping or correlated-instrument analysis, the summary-statistic vector and signed LD matrix were restricted to the same harmonized variants in identical order; matrices were screened for non-finite entries, duplicate or mismatched annotations, and allele inconsistencies before model fitting.
2.6. Fine-Mapping
The pediatric sepsis association and each eligible molecular phenotype were fine-mapped separately with SuSiE using summary association statistics and the corresponding signed 1000 Genomes European LD matrix. Models allowed up to 10 single-effect components, consistent with the locked fine-mapping outputs, and retained variant-level posterior inclusion probabilities (PIPs) and 95% credible sets. A fit was accepted only after model convergence, allele alignment, and exact variant-order checks had passed. Sensitivity fits varied the maximum number of SuSiE components when signal multiplicity was uncertain. Credible-set purity was assessed from within-set LD, including the minimum absolute pairwise correlation. Raw sets with very low within-set correlation were retained for audit but were not interpreted as localized signals.
2.7. Colocalization
Signal-level colocalization was performed with coloc using the SuSiE fits, thereby allowing comparison of multiple disease and molecular components. The posterior hypotheses were H0, neither trait associated; H1, disease only; H2, molecular phenotype only; H3, both traits associated through distinct causal variants; and H4, both traits associated through a shared causal variant. Primary priors were p1 = 1 × 10⁻⁴, p2 = 1 × 10⁻⁴, and p12 = 5 × 10⁻⁶. The shared-variant prior was varied to 1 × 10⁻⁶ and 1 × 10⁻⁵, with continuous prior-sensitivity assessment when necessary.
Colocalization was interpreted from the relative support for H3 and H4, stability across the prespecified prior range, credible-set composition, and local LD. No single PP.H4 cutoff was used in isolation. Strong sharing required H4 to predominate over H3 and to remain qualitatively stable across prior sensitivity analyses; supportive evidence denoted H4 predominance whose strength remained materially prior-sensitive; and H3 predominance indicated genetically distinct disease and regulatory signals. Analyses lacking sufficiently localized or compatible inputs were reported as not evaluable rather than assigned a biological conclusion. This classification framework was applied consistently across gene-state comparisons.
2.8. Genetic Instrument Selection and Outcome Harmonization
The MR inventory comprised 28 prespecified gene-probe-state exposures. Instrument selection used exposure data alone. For this locus-focused analysis, cis eligibility was restricted to variants available within the prespecified 2-Mb regional window for the corresponding gene-probe-state exposure. Candidate variants were required to be genome-wide significant (P < 5 × 10⁻⁸), strong (F ≥ 10, where F = β²/SE²), present in the European LD reference with minor-allele count ≥10, and LD-independent within an exposure at r² < 0.10. Eligible variants were ordered by ascending exposure P value, descending F statistic, and variant identifier. Exposure effects were then aligned to the 1000 Genomes alternate allele, and the complete instrument set was frozen before pediatric sepsis outcome extraction.
Frozen instruments were harmonized to the pediatric sepsis summary statistics by allele identity and effect direction. Palindromic variants were retained only when allele frequencies resolved orientation unambiguously; ambiguous palindromic variants, missing outcome variants, and allele-incompatible records were excluded with the reason recorded. No post-outcome replacement or proxy substitution was permitted. Reuse of the same variant across probes or states was audited explicitly and was not interpreted as independent replication; accordingly, MR tests sharing instruments were recognized as statistically dependent rather than independent pieces of genetic evidence.
2.9. Mendelian Randomization and Sensitivity Analyses
Single-instrument exposures were analyzed with the Wald ratio. Thirteen of the 14 evaluable exposures had a single usable cis instrument; the TMEM245 IFN-γ 24-hour exposure retained three instruments and was analyzed by inverse-variance weighting with multiplicative random effects (IVW-MRE). Estimates were expressed as odds ratios for pediatric sepsis with 95% confidence intervals. Cochran’s Q and its P value were calculated when more than one instrument was available, and the multiplicative overdispersion factor was reported. Conventional pleiotropy-robust estimators such as MR-Egger and weighted-median MR were not considered informative because the cis instrument sets were too small, particularly for the predominantly single-instrument exposures. All tests were two-sided. Benjamini-Hochberg adjustment was applied specifically to the family of 14 evaluable MR tests; a false-discovery-rate q value below 0.05 defined statistical significance. This MR multiplicity correction was not applied to the fine-mapping or colocalization analyses, which were evaluated using their prespecified posterior and sensitivity criteria. Because several MR exposures reused the same cis variant, the 14 tests were not statistically independent, and q values were interpreted in that context.
For the multi-instrument TMEM245 IFN-γ 24-hour exposure, sensitivity analyses included instrument-specific Wald estimates and leave-one-out IVW-MRE analyses. Moreover, a correlated-IVW model was fitted using the instrument variance-covariance matrix derived from signed European LD. Pairwise correlations, maximum r², the matrix condition number, Cochran’s Q, and the overdispersion factor were reviewed to assess whether residual instrument correlation or a single influential variant explained the association.
The MR assumptions of relevance, independence from exposure-outcome confounding, and absence of alternative variant-to-outcome pathways were considered alongside the regional genetic evidence. Because the latter assumptions cannot be verified from summary data alone, MR estimates were not interpreted as stand-alone proof of causality or locus mediation. In particular, a nominal or FDR-adjusted MR association without compatible colocalization was considered supportive of a gene-state association but insufficient to establish that the molecular exposure mediated the pediatric-sepsis GWAS signal.
2.10. Integrated Target Prioritization
Genes were adjudicated with a prespecified evidence hierarchy rather than a numerical score. The hierarchy considered disease-signal localization and credible-set structure; reproducible cis-expression or splice regulation; compatibility across developmental and inflammatory contexts; evidence for a shared rather than H3-dominant signal; availability and strength of a prospectively frozen cis-regulatory instrument; and consistency of the MR estimate with the preceding regulatory evidence. Accordingly, a gene could show strong stimulus-responsive biology yet remain genetically distinct from the pediatric-sepsis association. Conversely, a shared signal could be restricted to a particular developmental or inflammatory context.
Final classifications distinguished the leading disease-linked regulatory candidate, secondary MR-supported evidence with unresolved mediation, disease-distinct regulatory biology, disease-fine-map-discordant splicing biology, and insufficient localized evidence. This structure preserved informative negative or unresolved analyses without converting them into evidence of no biological effect.
2.11. Software Reproducibility and Ethical Considerations
Core analyses were performed in R and Python. Fine-mapping used susieR version 0.16.6, and colocalization used coloc version 5.2.3. Intermediate variant manifests, harmonized gene-state datasets, signed LD matrices, SuSiE fit objects, credible sets, colocalization outputs, instrument-level MR files, and quality-control summaries were retained as versioned analysis artifacts. Key pre-outcome, harmonization, MR, sensitivity, and final-adjudication outputs were frozen with analysis-lock files and SHA-256 checksums. MR reporting was structured with reference to STROBE-MR principles where applicable to a locus-focused summary-data analysis. All source datasets used in this study are publicly available from the cited studies and resources. Derived harmonized analysis files, analysis code, and audit artifacts are available from the corresponding author on reasonable request.
3. Results
3.1. The Pediatric Sepsis Signal Localized to an Rs28361152-Centered Credible Set
The regional analysis included 510 children with culture-proven bacterial sepsis and 994 population-matched controls (N = 1,504). All five prespecified genes were represented within the 2-Mb interval in each adult monocyte state, and rs28361152 was captured in every state. Depending on the stimulation condition, regional coverage comprised 8,732-8,866 unique variants for ELP1 and ABITRAM, 8,632-8,762 for CTNNAL1, 8,043-8,159 for FRRS1L, and 8,277-8,401 for TMEM245. After exact alignment to the 1000 Genomes European reference panel and exclusion of variants with European minor-allele count below 10, 6,255 variants entered the primary disease fine-map (Figure 1; Figure 2A).
SuSiE identified one active disease component and an 11-variant 95% credible set. rs28361152 carried most of the posterior probability (z = -6.293, P = 3.12 × 10⁻¹⁰; posterior inclusion probability [PIP] = 0.605). The remaining posterior mass was distributed chiefly across rs11788336 (PIP = 0.0465), rs7027263 (PIP = 0.0452), rs74384717 (PIP = 0.0448), rs113624314 (PIP = 0.0426), and rs744980 (PIP = 0.0426). The set also contained rs28361186 (PIP = 0.0231), which subsequently appeared within the stimulated ABITRAM regulatory architecture.
Reintroduction of 177 variants with European minor-allele counts of 1-9 increased the model to 6,432 variants but reproduced the same 11-member credible set (Jaccard index = 1.00). The PIP for rs28361152 changed by only 1.77 × 10⁻⁶, and no reintroduced low-frequency variant entered the credible set, indicating that the primary disease fine-map was stable to the minor-allele-count filter (Figure 2B).
3.2. The Susceptibility Allele Showed Stimulus-Dependent Regulatory Effects
At rs28361152, the direction and magnitude of the eQTL effects differed by gene and activation state (Figure 2C; Table 1; Supplementary Figure S2A–B). The sepsis-associated allele was associated with lower ABITRAM expression at rest (β = -0.181, SE = 0.0608; P = 0.00307), after 24-hour interferon-γ stimulation (β = -0.264, SE = 0.0643; P = 4.84 × 10⁻⁵), after 2-hour lipopolysaccharide stimulation (β = -0.279, SE = 0.0850; P = 0.00120), and after 24-hour lipopolysaccharide stimulation (β = -0.303, SE = 0.0681; P = 1.23 × 10⁻⁵). The direction was therefore consistent across states, while the largest effect occurred after sustained lipopolysaccharide exposure.
TMEM245 showed a different, activation-dependent pattern. The sentinel was not associated with TMEM245 expression in resting monocytes (β = 0.0277, SE = 0.0683; P = 0.685); however, a strong positive association emerged after 24-hour interferon-γ stimulation (β = 0.616, SE = 0.0645; P = 2.19 × 10⁻¹⁹) and remained evident after 2-hour (β = 0.369, SE = 0.0793; P = 5.36 × 10⁻⁶) and 24-hour (β = 0.481, SE = 0.0692; P = 2.06 × 10⁻¹¹) lipopolysaccharide stimulation. By contrast, rs28361152 was not materially associated with CTNNAL1 expression in any tested state (all P ≥ 0.184). FRRS1L showed no sentinel association at rest, after interferon-γ stimulation, or after 2-hour lipopolysaccharide exposure; a nominal association appeared after 24 hours of lipopolysaccharide stimulation (β = 0.252, SE = 0.0888; P = 0.00488).
Regulatory fine-mapping distinguished a sentinel eQTL association from localization of the underlying expression signal. For ABITRAM, credible sets were recovered at rest (34 variants), after interferon-γ stimulation (84 variants), and after 24-hour lipopolysaccharide exposure, but not after 2-hour exposure. In the final disease-regulatory harmonized LPS-24-hour analysis, the ABITRAM credible set contained 39 variants. All 11 variants in the pediatric-sepsis disease credible set were retained within this regulatory credible set (Jaccard index = 0.282). rs28361152 was not the leading regulatory variant; rs2275641 had the highest regulatory PIP in the harmonized LPS-24-hour model.
3.3. ABITRAM Showed the Strongest Shared Disease Regulatory Architecture
Formal colocalization separated the state-specific signals (Figure 2D; Figure 3A–D; Table 1; Supplementary Figure S2C–D; Supplementary Table S7). ABITRAM showed the strongest evidence of shared disease-regulatory architecture after 24-hour lipopolysaccharide stimulation (PP.H4 = 0.904; PP.H3 = 0.096). The interferon-γ 24-hour comparison also favored H4, although the magnitude of support remained sensitive to the shared-variant prior (PP.H4 = 0.716). By contrast, the resting ABITRAM signal was H3-dominant and therefore disease-distinct (PP.H3 = 0.985; PP.H4 = 0.015), while the 2-hour lipopolysaccharide state was not evaluable. Thus, the strongest evidence for shared genetic architecture was observed after sustained inflammatory stimulation; these colocalization results do not, by themselves, establish expression-mediated causality.
CTNNAL1 displayed strong regional cis regulation but a distinct disease-regulatory architecture. Formal comparison was feasible for both CTNNAL1 probes in all four monocyte states, with 5,577 variants shared between each regulatory model and the disease model. Every probe-state comparison was H3-dominant (PP.H3 ≈ 0.99974), with negligible support for H4 (PP.H4 ≈ 2.9 × 10⁻⁶). Varying the shared-signal prior from 1 × 10⁻⁶ to 1 × 10⁻⁵ did not change this classification.
For TMEM245, the evaluable 24-hour lipopolysaccharide comparison was also disease-distinct (PP.H3 = 0.99969; PP.H4 = 3.7 × 10⁻⁵). Colocalization was not evaluable for the interferon-γ, 2-hour lipopolysaccharide, or resting states. FRRS1L regulatory fine-mapping remained diffuse, so formal disease-regulatory colocalization was not evaluable in any state. These comparisons were retained as unresolved and were not interpreted as evidence for absence of regulation.
3.4. Prospectively Frozen Instrument Selection Yielded 14 Evaluable Mendelian Randomization Tests
The prespecified inventory contained 28 gene-probe-state exposures (Supplementary Figure S1; Supplementary Tables S2–S4). Fifteen exposures had at least one eligible exposure-side instrument, yielding 22 instrument records that were frozen before pediatric-sepsis outcome extraction. Outcome harmonization retained 16 records across 14 exposures, producing the final family of 14 evaluable MR tests. Six frozen records were excluded: four because the outcome variant was missing or allele-incompatible and two because palindromic orientation remained ambiguous. No excluded instrument was replaced with a proxy or post-outcome alternative. Thirteen of the 14 evaluable exposures were analyzed with a single cis instrument; TMEM245 after 24-hour interferon-γ stimulation retained three instruments.
The ABITRAM 24-hour lipopolysaccharide exposure, which showed the strongest colocalization with pediatric sepsis, was not evaluable by MR because its prospectively frozen instrument, rs2289480, was absent from the pediatric-sepsis outcome dataset; under the prespecified plan, no replacement instrument was introduced. The ABITRAM interferon-γ 24-hour instrument produced an odds ratio (OR) of 1.086 (95% confidence interval [CI] 1.000-1.179; P = 0.0509; false-discovery-rate [FDR] q = 0.178), whereas the resting estimate was null (OR = 1.029, 95% CI 0.939-1.127; P = 0.542; q = 0.977). None of the eight evaluable CTNNAL1 probe-state analyses remained associated after FDR correction; the ORs were close to 1.00 and all q values were 0.977 (Table 2). Several tests reused the same cis instrument across probes or states and therefore did not constitute independent replications. TMEM245 estimates after 2-hour and 24-hour lipopolysaccharide stimulation were nominally inverse (OR = 0.916 and 0.909, respectively; both P = 0.0195), but neither survived Benjamini-Hochberg correction across the 14 MR tests (both q = 0.0912) (Figure 4A–B; Table 2; Supplementary Table S5). The 24-hour lipopolysaccharide regulatory signal was disease-distinct and was therefore not considered evidence of mediation for that state.
3.5. TMEM245 After Interferon-γ Stimulation Was the Only FDR-significant MR Association
Higher genetically predicted TMEM245 expression after 24-hour interferon-γ stimulation was associated with lower pediatric sepsis susceptibility in the three-instrument inverse-variance-weighted multiplicative random-effects model (OR = 0.941, 95% CI 0.908-0.975; P = 8.11 × 10⁻⁴; q = 0.0114) (Figure 4A–D; Table 2; Supplementary Table S5). This was the only association that remained significant after Benjamini-Hochberg correction across the 14 evaluable MR tests. There was no statistical evidence of between-instrument heterogeneity (Q = 1.438, df = 2; P = 0.487), and the multiplicative overdispersion factor was 1.00.
All three instrument-specific Wald estimates were inverse: OR = 0.957 (95% CI 0.909-1.007; P = 0.0936) for rs7022699, OR = 0.914 (95% CI 0.862-0.970; P = 0.00285) for rs838820, and OR = 0.956 (95% CI 0.871-1.048; P = 0.332) for rs141184426. The direction remained inverse in every leave-one-out estimate (Supplementary Table S6; Supplementary Figure S3). However, omission of rs838820 produced the greatest attenuation and crossed the nominal significance threshold (OR = 0.9566, 95% CI 0.9146-1.0005; P = 0.0528), whereas omission of either remaining instrument retained nominal significance.
As a sensitivity analysis, modeling residual correlation between the three instruments yielded a concordant correlated-IVW estimate (OR = 0.938, 95% CI 0.910-0.967; P = 4.48 × 10⁻⁵), with Q = 1.014 (df = 2; P = 0.602) and an overdispersion factor of 1.00. The maximum pairwise |r| was 0.308, the maximum r² was 0.0947, and the LD matrix condition number was 5.52. Because interferon-γ 24-hour colocalization was not evaluable, the MR evidence supported TMEM245 as a secondary candidate but did not establish that TMEM245 expression mediated the pediatric-sepsis association.
3.6. ELP1 Splicing Was Reproducible but Discordant with the Disease Fine Map
The exact ELP1 junction at GRCh38 chr9:108,906,485-108,908,305 was represented by 13 records across 11 GTEx tissues, with one credible set in each tissue (Figure 5A; Table 1; Supplementary Table S9; Supplementary Figure S4D). The recurrent splice architecture included rs838825 and rs838838. However, rs28361152 was absent from every exact-junction credible set (0 of 11), and the recurrent splice haplotype did not align with the pediatric-sepsis fine-map. Complete variant-level effect estimates were unavailable from the posterior-only archive; therefore, formal H3/H4 colocalization could not be calculated. The result was classified as reproducible splice-QTL architecture that was discordant with the pediatric-sepsis fine-map.
3.7. Neonatal Models Did Not Provide a Purity-Qualified Regulatory Signal
All six neonatal myeloid models converged, including resting and lipopolysaccharide-stimulated analyses for ABITRAM, CTNNAL1, and ELP1 (Figure 5B–C; Table 1; Supplementary Tables S8A–B; Supplementary Figure S4A–C). Nevertheless, none produced a purity-qualified credible set. The largest model PIPs were 0.00892 and 0.01865 for ABITRAM at rest and after lipopolysaccharide stimulation, 0.00411 and 0 for CTNNAL1, and 0.00429 and 0 for ELP1. The stimulated CTNNAL1 and ELP1 fits yielded no non-null credible set.
Across the four non-null models, 24 raw 95% credible sets were recovered, containing approximately 2,400-3,700 variants with near-zero minimum absolute within-set correlations. These sets failed the within-set purity requirement used in the analysis and were excluded from colocalization and integrated adjudication. The neonatal analyses were therefore classified as unresolved; they did not provide evidence for absence of neonatal regulation.
3.8. Integrated Evidence Prioritized ABITRAM and Retained TMEM245 as a Secondary Candidate
The frozen adjudication integrated the disease fine-map, adult and neonatal regulatory localization, colocalization, MR using prospectively frozen exposure-side instruments, and sensitivity analyses without assigning a numerical score (Figure 5D; Table 3; Supplementary Table S10). ABITRAM was classified as the leading disease-linked regulatory candidate because its 24-hour lipopolysaccharide signal showed the strongest shared architecture with pediatric sepsis. An MR estimate could not be obtained for that state under the frozen instrument plan.
TMEM245 was classified as a secondary MR-supported candidate. Its interferon-γ 24-hour association survived FDR correction and remained directionally concordant in leave-one-out and correlated-IVW sensitivity analyses; however, mediation remained unresolved because formal colocalization was not evaluable. CTNNAL1 was classified as strong regulatory biology with a disease-distinct architecture. ELP1 showed reproducible splicing architecture that was discordant with the pediatric-sepsis fine-map. FRRS1L had insufficient localized genetic evidence for either colocalization or MR. These classifications summarize the prespecified evidence hierarchy without equating locus-linked evidence with demonstrated causal mediation.
4. Discussion
4.1. Principal Findings: Immune-State Resolution Refines the Biology of the Pediatric Sepsis Locus
This study advances the recently identified chromosome 9 pediatric-sepsis susceptibility locus from regional association toward a more genetically resolved model of host susceptibility. The original Swiss Pediatric Sepsis Study established the CTNNAL1–ELP1 region as a genome-wide significant locus in culture-proven bacterial sepsis but could not identify the relevant effector gene or molecular phenotype [13]. By integrating disease fine-mapping with immune-state-specific cis-regulation, neonatal myeloid regulation, exact-junction splicing, signal-level colocalization, and prospectively specified cis-Mendelian randomization, we identified a coherent hierarchy among the five prespecified genes. ABITRAM emerged as the leading disease-linked regulatory candidate; TMEM245 provided an FDR-significant, directionally consistent MR-supported signal whose mediation remains unresolved; CTNNAL1 and ELP1 showed biologically credible regional regulation that was genetically discordant with the disease signal; and FRRS1L remained insufficiently localized for mechanistic assignment.
The principal contribution is therefore not simply the nomination of another gene at a GWAS locus, but the demonstration that inherited susceptibility and its molecular readout must be interpreted in biological context. The disease signal was comparatively stable: rs28361152 dominated an 11-variant credible set, and the credible-set composition was unchanged after low-frequency variants were reintroduced. In contrast, the surrounding regulatory architecture differed across resting, IFN-γ-stimulated, early-LPS, and sustained-LPS conditions. This separation between a stable germline susceptibility signal and a context-dependent regulatory readout is especially pertinent to sepsis, where host-response heterogeneity is recognized across pediatric endotypes, precision-immunology frameworks, adult genomic response signatures, multicohort transcriptomic classifications, and clinically derived sepsis phenotypes [26,27,28,29,30,31,32].
4.2. ABITRAM Reveals a State-Dependent Disease-Linked Regulatory Signal During Sustained Innate Immune Activation
The strongest locus-linked evidence concerned ABITRAM. The sepsis-associated rs28361152 allele was associated with lower ABITRAM expression in every tested adult-monocyte state, with the largest effect after 24 hours of LPS exposure. More importantly, the relationship between ABITRAM regulation and the disease signal changed qualitatively across states: the resting eQTL was disease-distinct, whereas the sustained-LPS signal strongly favored shared genetic architecture (PP.H4=0.904). In the final harmonized LPS-24-hour analysis, the ABITRAM credible set contained 39 variants and retained all 11 variants from the pediatric-sepsis credible set (Jaccard index=0.282). IFN-γ stimulation provided additional, although prior-sensitive, evidence for sharing. Accordingly, this locus cannot be interpreted adequately from resting-cell regulation alone.
This state dependence is consistent with the broader response-eQTL literature. Fairfax et al. demonstrated that monocyte stimulation with LPS or IFN-γ exposes regulatory effects that are absent or substantially altered at baseline, with additional differences between early and sustained stimulation [15]. Lee et al. independently showed that common variants modify pathogen- and interferon-responsive transcription in dendritic cells [16]. Mechanistically, Alasoo et al. showed that inherited differences in enhancer accessibility can precede stimulation and shape subsequent transcriptional responses, providing a plausible framework for regulatory effects that become apparent only after immune activation [17]. Taken together, these observations support a proposed model of latent regulatory susceptibility: a disease-associated haplotype may appear weakly informative, or even genetically distinct, in quiescent cells yet become genetically aligned with the disease architecture when a relevant inflammatory program is engaged.
The concentration of ABITRAM sharing after sustained rather than early LPS stimulation is particularly notable. Pediatric septic shock is temporally dynamic; molecular endotypes can change during the first days of illness, and those transitions alter both prognosis and apparent treatment response [26]. Likewise, adult sepsis studies have identified genomic response signatures that differ in immune suppression, endotoxin tolerance, HLA expression, and outcome, while regulatory effects themselves can vary across host-response states [29,30]. Consequently, the 24-hour LPS result should not be interpreted as merely a larger resting eQTL. Rather, it identifies a distinct inflammatory context in which ABITRAM regulation is most closely aligned with the inherited pediatric-sepsis susceptibility signal.
Nevertheless, the present data do not establish that reduced ABITRAM expression is itself the causal intermediate. Colocalization supports shared genetic architecture, not expression-mediated causality, and there is not yet a mature experimental literature directly linking ABITRAM to sepsis. That limited prior biological expectation is informative rather than problematic: post-GWAS functional genetics is most valuable when concordant genetic evidence can prioritize an under-characterized effector over a more familiar neighboring gene without relying on proximity or narrative plausibility. ABITRAM should therefore be regarded as a high-priority mechanistic hypothesis for experimental validation, not as a validated therapeutic target.
4.3. TMEM245 Defines a Second Genetic Axis: FDR-Significant MR Evidence with Unresolved Locus Mediation
TMEM245 produced a distinctly different pattern. rs28361152 had essentially no effect on TMEM245 expression in resting monocytes, whereas strong positive regulatory effects emerged after IFN-γ and LPS stimulation. Higher genetically predicted TMEM245 expression after 24-hour IFN-γ stimulation was associated with lower pediatric-sepsis susceptibility (OR=0.941, 95% CI 0.908-0.975; q=0.0114), and this was the only MR association that survived correction across the 14 evaluable exposures. The effect direction was inverse for all three instruments and remained inverse in every leave-one-out analysis; an LD-aware correlated-IVW model yielded a closely concordant estimate. However, omission of rs838820 reduced precision and attenuated the nominal P value to 0.0528, indicating that the strength of statistical evidence was not uniform across all leave-one-out models.
These sensitivity analyses therefore support directional consistency but do not justify treating the association as independent of instrument influence. More importantly, the interpretation must remain narrower than the statistical result. Formal colocalization for the IFN-γ state was not evaluable, whereas the evaluable TMEM245 LPS-24-hour signal was strongly disease-distinct. Zuber et al. emphasized that MR and colocalization address related but non-equivalent questions: a non-zero MR association can arise when exposure and outcome are influenced by distinct but correlated variants within the same region [33]. Consequently, TMEM245 is best classified as a secondary genetically supported susceptibility or target hypothesis, not as the demonstrated mediator of rs28361152.
This distinction is consequential for translation. Human genetic support can inform therapeutic target selection, and genetically supported drug mechanisms have been associated with higher probabilities of clinical development success [24,25,34,35]. Nevertheless, that advantage depends on correct gene assignment, valid instruments, relevant biological context, and concordance between regional genetics and the proposed molecular mechanism. In the present locus, TMEM245 warrants further study in IFN-γ-responsive monocytes, but replication of the regulatory signal and demonstration of compatible colocalization should precede any claim of locus mediation. Functional experiments can then determine whether altered TMEM245 expression influences antimicrobial effector functions, cytokine signaling, metabolic adaptation, phagocytosis, or cell survival. Until those steps are completed, the MR result should not be translated into therapeutic efficacy.
4.4. Biological Plausibility Is Not Genetic Mediation: Discordant and Unresolved Evidence for CTNNAL1, ELP1, and FRRS1L
The discordant and unresolved findings are equally important to the mechanistic interpretation of this locus. CTNNAL1 is biologically attractive because its encoded protein, alpha-catulin, can bind IKK-β, augment NF-κB activity, influence migration, and modulate resistance to apoptosis [36]. Such biology could readily encourage proximity-based prioritization in an inflammatory disorder. Yet rs28361152 was not materially associated with CTNNAL1 expression in the tested monocyte states, and every evaluable CTNNAL1 probe-state comparison overwhelmingly favored distinct disease and regulatory signals. Thus, the genetic evidence argues against mediation through the CTNNAL1 cis-expression signals tested here, despite credible inflammatory biology.
ELP1 provides a different lesson. For the exact splice junction examined, the splice architecture was reproducible across 11 GTEx tissues, yet rs28361152 was absent from every credible set and the recurrent splice haplotype did not align with the pediatric-sepsis fine-map. This discordance should not be interpreted as evidence that ELP1 splicing is biologically irrelevant. Alternative splicing is an important regulatory layer of human gene control [19,20], and sepsis-specific genome-wide analyses have identified widespread splicing changes linked to immune-cell regulatory networks [37]. Rather, the present findings demonstrate that reproducible molecular biology within a GWAS interval can remain genetically orthogonal to the disease association under study.
FRRS1L represents a third category: unresolved localization. Its regulatory signal remained too diffuse for formal colocalization or MR, and this failure to localize may reflect insufficient signal resolution rather than absence of biological regulation. The integrated classification therefore preserves four distinct evidence states across the locus: disease-linked regulation, MR-supported but unresolved mediation, disease-distinct regulation, and unresolved localization. This hierarchy is more informative than a binary candidate/noncandidate label and illustrates why post-GWAS interpretation should resist both nearest-gene assumptions and single-assay overinterpretation.
4.5. Developmental Immunogenomics: Neonatal Uncertainty Is Informative, Not Negative
Neonatal analyses require similarly careful interpretation. All six neonatal myeloid models converged, yet none generated a purity-qualified credible set. The raw credible sets were exceptionally broad, and the stimulated CTNNAL1 and ELP1 models yielded no non-null credible set. Accordingly, the available neonatal data did not localize the pediatric-sepsis susceptibility signal with sufficient precision for colocalization. They therefore do not demonstrate absence of neonatal regulation.
That distinction is biologically important. Transcriptomic studies have repeatedly shown that the septic host response differs across neonates, infants, children, and adults [10,12], while neonatal eQTL mapping demonstrates that immune regulation depends strongly on both developmental stage and stimulation state [18]. Contemporary pediatric sepsis reviews likewise emphasize age-specific epidemiology, clinical context, and long-term consequences, reinforcing that pediatric sepsis should not be assumed to mirror adult disease [38]. Furthermore, recent pediatric host-response RNA-sequencing has shown that gene-expression signatures can identify infection type and organ dysfunction in children, underscoring the potential clinical importance of developmentally appropriate molecular phenotyping [39].
The appropriate conclusion is therefore developmental uncertainty, not developmental equivalence. Adult monocyte data provide a biologically informative context for locus interpretation, but they cannot be assumed to reproduce neonatal myeloid regulation. Importantly, the present analysis cannot distinguish true developmental divergence from insufficient neonatal fine-mapping power. Larger neonatal and early-childhood regulatory datasets, ideally with denser genotype data, ancestry-matched LD, single-cell resolution, and time-resolved stimulation, will be required to determine whether the activated ABITRAM architecture is present from birth, emerges later in development, or is replaced by a different regulatory mechanism in early life.
4.6. Genetic Triangulation: Fine-Mapping, Colocalization, and cis-MR Answer Different Questions
The study was deliberately structured so that no single statistical method was asked to answer every causal question. Fine-mapping identifies variants that remain plausible contributors to an association; colocalization evaluates whether disease and molecular traits are compatible with the same underlying causal signal; and MR tests whether genetically predicted variation in a molecular exposure is associated with disease. These questions are complementary, but they are not interchangeable [14,21,22,23,24,25,33].
The present locus illustrates why that distinction matters. ABITRAM has the strongest evidence for disease sharing but lacks an evaluable MR estimate in its strongest LPS-24-hour state because the prospectively frozen instrument was absent from the outcome dataset. TMEM245 has the strongest MR association, yet mediation remains unresolved because the relevant colocalization could not be evaluated. CTNNAL1 has robust cis-regulatory biology but compelling evidence for a distinct disease architecture, whereas ELP1 has reproducible splicing with disease-fine-map discordance. No single observation, considered alone, would have produced the same hierarchy as the integrated analysis.
Methodological restraint is therefore integral to the inference. The decision not to replace the missing ABITRAM instrument preserved the prespecified instrument-selection framework, even though it left the strongest colocalizing state without an MR estimate. Likewise, non-evaluable colocalization and low-purity neonatal credible sets were retained as unresolved rather than converted into negative biological claims. In a field vulnerable to target over-prioritization, an informative ‘not evaluable’ result is preferable to post hoc optimization that yields a more complete but less credible causal narrative.
4.7. From Susceptibility Genetics to Precision Sepsis Biology
The broader implications extend beyond this chromosome 9 interval. Sepsis is a biologically heterogeneous syndrome in which similar clinical presentations can reflect markedly different host-response states. Adult studies have identified sepsis response signatures and genomic endotypes with reproducible differences in immune suppression, coagulation, metabolism, and mortality [29,30,31]. Clinical clustering likewise yields distinct sepsis phenotypes with different biomarker profiles, outcomes, and potential treatment effects [32]. More recently, single-cell profiling has resolved cell-specific immune states in bacterial sepsis that are obscured in bulk blood measurements [40], while global multicohort analyses have shown that transcriptionally defined endotypes recur across geographically diverse populations [28].
These observations have a direct pediatric counterpart. Molecular endotypes in pediatric septic shock are dynamic rather than fixed, and their transitions track changing risk and corticosteroid response [26]. Biomarker-based enrichment models such as PERSEVERE and PERSEVERE-II further demonstrate that biologically informed stratification can estimate mortality risk and support precision-trial design in children [41,42]. Subsequent work integrating endothelial and host-response biomarkers has improved prediction of death or persistent multiple-organ dysfunction, while risk-stratified analyses suggest that treatment associations can differ across pediatric biological risk groups [43,44]. Thus, precision sepsis medicine increasingly depends on matching the biological measurement to the question being asked: susceptibility, prognosis, mechanism, or treatment response.
Our findings add a germline-regulatory layer to that framework. If validated, we propose that the clinically relevant interpretation of inherited susceptibility at this locus may require genotype to be considered together with regulatory target, immune state, and developmental stage. This concept is consistent with contemporary precision-immunotherapy models, which argue that broad immunomodulation has repeatedly failed in sepsis partly because patients with fundamentally different immune states were treated as though they shared a single mechanism [27,45]. Germline genetics should not replace dynamic biomarkers; rather, the two may ultimately provide complementary information, with inherited variation defining predisposition and acquired host-response states describing the biological context in which that predisposition is expressed.
4.8. Strengths and Limitations
Several features strengthen the present study. The disease association was derived from culture-proven pediatric bacterial sepsis rather than an administrative or broadly syndromic phenotype. Candidate genes and analytical stages were prespecified, and physical proximity was explicitly not treated as evidence of mediation. Build-aware and allele-aware harmonization preceded LD-dependent inference; the disease fine-map was stable to reintroduction of low-frequency variants; and adult regulatory phenotypes were analyzed by immune state rather than pooled. Most importantly, molecular signals were interpreted through shared versus distinct regional architecture rather than from sentinel eQTL P values alone.
The MR framework was similarly conservative. Exposure-side instruments were selected and frozen before pediatric-sepsis outcome associations were accessed; missing or incompatible instruments were not replaced post hoc; residual correlation in the multi-instrument TMEM245 model was explicitly modeled; and leave-one-out estimates were directionally concordant. These design choices reduce avoidable sources of bias and make the distinction between ABITRAM and TMEM245 more credible. Nevertheless, the predominantly cis-regulatory design also imposes important constraints: 13 of the 14 evaluable exposures were represented by a single instrument, and several tests reused the same cis variant across probes or states.
There are, nevertheless, important limitations. First, the pediatric GWAS is modest by contemporary complex-trait standards and is restricted to European ancestry, limiting both fine-mapping resolution and generalizability. LD was therefore estimated from an ancestry-matched but external 1000 Genomes European reference rather than from study-specific genotypes, leaving some potential for reference-panel mismatch. Second, adult monocyte eQTLs are experimental proxies rather than measurements from children during active sepsis. Third, the neonatal datasets were underpowered for localized fine-mapping at this locus. Fourth, several biologically important comparisons were not formally evaluable, including TMEM245 colocalization after IFN-γ stimulation and exact-junction ELP1 colocalization. Fifth, the strongest ABITRAM state lacked an outcome-side MR instrument under the frozen analysis plan. Sixth, because most MR analyses were single-instrument Wald-ratio estimates, conventional pleiotropy-robust estimators such as MR-Egger and weighted-median MR were not informative; moreover, reused instruments mean that the 14 MR tests were not statistically independent. Finally, MR represents lifelong germline perturbation and cannot be translated quantitatively into the magnitude, timing, reversibility, efficacy, or safety of acute pharmacological modulation during sepsis. These constraints are especially important in a syndrome in which both immune state and treatment response can change rapidly over time [26,27,45].
4.9. Experimental Validation and Future Directions
The evidence hierarchy yields a focused experimental agenda. For ABITRAM, the first priority is allele- and haplotype-aware perturbation in primary monocytes or macrophages under resting, early-LPS, and sustained-LPS conditions. CRISPR interference or activation, targeted enhancer perturbation, chromatin-accessibility profiling, H3K27ac mapping, transcription-factor occupancy, and chromosome-conformation assays could determine whether disease-credible-set variants physically and functionally regulate ABITRAM. Single-cell RNA sequencing after perturbation would be particularly valuable for establishing whether the effect is confined to a specific monocyte state.
For TMEM245, replication should precede mechanistic escalation. The IFN-γ-responsive eQTL and the inverse MR association should first be reproduced in a larger stimulated-monocyte resource, followed by formal colocalization to determine whether the regulatory and pediatric-sepsis signals share causal architecture. Only then should functional studies test whether altered TMEM245 expression modifies bacterial killing, cytokine output, phagocytosis, metabolic rewiring, or survival under inflammatory challenge. This sequence is important because higher genetically predicted expression was associated with lower susceptibility, yet locus mediation remains unresolved.
For CTNNAL1, the experimental question is different. Its relationship with IKK-β/NF-κB provides genuine biological plausibility [36], but the genetics argue strongly that the tested cis-regulatory signal is not the disease-driving signal at this locus. Functional experiments should therefore ask whether CTNNAL1 modifies downstream sepsis biology independently of the rs28361152 susceptibility mechanism. Likewise, ELP1 should be investigated as a potentially relevant splicing-regulatory gene without conflating recurrent splice biology with mediation of the pediatric-sepsis GWAS signal.
Ultimately, the most informative next-generation resource would combine larger multi-ancestry pediatric sepsis GWAS with age-stratified single-cell response-eQTL, chromatin-QTL, splice-QTL, and proteomic-QTL data. Such datasets would permit the locus to be traced across variant, chromatin, transcript, protein, cellular phenotype, and clinical susceptibility. That level of integration will be required to move from statistical prioritization to experimentally grounded precision medicine.
4.10. Conclusions
This study refines an ambiguous pediatric-sepsis susceptibility locus into a biologically differentiated set of hypotheses. ABITRAM emerged as the leading disease-linked regulatory candidate, with strong colocalization concentrated after sustained LPS stimulation, supporting an inflammation-dependent rather than purely constitutive model of locus function. TMEM245 provided a separate FDR-significant and directionally consistent MR signal after IFN-γ stimulation; however, the absence of evaluable colocalization prevents attribution of the GWAS association to TMEM245 expression. Conversely, CTNNAL1 and ELP1 demonstrate that compelling regional biology does not necessarily identify the mediator of a disease locus, while FRRS1L illustrates the importance of preserving unresolved evidence as unresolved.
More broadly, the findings support a conditional model of pediatric-sepsis genetics in which the functional interpretation of a susceptibility haplotype depends on molecular phenotype, immune-cell state, duration of inflammatory exposure, and potentially developmental stage. Integrating these dimensions through fine-mapping, colocalization, and prospectively specified cis-MR provides a more disciplined route from genomic association to experimentally testable mechanism. Replication and functional validation are required before these candidates can be advanced toward therapeutic inference; nevertheless, this framework may help connect inherited susceptibility with the emerging precision-medicine architecture of pediatric sepsis.
Supplementary Materials
The following supporting information can be downloaded at the website of this paper posted on Preprints.org, Supplementary Figures S1–S4 and Supplementary Tables S1–S10, covering the prespecified MR exposure framework, regulatory and colocalization analyses, TMEM245 sensitivity analyses, neonatal and ELP1 splicing analyses, source datasets, instrument selection and harmonization, MR results, fine-mapping, and integrated gene prioritization.
Author Contributions
Conceptualization, A.A.A. and M.S.U.; methodology, M.S.U.; formal analysis, M.S.U.; investigation, A.A.A. and M.S.U.; data curation, M.S.U.; validation, M.S.U.; visualization, M.S.U.; writing—original draft preparation, A.A.A. and M.S.U.; writing—review and editing, A.A.A. and M.S.U.; supervision, M.S.U.; project administration, A.A.A. and M.S.U. All authors have read and agreed to the published version of the manuscript.
Funding
This research received no external funding.
Institutional Review Board Statement
This study involved secondary analysis of previously collected, de-identified, summary-level genetic and molecular data obtained from publicly available resources and published studies. No new participants were recruited, and no new identifiable human data were collected. Ethical approval and institutional review procedures for the original contributing studies were reported by their respective investigators. Accordingly, no additional institutional review board approval was required for the present secondary analysis.
Informed Consent Statement
No new informed consent was obtained for the present study because it was based exclusively on de-identified, summary-level data from previously published studies and public genomic resources. Informed consent for participation in the original studies was obtained by the respective investigators in accordance with their approved study protocols.
Data Availability Statement
The datasets analyzed in this study were obtained from publicly available repositories and previously published studies. Pediatric sepsis genome-wide association summary statistics from the Swiss Pediatric Sepsis Study are available through the NHGRI-EBI GWAS Catalog (accession GCST90726424) and Zenodo (DOI: 10.5281/zenodo.15675718). Adult monocyte cis-eQTL data were obtained from the EMBL-EBI eQTL Catalogue, including the Fairfax datasets QTD000081–QTD000084. Neonatal myeloid cis-eQTL data were obtained from ArrayExpress (accession E-MTAB-8977). ELP1 splicing-QTL fine-mapping data were obtained from the GTEx Portal, Analysis V10, including publicly available SuSiE sQTL resources. Linkage disequilibrium estimates were derived from the 1000 Genomes Project Phase 3 European reference populations. Derived harmonized datasets, analysis scripts, quality-control outputs, fine-mapping results, colocalization results, Mendelian randomization files, and audit artifacts generated for this study are available from the corresponding author on reasonable request, subject to the terms of use of the original data providers.
Acknowledgments
The authors gratefully acknowledge the participants, investigators, and data-generating consortia whose publicly available genomic and molecular resources contributed to this study, including the Swiss Pediatric Sepsis Study, the eQTL Catalogue, the Genotype-Tissue Expression Project, neonatal myeloid regulatory datasets, and the 1000 Genomes Project.
Conflicts of Interest
The authors declare no conflicts of interest.
References
- Schlapbach, L.J.; Watson, R.S.; Sorce, L.R.; et al. International Consensus Criteria for Pediatric Sepsis and Septic Shock. JAMA 2024, 331(8), 665–674. [Google Scholar] [CrossRef] [PubMed]
- Nadel, S.; Newport, M.J.; Booy, R.; Levin, M. Variation in the tumor necrosis factor-alpha gene promoter region may be associated with death from meningococcal disease. J. Infect. Dis. 1996, 174(4), 878–880. [Google Scholar] [CrossRef] [PubMed]
- Westendorp, R.G.J.; Langermans, J.A.M.; Huizinga, T.W.J.; et al. Genetic influence on cytokine production and fatal meningococcal disease. Lancet 1997, 349, 170–173. [Google Scholar] [CrossRef] [PubMed]
- Holmes, C.L.; Russell, J.A.; Walley, K.R. Genetic polymorphisms in sepsis and septic shock: role in prognosis and potential for therapy. Chest 2003, 124(3), 1103–1115. [Google Scholar] [CrossRef] [PubMed]
- Wong, H.R.; Shanley, T.P.; Sakthivel, B.; et al. Genome-level expression profiles in pediatric septic shock indicate a role for altered zinc homeostasis in poor outcome. Physiol. Genom. 2007, 30(2), 146–155. [Google Scholar] [CrossRef] [PubMed]
- Shanley, T.P.; Cvijanovich, N.; Lin, R.; et al. Genome-level longitudinal expression of signaling pathways and gene networks in pediatric septic shock. Mol. Med. 2007, 13(9-10), 495–508. [Google Scholar] [CrossRef] [PubMed]
- Wong, H.R.; Cvijanovich, N.; Allen, G.L.; et al. Genomic expression profiling across the pediatric systemic inflammatory response syndrome, sepsis, and septic shock spectrum. Crit. Care Med. 2009, 37(5), 1558–1566. [Google Scholar] [CrossRef] [PubMed]
- Wong, H.R.; Cvijanovich, N.; Lin, R.; et al. Identification of pediatric septic shock subclasses based on genome-wide expression profiling. BMC Med. 2009, 7, 34. [Google Scholar] [CrossRef] [PubMed]
- Ardura, M.I.; Banchereau, R.; Mejias, A.; et al. Enhanced monocyte response and decreased central memory T cells in children with invasive Staphylococcus aureus infections. PLoS ONE 2009, 4(5), e5446. [Google Scholar] [CrossRef] [PubMed]
- Wynn, J.L.; Cvijanovich, N.Z.; Allen, G.L.; et al. The influence of developmental age on the early transcriptomic response of children with septic shock. Mol. Med. 2011, 17(11-12), 1146–1156. [Google Scholar] [CrossRef] [PubMed]
- Kwan, A.; Hubank, M.; Rashid, A.; Klein, N.; Peters, M.J. Transcriptional instability during evolving sepsis may limit biomarker based risk stratification. PLoS ONE 2013, 8(3), e60501. [Google Scholar] [CrossRef] [PubMed]
- Raymond, S.L.; López, M.C.; Baker, H.V.; et al. Unique transcriptomic response to sepsis is observed among patients of different age groups. PLoS ONE 2017, 12(9), e0184159. [Google Scholar] [CrossRef] [PubMed]
- Lawless, D.; Hodel, F.A.; Thorball, C.W.; et al. Swiss Pediatric Sepsis Study Group. Genome-wide association study of paediatric bacteraemia and sepsis. EBioMedicine 2026, 129, 106320. [Google Scholar] [CrossRef] [PubMed]
- Schaid, D.J.; Chen, W.; Larson, N.B. From genome-wide associations to candidate causal variants by statistical fine-mapping. Nat. Rev. Genet. 2018, 19(8), 491–504. [Google Scholar] [CrossRef] [PubMed]
- Fairfax, B.P.; Humburg, P.; Makino, S.; et al. Innate immune activity conditions the effect of regulatory variants upon monocyte gene expression. Science 2014, 343(6175), 1246949. [Google Scholar] [CrossRef] [PubMed]
- Lee, M.N.; Ye, C.; Villani, A.C.; et al. Common genetic variants modulate pathogen-sensing responses in human dendritic cells. Science 2014, 343(6175), 1246980. [Google Scholar] [CrossRef] [PubMed]
- Alasoo, K.; Rodrigues, J.; Mukhopadhyay, S.; et al. Shared genetic effects on chromatin and gene expression indicate a role for enhancer priming in immune response. Nat. Genet. 2018, 50(3), 424–431. [Google Scholar] [CrossRef] [PubMed]
- Huang, Q.Q.; Tang, H.H.F.; Teo, S.M.; et al. Neonatal genetics of gene expression reveal potential origins of autoimmune and allergic disease risk. Nat. Commun. 2020, 11(1), 3761. [Google Scholar] [CrossRef] [PubMed]
- GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science 2020, 369(6509), 1318–1330. [Google Scholar] [CrossRef] [PubMed]
- Kerimov, N.; Hayhurst, J.D.; Peikova, K.; et al. A compendium of uniformly processed human gene expression and splicing quantitative trait loci. Nat. Genet. 2021, 53(9), 1290–1299. [Google Scholar] [CrossRef] [PubMed]
- Giambartolomei, C.; Vukcevic, D.; Schadt, E.E.; et al. Bayesian test for colocalisation between pairs of genetic association studies using summary statistics. PLoS Genet. 2014, 10(5), e1004383. [Google Scholar] [CrossRef] [PubMed]
- Wang, G.; Sarkar, A.; Carbonetto, P.; Stephens, M. A simple new approach to variable selection in regression, with application to genetic fine-mapping. J. R Stat. Soc. Ser. B Stat. Methodol. 2020, 82(5), 1273–1300. [Google Scholar] [CrossRef] [PubMed]
- Zou, Y.; Carbonetto, P.; Wang, G.; Stephens, M. Fine-mapping from summary data with the “Sum of Single Effects” model. PLoS Genet. 2022, 18(7), e1010299. [Google Scholar] [CrossRef] [PubMed]
- Daghlas, I.; Gill, D. Mendelian randomization as a tool to inform drug development using human genetics. Camb. Prism Precis Med. 2023, 1, e16. [Google Scholar] [CrossRef] [PubMed]
- Burgess, S.; Mason, A.M.; Grant, A.J.; et al. Using genetic association data to guide drug discovery and development: review of methods and applications. Am. J. Hum. Genet. 2023, 110(2), 195–214. [Google Scholar] [CrossRef] [PubMed]
- Wong, H.R.; Cvijanovich, N.Z.; Anas, N.; et al. Endotype Transitions During the Acute Phase of Pediatric Septic Shock Reflect Changing Risk and Treatment Response. Crit. Care Med. 2018, 46(3), e242–e249. [Google Scholar] [CrossRef] [PubMed]
- Giamarellos-Bourboulis, E.J.; Aschenbrenner, A.C.; Bauer, M.; et al. The pathophysiology of sepsis and precision-medicine-based immunotherapy. Nat. Immunol. 2024, 25, 19–28. [Google Scholar] [CrossRef] [PubMed]
- Chenoweth, J.G.; Brandsma, J.; Striegel, D.A.; et al. Sepsis endotypes identified by host gene expression across global cohorts. Commun. Med. (Lond) 2024, 4, 120. [Google Scholar] [CrossRef] [PubMed]
- Davenport, E.E.; Burnham, K.L.; Radhakrishnan, J.; et al. Genomic landscape of the individual host response and outcomes in sepsis: a prospective cohort study. Lancet Respir. Med. 2016, 4(4), 259–271. [Google Scholar] [CrossRef] [PubMed]
- Scicluna, B.P.; van Vught, L.A.; Zwinderman, A.H.; Consortium, M.A.R.S.; et al. Classification of patients with sepsis according to blood genomic endotype: a prospective cohort study. Lancet Respir. Med. 2017, 5(10), 816–826. [Google Scholar] [CrossRef] [PubMed]
- Sweeney, T.E.; Azad, T.D.; Donato, M.; et al. Unsupervised Analysis of Transcriptomics in Bacterial Sepsis Across Multiple Datasets Reveals Three Robust Clusters. Crit. Care Med. 2018, 46(6), 915–925. [Google Scholar] [CrossRef] [PubMed]
- Seymour, C.W.; Kennedy, J.N.; Wang, S.; et al. Derivation, Validation, and Potential Treatment Implications of Novel Clinical Phenotypes for Sepsis. JAMA 2019, 321(20), 2003–2017. [Google Scholar] [CrossRef] [PubMed]
- Zuber, V.; Grinberg, N.F.; Gill, D.; et al. Combining evidence from Mendelian randomization and colocalization: review and comparison of approaches. Am. J. Hum. Genet. 2022, 109(5), 767–782. [Google Scholar] [CrossRef] [PubMed]
- Nelson, M.R.; Tipney, H.; Painter, J.L.; et al. The support of human genetic evidence for approved drug indications. Nat. Genet. 2015, 47(8), 856–860. [Google Scholar] [CrossRef] [PubMed]
- King, E.A.; Davis, J.W.; Degner, J.F. Are drug targets with genetic support twice as likely to be approved? Revised estimates of the impact of genetic support for drug mechanisms on the probability of drug approval. PLoS Genet. 2019, 15(12), e1008489. [Google Scholar] [CrossRef] [PubMed]
- Wiesner, C.; Winsauer, G.; Resch, U.; et al. Alpha-catulin, a Rho signalling component, can regulate NF-κB through binding to IKK-β, and confers resistance to apoptosis. Oncogene 2008, 27(15), 2159–2169. [Google Scholar] [CrossRef] [PubMed]
- Tuerdimaimaiti, D.; Abuduaini, B.; Kang, S.; et al. Genome-wide identification and functional analysis of dysregulated alternative splicing profiles in sepsis. J. Inflamm. (Lond) 2023, 20(1), 31. [Google Scholar] [CrossRef] [PubMed]
- Watson, R.S.; Carrol, E.D.; Carter, M.J.; Kissoon, N.; Ranjit, S.; Schlapbach, L.J. The burden and contemporary epidemiology of sepsis in children. Lancet Child Adolesc. Health 2024, 8(9), 670–681. [Google Scholar] [CrossRef] [PubMed]
- Schlapbach, L.J.; Ganesamoorthy, D.; Wilson, C.; et al. Host gene expression signatures to identify infection type and organ dysfunction in children evaluated for sepsis: a multicentre cohort study. Lancet Child Adolesc. Health 2024, 8(5), 325–338. [Google Scholar] [CrossRef] [PubMed]
- Reyes, M.; Filbin, M.R.; Bhattacharyya, R.P.; et al. An immune-cell signature of bacterial sepsis. Nat. Med. 2020, 26(3), 333–340. [Google Scholar] [CrossRef] [PubMed]
- Wong, H.R.; Salisbury, S.; Xiao, Q.; et al. The pediatric sepsis biomarker risk model. Crit. Care 2012, 16(5), R174. [Google Scholar] [CrossRef] [PubMed]
- Wong, H.R.; Caldwell, J.T.; Cvijanovich, N.Z.; et al. Prospective clinical testing and experimental validation of the Pediatric Sepsis Biomarker Risk Model. Sci. Transl. Med. 2019, 11(518), eaax9000. [Google Scholar] [CrossRef] [PubMed]
- Atreya, M.R.; Cvijanovich, N.Z.; Fitzgerald, J.C.; et al. Integrated PERSEVERE and endothelial biomarker risk model predicts death and persistent MODS in pediatric septic shock: a secondary analysis of a prospective observational study. Crit. Care 2022, 26(1), 210. [Google Scholar] [CrossRef] [PubMed]
- Klowak, J.A.; Bijelić, V.; Barrowman, N.; Menon, K. Genomics of Pediatric Septic Shock Investigators. The Association of Corticosteroids and Pediatric Sepsis Biomarker Risk Model (PERSEVERE)-II Biomarker Risk Stratification With Mortality in Pediatric Septic Shock. Pediatr. Crit. Care Med. 2023, 24(3), 186–193. [Google Scholar] [CrossRef] [PubMed]
- Hotchkiss, R.S.; Monneret, G.; Payen, D. Sepsis-induced immunosuppression: from cellular dysfunctions to immunotherapy. Nat. Rev. Immunol. 2013, 13(12), 862–874. [Google Scholar] [CrossRef] [PubMed]
Figure 1.
Developmental regulatory genetics framework for the pediatric sepsis susceptibility locus. A, A Swiss genome-wide association study of culture-proven bacterial sepsis included 510 cases and 994 population-matched controls. Fine-mapping of 6,255 variants identified one active component, with rs28361152 showing the highest posterior inclusion probability (PIP = 0.605). The five prespecified genes are shown schematically. B, Regulatory evidence was evaluated across resting and stimulated adult monocytes, neonatal myeloid cells, and GTEx ELP1 splicing data. C, Disease fine-mapping and colocalization were completed as separate locus-resolution analyses; exposure-side MR instruments were then selected and frozen before pediatric-sepsis outcome extraction for MR estimation. The frozen exposure-side set comprised 22 instrument records across 15 exposures; after outcome harmonization, 16 records across 14 exposures remained for 14 MR tests. Estimates were derived using Wald-ratio or multiplicative random-effects inverse-variance-weighted models, with Benjamini–Hochberg false-discovery-rate correction and prespecified sensitivity analyses. D, Evidence was integrated hierarchically across disease fine-mapping, colocalization, Mendelian randomization, and sensitivity analyses. ABITRAM emerged as the leading locus-linked candidate, whereas TMEM245 represented a secondary signal with unresolved mediation. Abbreviation: eQTL, expression quantitative trait locus; FDR, false-discovery rate; GTEx, Genotype-Tissue Expression project; IFN-γ, interferon-γ; LD, linkage disequilibrium; LPS, lipopolysaccharide; MR, Mendelian randomization; PIP, posterior inclusion probability.
Figure 1.
Developmental regulatory genetics framework for the pediatric sepsis susceptibility locus. A, A Swiss genome-wide association study of culture-proven bacterial sepsis included 510 cases and 994 population-matched controls. Fine-mapping of 6,255 variants identified one active component, with rs28361152 showing the highest posterior inclusion probability (PIP = 0.605). The five prespecified genes are shown schematically. B, Regulatory evidence was evaluated across resting and stimulated adult monocytes, neonatal myeloid cells, and GTEx ELP1 splicing data. C, Disease fine-mapping and colocalization were completed as separate locus-resolution analyses; exposure-side MR instruments were then selected and frozen before pediatric-sepsis outcome extraction for MR estimation. The frozen exposure-side set comprised 22 instrument records across 15 exposures; after outcome harmonization, 16 records across 14 exposures remained for 14 MR tests. Estimates were derived using Wald-ratio or multiplicative random-effects inverse-variance-weighted models, with Benjamini–Hochberg false-discovery-rate correction and prespecified sensitivity analyses. D, Evidence was integrated hierarchically across disease fine-mapping, colocalization, Mendelian randomization, and sensitivity analyses. ABITRAM emerged as the leading locus-linked candidate, whereas TMEM245 represented a secondary signal with unresolved mediation. Abbreviation: eQTL, expression quantitative trait locus; FDR, false-discovery rate; GTEx, Genotype-Tissue Expression project; IFN-γ, interferon-γ; LD, linkage disequilibrium; LPS, lipopolysaccharide; MR, Mendelian randomization; PIP, posterior inclusion probability.

Figure 2.
Disease fine-mapping and cell-state-specific regulatory architecture. A, Bayesian fine-mapping identified one active component and an 11-variant 95% credible set; rs28361152 was the leading variant (PIP = 0.605). B, Posterior estimates were stable after re-inclusion of low-frequency variants (credible-set Jaccard index = 1.00; change in sentinel PIP = 1.77 × 10^−6; maximum low-MAC PIP = 5.48 × 10^−7). C, Effect estimates for rs28361152 on gene expression in resting, IFN-γ–stimulated, and LPS-stimulated adult monocytes, reported as β coefficients with 95% confidence intervals. Regulatory effects varied by gene and activation state. D, Posterior probabilities of a shared causal variant (PP.H4) between pediatric sepsis and each regulatory phenotype. Sharing was strongest for ABITRAM after 24-h LPS stimulation (PP.H4 = 0.904) and was supportive after 24-h IFN-γ stimulation (PP.H4 = 0.716); other evaluable gene–state combinations were disease-distinct. Grey cells indicate comparisons that were not formally evaluable. Abbreviations: CI, confidence interval; CS, credible set; IFN-γ, interferon-γ; LPS, lipopolysaccharide; MAC, minor allele count; NE, not evaluable; PIP, posterior inclusion probability; PP.H4, posterior probability that both traits share a causal variant.
Figure 2.
Disease fine-mapping and cell-state-specific regulatory architecture. A, Bayesian fine-mapping identified one active component and an 11-variant 95% credible set; rs28361152 was the leading variant (PIP = 0.605). B, Posterior estimates were stable after re-inclusion of low-frequency variants (credible-set Jaccard index = 1.00; change in sentinel PIP = 1.77 × 10^−6; maximum low-MAC PIP = 5.48 × 10^−7). C, Effect estimates for rs28361152 on gene expression in resting, IFN-γ–stimulated, and LPS-stimulated adult monocytes, reported as β coefficients with 95% confidence intervals. Regulatory effects varied by gene and activation state. D, Posterior probabilities of a shared causal variant (PP.H4) between pediatric sepsis and each regulatory phenotype. Sharing was strongest for ABITRAM after 24-h LPS stimulation (PP.H4 = 0.904) and was supportive after 24-h IFN-γ stimulation (PP.H4 = 0.716); other evaluable gene–state combinations were disease-distinct. Grey cells indicate comparisons that were not formally evaluable. Abbreviations: CI, confidence interval; CS, credible set; IFN-γ, interferon-γ; LPS, lipopolysaccharide; MAC, minor allele count; NE, not evaluable; PIP, posterior inclusion probability; PP.H4, posterior probability that both traits share a causal variant.

Figure 3.
Activated ABITRAM shares the pediatric sepsis susceptibility signal. A, Regional association profiles for pediatric sepsis and the ABITRAM LPS-24-hour eQTL across 5,624 harmonized variants. rs28361152 was the leading disease variant (PIP = 0.605), whereas rs2275641 had the highest regulatory PIP. B, Comparison of disease and regulatory association strength. Colors distinguish variants outside the credible sets, variants confined to the ABITRAM credible set, and variants shared by both credible sets; point size reflects the larger SuSiE PIP. C, PIP architecture of the 39-variant ABITRAM credible set. All 11 disease credible-set variants were retained in the regulatory credible set (Jaccard index = 0.282); gold outlines identify disease credible-set members, and stars denote rs28361152. D, Continuous sensitivity of PP.H3 and PP.H4 to the shared-association prior \(p_{12}\). At the prespecified primary prior \(p_{12}=5\times10^{-6}\), PP.H4 was 0.904 for LPS-24-hour, 0.716 for IFN-γ–24-hour, and 0.015 for resting monocytes. LPS-2-hour regulation was not evaluable because no purity-qualified credible set was identified. Colocalization supports shared genetic architecture but does not establish expression-mediated causality. Abbreviations: CS, credible set; eQTL, expression quantitative trait locus; IFN-γ, interferon-γ; LPS, lipopolysaccharide; PIP, posterior inclusion probability.
Figure 3.
Activated ABITRAM shares the pediatric sepsis susceptibility signal. A, Regional association profiles for pediatric sepsis and the ABITRAM LPS-24-hour eQTL across 5,624 harmonized variants. rs28361152 was the leading disease variant (PIP = 0.605), whereas rs2275641 had the highest regulatory PIP. B, Comparison of disease and regulatory association strength. Colors distinguish variants outside the credible sets, variants confined to the ABITRAM credible set, and variants shared by both credible sets; point size reflects the larger SuSiE PIP. C, PIP architecture of the 39-variant ABITRAM credible set. All 11 disease credible-set variants were retained in the regulatory credible set (Jaccard index = 0.282); gold outlines identify disease credible-set members, and stars denote rs28361152. D, Continuous sensitivity of PP.H3 and PP.H4 to the shared-association prior \(p_{12}\). At the prespecified primary prior \(p_{12}=5\times10^{-6}\), PP.H4 was 0.904 for LPS-24-hour, 0.716 for IFN-γ–24-hour, and 0.015 for resting monocytes. LPS-2-hour regulation was not evaluable because no purity-qualified credible set was identified. Colocalization supports shared genetic architecture but does not establish expression-mediated causality. Abbreviations: CS, credible set; eQTL, expression quantitative trait locus; IFN-γ, interferon-γ; LPS, lipopolysaccharide; PIP, posterior inclusion probability.

Figure 4.
Mendelian-randomization estimates and robustness analyses. A, Odds ratios and 95% confidence intervals for the 14 evaluable expression exposures. Only IFN-γ–responsive TMEM245 remained associated with pediatric sepsis after multiplicity correction. B, Comparison of nominal and Benjamini–Hochberg–adjusted evidence; point size represents the number of instruments. The TMEM245 IFN-γ 24-h association survived false-discovery-rate correction (q = 0.0114). C, Individual Wald-ratio, inverse-variance-weighted, leave-one-out, and correlated-instrument estimates for TMEM245. Effect direction was consistent across analyses; Q statistics provided no evidence of between-instrument heterogeneity (independent IVW: Q = 1.438, P = 0.487; correlated IVW: Q = 1.014, P = 0.602). D, Instrument-level associations between variant effects on TMEM245 expression and pediatric-sepsis log-odds. Independent and correlated IVW models produced concordant inverse slopes; pairwise instrument correlation was modest (maximum |r| = 0.308; maximum r² = 0.0947; condition number = 5.52). Abbreviation: BH, Benjamini–Hochberg; CI, confidence interval; FDR, false-discovery rate; IFN-γ, interferon-γ; IVW, inverse-variance weighted; LPS, lipopolysaccharide; MR, Mendelian randomization; Q, Cochran’s heterogeneity statistic.
Figure 4.
Mendelian-randomization estimates and robustness analyses. A, Odds ratios and 95% confidence intervals for the 14 evaluable expression exposures. Only IFN-γ–responsive TMEM245 remained associated with pediatric sepsis after multiplicity correction. B, Comparison of nominal and Benjamini–Hochberg–adjusted evidence; point size represents the number of instruments. The TMEM245 IFN-γ 24-h association survived false-discovery-rate correction (q = 0.0114). C, Individual Wald-ratio, inverse-variance-weighted, leave-one-out, and correlated-instrument estimates for TMEM245. Effect direction was consistent across analyses; Q statistics provided no evidence of between-instrument heterogeneity (independent IVW: Q = 1.438, P = 0.487; correlated IVW: Q = 1.014, P = 0.602). D, Instrument-level associations between variant effects on TMEM245 expression and pediatric-sepsis log-odds. Independent and correlated IVW models produced concordant inverse slopes; pairwise instrument correlation was modest (maximum |r| = 0.308; maximum r² = 0.0947; condition number = 5.52). Abbreviation: BH, Benjamini–Hochberg; CI, confidence interval; FDR, false-discovery rate; IFN-γ, interferon-γ; IVW, inverse-variance weighted; LPS, lipopolysaccharide; MR, Mendelian randomization; Q, Cochran’s heterogeneity statistic.

Figure 5.
Developmental evidence and integrated five-gene adjudication. (A) Cross-tissue audit of the ELP1 exact-junction signal identified 13 records across 11 tissues, with one credible set per tissue. rs838825 and rs838838 marked the recurrent splice-associated haplotype, whereas rs28361152 was absent from every credible set (0/11). Only verified tissue-level membership is shown; unavailable tissue-by-variant assignments were not inferred. (B) Maximum model and rs28361152 PIPs across six neonatal myeloid models. Filled and open symbols denote maximum and sentinel PIPs, respectively; marker shape distinguishes resting and LPS-stimulated states. The ELP1 and CTNNAL1 LPS models yielded no non-null credible set. (C) Size and minimum absolute within-set correlation of 24 raw neonatal 95% credible sets. None met the prespecified purity threshold, and the two corresponding stimulated models yielded no credible set. (D) Integrated matrix of the locked five-gene adjudication. ABITRAM was the leading disease-linked regulatory candidate; TMEM245 represented a secondary MR-supported signal with unresolved mediation. CTNNAL1 showed disease-distinct regulatory activity, ELP1 showed reproducible but disease-discordant splicing, and FRRS1L lacked sufficient localized genetic evidence. The panel summarizes the developmental evidence and the final prespecified five-gene classification. CS, credible set; FDR+, significant after Benjamini–Hochberg false-discovery-rate correction; IFN-γ, interferon-γ; LPS, lipopolysaccharide; MR, Mendelian randomization; NE, not evaluable; PIP, posterior inclusion probability.
Figure 5.
Developmental evidence and integrated five-gene adjudication. (A) Cross-tissue audit of the ELP1 exact-junction signal identified 13 records across 11 tissues, with one credible set per tissue. rs838825 and rs838838 marked the recurrent splice-associated haplotype, whereas rs28361152 was absent from every credible set (0/11). Only verified tissue-level membership is shown; unavailable tissue-by-variant assignments were not inferred. (B) Maximum model and rs28361152 PIPs across six neonatal myeloid models. Filled and open symbols denote maximum and sentinel PIPs, respectively; marker shape distinguishes resting and LPS-stimulated states. The ELP1 and CTNNAL1 LPS models yielded no non-null credible set. (C) Size and minimum absolute within-set correlation of 24 raw neonatal 95% credible sets. None met the prespecified purity threshold, and the two corresponding stimulated models yielded no credible set. (D) Integrated matrix of the locked five-gene adjudication. ABITRAM was the leading disease-linked regulatory candidate; TMEM245 represented a secondary MR-supported signal with unresolved mediation. CTNNAL1 showed disease-distinct regulatory activity, ELP1 showed reproducible but disease-discordant splicing, and FRRS1L lacked sufficient localized genetic evidence. The panel summarizes the developmental evidence and the final prespecified five-gene classification. CS, credible set; FDR+, significant after Benjamini–Hochberg false-discovery-rate correction; IFN-γ, interferon-γ; LPS, lipopolysaccharide; MR, Mendelian randomization; NE, not evaluable; PIP, posterior inclusion probability.

Table 1.
Regulatory, colocalization, and developmental evidence across the five-gene locus.
| Gene | Adult regulation | Disease sharing | Neonatal evidence | Interpretation |
| ABITRAM | Lower expression in all four monocyte states; strongest after LPS 24 h (β=-0.303; P=1.23×10⁻⁵). | Shared: LPS 24 h (PP.H4=0.904); supportive/prior-sensitive: IFN-γ 24 h (PP.H4=0.716); distinct: resting (PP.H3=0.985); LPS 2 h NE. | No purity-qualified CS; maximum PIP 0.00892 (resting) and 0.01865 (LPS). | Leading disease-linked candidate; sharing concentrated in sustained LPS activation. |
| TMEM245 | Activation-dependent effects: IFN-γ 24 h β=0.616 (P=2.19×10⁻¹⁹); LPS 2 h β=0.369 (P=5.36×10⁻⁶); LPS 24 h β=0.481 (P=2.06×10⁻¹¹). | LPS 24 h disease-distinct (PP.H3=0.999691; PP.H4=0.000037); other states NE. | Not assessed in the six neonatal models. | Secondary MR-supported candidate; mediation unresolved. |
| CTNNAL1 | Reproducible cis regulation across probes and states; rs28361152 was not significant. | All evaluable probe-state comparisons were distinct (PP.H3≈0.999740; PP.H4≈2.86×10⁻⁶). | No purity-qualified CS; maximum PIP 0.00411 (resting) and 0 (LPS). | Strong regulatory biology; disease distinct. |
| ELP1 | Exact-junction splicing: 13 records across 11 tissues; recurrent variants rs838825 and rs838838. | Formal colocalization NE; rs28361152 absent from all 11 CS; splice architecture disease discordant. | No purity-qualified CS; maximum PIP 0.00429 (resting) and 0 (LPS). | Reproducible splicing biology; disease-fine-map discordant. |
| FRRS1L | Nominal rs28361152 association after LPS 24 h (β=0.252; P=0.00488); fine-mapping remained diffuse. | NE because no localized regulatory signal was obtained. | Not assessed in the six neonatal models. | Insufficient localized evidence for mediation or MR. |
Note. Adult sentinel effects refer to rs28361152. PP.H3 and PP.H4 denote distinct and shared causal variants, respectively. All six neonatal models converged, but none yielded a purity-qualified 95% credible set. CS, credible set; IFN-γ, interferon-γ; LPS, lipopolysaccharide; MR, Mendelian randomization; NE, not evaluable; PIP, posterior inclusion probability.
Table 2.
Mendelian randomization estimates for the 14 evaluable exposures.
| Exposure | IVs / method | OR (95% CI) | P | BH q | Q P | Colocalization |
| ABITRAM IFN-γ 24 h ILMN_2191929 | 1 / Wald | 1.086 (1.000-1.179) | 0.0509 | 0.178 | NA | Supportive; prior-sensitive |
|
ABITRAM Resting ILMN_2191929 |
1 / Wald | 1.029 (0.939-1.127) | 0.542 | 0.977 | NA | Disease-distinct |
|
TMEM245 IFN-γ 24 h ILMN_1788384 |
3 / IVW-MRE | 0.941 (0.908-0.975) | 0.000811 | 0.0114 | 0.487 | NE |
|
TMEM245 LPS 24 h ILMN_1788384 |
1 / Wald | 0.909 (0.838-0.985) | 0.0195 | 0.0912 | NA | Disease-distinct |
|
TMEM245 LPS 2 h ILMN_1788384 |
1 / Wald | 0.916 (0.851-0.986) | 0.0195 | 0.0912 | NA | NE |
|
TMEM245 Resting ILMN_1788384 |
1 / Wald | 1.002 (0.903-1.111) | 0.977 | 0.977 | NA | NE |
|
CTNNAL1 IFN-γ 24 h ILMN_1721901 |
1 / Wald | 1.015 (0.961-1.073) | 0.589 | 0.977 | NA | Disease-distinct |
|
CTNNAL1 IFN-γ 24 h ILMN_2136446 |
1 / Wald | 0.999 (0.954-1.045) | 0.949 | 0.977 | NA | Disease-distinct |
|
CTNNAL1 LPS 24 h ILMN_1721901 |
1 / Wald | 0.998 (0.946-1.053) | 0.949 | 0.977 | NA | Disease-distinct |
|
CTNNAL1 LPS 24 h ILMN_2136446 |
1 / Wald | 0.998 (0.948-1.051) | 0.949 | 0.977 | NA | Disease-distinct |
|
CTNNAL1 LPS 2 h ILMN_1721901 |
1 / Wald | 0.998 (0.948-1.051) | 0.949 | 0.977 | NA | Disease-distinct |
|
CTNNAL1 LPS 2 h ILMN_2136446 |
1 / Wald | 0.998 (0.942-1.057) | 0.949 | 0.977 | NA | Disease-distinct |
| CTNNAL1 Resting ILMN_1721901 | 1 / Wald | 1.002 (0.952-1.055) | 0.926 | 0.977 | NA | Disease-distinct |
|
CTNNAL1 Resting ILMN_2136446 |
1 / Wald | 1.001 (0.954-1.050) | 0.967 | 0.977 | NA | Disease-distinct |
Foot Note. ORs are per unit increase in genetically predicted expression. Benjamini-Hochberg correction covered all 14 tests; Q applies only to the three-instrument TMEM245 IFN-γ 24 h model. ABITRAM LPS 24 h was not evaluable because the frozen instrument rs2289480 was absent from the outcome dataset and was not replaced. BH, Benjamini-Hochberg; CI, confidence interval; IFN-γ, interferon-γ; IV, instrumental variable; IVW-MRE, inverse-variance-weighted multiplicative random-effects; LPS, lipopolysaccharide; NA, not applicable; NE, not evaluable; OR, odds ratio.
Table 3.
Integrated evidence adjudication and final classification of the five candidate genes.
| Gene | Disease sharing | MR evidence | Key constraint | Final classification |
| ABITRAM | Shared after LPS 24 h (PP.H4=0.904); supportive/prior-sensitive after IFN-γ 24 h; resting disease-distinct. | LPS 24 h NE. IFN-γ 24 h: OR=1.086 (1.000-1.179), P=0.0509, q=0.178. | Frozen LPS 24 h instrument rs2289480 was absent from the outcome GWAS; no replacement was permitted. | Leading disease-linked regulatory candidate; strongest shared state lacked an evaluable MR estimate. |
| TMEM245 | LPS 24 h disease-distinct (PP.H3=0.999691; PP.H4=0.000037); IFN-γ 24 h NE. | IFN-γ 24 h: OR=0.941 (0.908-0.975), P=0.000811, q=0.0114; correlated IVW OR=0.938 (0.910-0.967). | Direction was consistent, but omission of rs838820 attenuated P to 0.0528; colocalization remained NE. | Secondary MR-supported candidate; mediation unresolved. |
| CTNNAL1 | All evaluable probe-state comparisons were disease-distinct (PP.H3≈0.999740; negligible PP.H4). | No persuasive association across eight evaluable exposures. | Several tests reused the same cis instrument and were not independent replications. | Strong regulatory biology; disease distinct. |
| ELP1 | Exact-junction splice architecture was disease-discordant; rs28361152 was absent from all 11 CS. | No genome-wide-significant Fairfax expression instrument; MR NE. | Discordance was consistent across 11 GTEx tissues. | Reproducible splicing biology; disease-fine-map discordant. |
| FRRS1L | Disease-regulatory sharing NE because regulatory fine-mapping remained diffuse. | No eligible genome-wide-significant cis-eQTL instrument. | Mismatch-aware fine-mapping did not localize a signal. | Insufficient localized genetic evidence for mediation or MR. |
Note. Classifications integrate disease fine-mapping, regulatory localization, colocalization, developmental evidence, MR using prospectively frozen exposure-side instruments, and prespecified sensitivity analyses. Categories are qualitative and do not establish therapeutic efficacy or safety. CS, credible set; eQTL, expression quantitative trait locus; GWAS, genome-wide association study; IFN-γ, interferon-γ; IVW, inverse-variance weighted; LPS, lipopolysaccharide; MR, Mendelian randomization; NE, not evaluable; PP.H3/PP.H4, posterior probabilities for distinct/shared causal variants.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.