Multiomics integration and causal inference in aging research—can the molecular gap be bridged?

Integration of under-studied molecular layers

Building on this multi-layer framework, previous research on aging-related QTL has predominantly concentrated on eQTLs^[7]. In the present study, the integration of sQTLs, apaQTLs, pQTLs, and mQTLs uncovers previously unrecognized biological insights. In total, Gai et al. identified 30 genes, 11 splicing events, 5 proteins, 3 APA events, and 39 metabolites associated with aging-related traits, including the metabolite oleoyl-linoleoyl-glycerol, which was negatively associated with longevity and showed unfavorable relationships with high-density lipoprotein and low-density lipoprotein^[8]. The OPERA analysis further identified the microtubule-associated protein tau locus, a region well known for its association with neurodegenerative diseases, as a colocalization locus involving three types of molecular phenotypes (posterior probability of association > 0.7), suggesting that its regulatory landscape may also influence longevity and epigenetic aging. Cross-tissue analysis reported by Gai et al.^[8] revealed that CASP8 was positively associated with longevity in the spleen, lung, and heart, and secretagogin (SCGN) showed a significant positive SMR association with Intrinsic Epigenetic Age Acceleration in adipose and thyroid tissues, with age-associated downregulation of SCGN expression. SORT1 was positively associated with longevity specifically in the liver but not in whole blood, highlighting the importance of tissue context. Of particular significance, the gene FBXL5 demonstrated a robust association with longevity exclusively through apaQTLs, with no corresponding association observed via eQTLs. This finding suggests a potential role for alternative polyadenylation as a regulatory mechanism in lifespan determination that has hitherto been insufficiently acknowledged, although functional validation is required. The absence of a corresponding eQTL signal does not by itself prove APA-mediated causality, and experimental approaches such as 3’ RACE, APA reporter assays, clustered regularly interspaced short palindromic repeats (CRISPR) editing of the polyadenylation signal, and RNA stability measurements are needed to validate this hypothesis. Consistent with this, APA-mediated 3’ UTR lengthening has been observed in cellular senescence, and APA is increasingly recognized as a key modulator of gene expression in age-related pathologies^[9]. It is important to note that no single QTL layer can capture the full regulatory architecture of aging. For example, eQTLs are often tissue-specific and may not reflect protein abundance or post-transcriptional regulation^[10]; sQTLs and apaQTLs are highly context-dependent and remain under-studied in aging^[7]; pQTLs and mQTLs are influenced by sample source, platform, and ancestry^[11]. In addition, these QTL datasets differ substantially in sample size, age distribution, sex composition, genetic ancestry, tissue or cell type, and experimental platform^[8,12]. For example, eQTLGen is based on whole blood (n = 31,684), BrainMeta on brain (n = 2,865), GTEx apaQTL on whole blood (n = 670), deCODE pQTL on plasma (n = 35,559), and Canadian longitudinal study on aging mQTL on metabolomics (n = 8,299). Most datasets are predominantly of European ancestry and are not uniformly designed for aging-specific analyses^[13]. Such heterogeneity can affect LD structure, instrument strength, winner’s curse, sample overlap, tissue-specific causal inference, and external validity, thereby complicating cross-layer integration and the portability of causal inference^[14,15]. Nevertheless, these cross-layer signals provide valuable candidate directions for future functional and epidemiological studies.

Causal inference across molecular QTL layers: a tiered framework and its caveats

In molecular epidemiological investigations of longevity and epigenetic aging, a primary challenge lies in differentiating putative causal relationships from confounding factors, reverse causation, or spurious signals arising from LD imbalance^[16,17]. In this context, Gai et al. have made notable methodological contributions by systematically integrating multi-level molecular QTL data with aging-related GWAS datasets, and by implementing a tiered statistical causal inference framework^[8].

Initially, the GSMR approach was applied, incorporating LD clustering and the removal of HEIDI outliers to minimize the confounding effects of horizontal pleiotropy^[18,19]. Subsequently, for each significant association identified, up to nine alternative Mendelian randomization (MR) methods were employed for cross-validation. This multi-method consensus approach substantially mitigates the risk of false positives that may arise from reliance on a single MR model assumption^[19,20]. Of particular note is the stringent criterion applied in the SMR analysis, wherein genes and proteins were retained only if they satisfied a HEIDI test p-value greater than 0.01, thereby effectively distinguishing genuine pleiotropic effects from chance associations driven solely by LD^[18].

This tiered analytical framework—progressing from genetic correlation assessment, through primary GSMR analysis, to sensitivity verification via multiple MR methods, and culminating in HEIDI filtering—establishes a rigorous analytical paradigm for aging research. It enhances the causal interpretability and reliability of subsequently identified candidate targets, including 30 genes, 11 splicing events, and 39 metabolites, surpassing the inferential power of simple correlation or co-localization analyses. However, these methods rest on several strong assumptions. GSMR and SMR require genetic variants to be valid instruments—associated with the molecular phenotype, independent of confounders, and affecting the outcome only through the phenotype. Horizontal pleiotropy, LD, winner’s curse, and sample overlap can violate these assumptions and bias effect estimates^[14,15,21]; for instance, simulations have shown that GSMR type I error rates are inflated when sample overlap increases, and GSMR may ignore weak pleiotropic effects, potentially leading to inflated type I errors^[22]. HEIDI reduces LD-driven false positives but can be conservative and may not fully exclude pleiotropy; the test can output NA when its assumptions are violated, and its power to distinguish causality from pleiotropy is limited when causal variants are in close to perfect LD^[23]. OPERA’s posterior probability of association depends on model priors and should be viewed as evidence for shared causal variants rather than proof of causality; the associations identified can occur even if the causal variants for the molecular phenotypes are distinct from, but in LD with, the causal variants for the trait, and OPERA only provides gene-level tests without pinpointing specific variants^[23]. Moreover, tissue heterogeneity and limited ancestry diversity may reduce the portability of these inferences. It should also be noted that the layer-specific findings reported above, including 30 genes, 11 splicing events, and 39 metabolites, were derived from individual SMR analyses rather than being jointly supported by multiple QTL layers. Multi-layer-supported signals were identified by the OPERA analysis, which detected 44 longevity-associated loci with joint evidence from at least two molecular QTL layers^[8]. In addition, this analytical hierarchy does not imply a sequential biological mediation pathway across molecular layers. Formal mediation analysis, longitudinal multi-omics data and experimental perturbation are required to establish sequential mediation. Nonetheless, it is important to acknowledge that this methodology cannot eliminate the influence of unknown horizontal pleiotropy or biases introduced by sample overlap^[22,24]. Despite these limitations, the approach represents one of the most robust statistical causal inference frameworks currently achievable given existing data constraints.

Mechanistic gaps will require further investigation

Through the application of SMR and OPERA analyses, genes including CASP8, PSRC1, and SORT1 were identified^[8]. Notably, CASP8, traditionally recognized for its role in apoptosis, exhibits a positive correlation with longevity, suggesting potential non-apoptotic functions, possibly related to immune homeostasis or cellular quality control^[25]. The liver-specific association of PSRC1 and SORT1 further underscores the importance of tissue-specific regulation in longevity. While these findings are significant, functional validation remains essential. As Gai et al. have noted^[8], integrative approaches based on xQTL data cannot substitute for direct experimental interventions. For instance, the identification of an apaQTL signal affecting FBXL5, absent in eQTL datasets, prompts the question of how alternative polyadenylation of this gene may influence lifespan. Addressing such mechanistic gaps will require further investigation using model organisms or cellular systems. For instance, CRISPR-Cas9 or prime editing of the 3’ UTR polyadenylation signal in hepatic organoids or hepatocyte cell lines (e.g., HepG2 or Huh7), followed by 3’ RACE, APA reporter assays, and RNA stability measurements, as well as mouse models with altered FBXL5 APA isoform usage. To resolve tissue-specific signals, single-cell RNA sequencing and spatial transcriptomics in liver, brain, blood, adipose, and thyroid tissues could identify the genes of cell types driving tissue-specific QTL effects^[26]. In addition, the sample size and ancestry limitations of the QTL datasets used here—most of which are predominantly European and based on whole blood, plasma, or bulk brain tissue—should be addressed by generating larger, cross-ancestry, and tissue-specific QTL datasets.

Enhancement of the association map linking pharmaceutical compounds and their biological targets

The mapping of associations between drugs and their molecular targets constitutes a critical component in longevity research, bridging the gap from molecular discovery to therapeutic intervention^[27]. Gai et al. employed the DrugBank database to link genes and proteins identified through SMR analyses to established or experimental pharmacological agents, thereby offering direct insights for drug repurposing efforts^[8]. For example, NFKB1, a central regulator of inflammatory processes, has been reported as a potential mediator of aspirin’s anti-inflammatory effects in atherosclerosis, although this evidence is context-dependent and requires further validation^[28]; similarly, the interaction between feline sarcoma oncogene and fostamatinib suggests potential therapeutic avenues for modulating immune-related aging pathways^[29]. The principal advantage of this supplementary approach lies in its capacity not only to catalog molecular phenotypes but also to actively associate these with existing drug resources, thereby expediting the translation from genetic findings to clinical evaluation. This is particularly pertinent for targets such as CASP8 and PSRC1, which exhibit significant associations with longevity^[8]; the presence of drugs targeting their homologous proteins or related pathways could prioritize these candidates for experimental validation or dosage optimization. Nonetheless, the current drug-target association map predominantly reflects correlative relationships and lacks critical annotations regarding tissue specificity, dosage effects, and adverse reactions^[30-32]. Moreover, most associations derive from literature reports or structural similarity assessments rather than direct in vivo intervention evidence^[33,34]. Additionally, the small-molecule drugs cataloged in DrugBank primarily target enzymes or receptors, with minimal representation of splicing events or mQTLs^[35]. Consequently, while this supplementary strategy offers valuable hypotheses, it should be considered a tool for generating assumptions rather than a definitive therapeutic framework^[36,37]. Future integration of drug-target associations with tissue-specific expression profiles, age-related expression dynamics, and multi-omics co-localization data, together with in vivo intervention evidence, holds promise for more effectively guiding precision intervention trials in longevity research.