MDD represents a substantial global health burden, yet approximately one-third of patients develop TRD following standard antidepressant therapy, facing prolonged suffering and increased healthcare costs. The current clinical diagnosis of TRD relies exclusively on retrospective assessment of treatment response and subjective symptom evaluation, lacking objective biomarkers to guide early identification and intervention. In this study, we integrated untargeted serum metabolomics with routine clinical indicators and employed a multi-step machine learning strategy to construct a dual-modal diagnostic model for TRD. Our results demonstrated that the random forest-based model combining eight core metabolic biomarkers and three clinical risk factors achieved an AUC of 0.911 (95%CI: 0.776–1.000) in the independent validation set, significantly outperforming single-modal approaches. These findings provide a novel, quantifiable tool for objective TRD identification and offer insights into the metabolic pathophysiology underlying treatment resistance.

The eight core metabolic biomarkers identified through our multi-step feature selection strategy reveal distinct biological processes implicated in TRD pathogenesis. Notably, three endogenous fatty acid amides-oleamide, linoleamide (18:2n6), and palmitamide (16:0)-were significantly downregulated in TRD patients serum. These lipid mediators function as critical signaling molecules within the endocannabinoid system (ECS), modulating neuroinflammation, synaptic plasticity, sleep-wake cycles, and stress responses (9–11). Reduced circulating levels of these amides suggest impaired ECS signaling in treatment-resistant depression, potentially reflecting chronic depletion of neuroprotective lipid mediators due to sustained stress, inflammation, or prolonged disease duration (12). Preclinical studies have demonstrated that oleamide possesses anxiolytic, sleep-promoting, and anti-inflammatory properties; its deficiency may compromise the endogenous mechanisms that normally limit neuroinflammatory responses and maintain synaptic homeostasis (13–15). Our findings extend these observations to the clinical context of TRD, suggesting that ECS hypofunction—rather than hyperactivation—characterizes treatment-resistant cases, aligning with the “clinical endocannabinoid deficiency syndrome” hypothesis proposed in chronic stress-related disorders (16).

Conversely, the elevation of 12-hydroxyeicosatetraenoic acid (12-HETE) and fibrinopeptide A in TRD patients supports the involvement of pro-inflammatory and pro-thrombotic processes in treatment resistance (17, 18). 12-HETE, a product of arachidonic acid metabolism via 12-lipoxygenase, exhibits potent pro-inflammatory properties and has been implicated in blood-brain barrier dysfunction (19, 20). The reciprocal pattern of reduced anti-inflammatory amides and elevated pro-inflammatory eicosanoids suggests a disrupted lipid mediator balance in TRD, favoring neuroinflammatory and pro-oxidative states. This imbalance characterized by deficiency of endogenous anti-inflammatory counter-regulatory mechanisms alongside activation of innate immune responses may constitute a metabolic signature of treatment refractoriness. Fibrinopeptide A, released during thrombin-mediated fibrinogen cleavage, indicates activation of the coagulation cascade, which interacts with neuroinflammatory pathways and may contribute to cerebrovascular dysfunction in refractory depression.

The reduction in beta-cryptoxanthin, a dietary carotenoid with antioxidant and anti-inflammatory properties, further supports the presence of compromised antioxidant defenses in TRD (21). This finding aligns with accumulating evidence linking oxidative stress and reduced antioxidant capacity to depression severity and treatment resistance (22, 23). The concomitant decrease in alliin, a compound associated with gut microbiota metabolism and sulfur-containing amino acid pathways, hints at potential alterations in gut-brain axis communication and microbial-host metabolic crosstalk in treatment resistance (24, 25). The glucuronide conjugate of C10H18O2, while structurally requiring further validation, may represent altered xenobiotic metabolism or phase II detoxification processes. Collectively, these metabolic alterations depict TRD as a state characterized by deficiency of neuroprotective and anti-inflammatory mediators alongside activation of pro-inflammatory and pro-oxidative pathways. This metabolic phenotype may impede therapeutic response.

The three identified clinical risk factors (medical history, HDL, FBG) are readily accessible and have important clinical and pathological significance for TRD. Low HDL is associated with chronic low-grade inflammation and impaired reverse cholesterol transport, and as a carrier for ECS lipid mediators (26), its reduction further exacerbates ECS hypofunction in TRD. Reduced FBG in the TRD group may reflect metabolic dysregulation or altered glucose homeostasis associated with chronic stress and depressive pathology, which has been linked to poor antidepressant response. Longer medical history reflects the accumulation of biological alterations such as progressive ECS dysfunction and inflammatory burden in chronic TRD cases. These factors enable initial clinical risk stratification for TRD, forming a simple and feasible preliminary screening tool for clinical practice.

It is critical to emphasize that the cross-sectional design and machine learning framework employed here identify statistical associations, not causal relationships. The metabolic alterations observed may represent consequences of treatment resistance, compensatory responses, or epiphenomena of chronic illness. Longitudinal and interventional studies are required to disentangle these possibilities and to validate whether the identified pathways are causally involved in treatment refractoriness.

The integration of metabolic biomarkers and clinical risk factors is the key to the superior diagnostic performance of the dual-modal model, as the two data types exhibit complementary advantages. Metabolomic profiling captures the intrinsic pathophysiological state of TRD patients, reflecting the downstream effects of genetic, environmental and microbial factors, while clinical risk factors provide rapid and accessible information on disease chronicity and metabolic comorbidities. SHAP analysis confirmed that Fibrinopeptide A516, 12-HETE and the three clinical factors jointly drive model predictions, with no single feature dominating classification. Methodologically, the multi-step machine learning pipeline effectively minimized overfitting, and the random forest algorithm ensured good generalization ability. This dual-modal strategy establishes a tiered diagnostic approach for TRD, and the identified ECS hypofunction also provides a promising therapeutic target—preclinical studies have shown that ECS signaling enhancement exerts antidepressant-like effects (27), suggesting potential for TRD targeted therapy.

Recent efforts in precision psychiatry have increasingly relied on multi-modal data integration to improve depression prediction beyond single-modality approaches. For instance, Zha et al. constructed a machine learning-based predictive model for depression in patients with advanced cardiovascular-kidney-metabolic syndrome by integrating sleep disorders, waist circumference, and gamma-glutamyl transferase with demographic features, achieving an AUC of 0.768 in the test set and consistent performance across three external validation cohorts (AUC 0.765–0.794) (28). While these studies demonstrate the feasibility of combining clinical and biological variables for risk stratification, they predominantly employ structured clinical data or targeted laboratory panels rather than untargeted metabolomics that captures the global metabolic network state. Our study addresses this methodological gap by integrating serum untargeted metabolomics with routine clinical indicators and employing SHAP-based interpretability to dissect the relative contributions of metabolic versus clinical drivers, thereby moving beyond “black-box” prediction toward mechanistically informed biomarker panels. This distinction is particularly relevant given the growing emphasis on metabolomics as the functional endpoint of gene-environment interactions in psychiatric disorders. Cavaleri et al. recently underscored that metabolomic biomarkers offer actionable targets for individualized treatment in precision psychiatry (29). our identified endocannabinoid deficiency, inflammatory lipid mediator imbalance, and antioxidant depletion align closely with this framework. Furthermore, the clinical utility of objective TRD identification extends beyond diagnosis to therapeutic stratification. Sun et al. reviewed the emerging evidence for transcutaneous auricular vagus nerve stimulation as a non-invasive intervention for MDD (30), whereas Wang et al. demonstrated the efficacy of school-based narrative writing interventions for adolescent anxiety and depression in a large cluster RCT (31). These advances in therapeutic modalities highlight the unmet need for biomarker-guided patient selection-an objective that our dual-modal model is positioned to address in future prospective trials.

We acknowledge that the primary validation relied on a single-center internal dataset with a modest sample size, and that multi-step feature selection in high-dimensional metabolomic data carries inherent risks of overfitting. To address this transparently, we performed nested cross-validation, which yielded a mean AUC of 0.949. This convergence between the primary and secondary validation estimates suggests that the original performance metrics were not substantially inflated by information leakage. Furthermore, bootstrap resampling confirmed high feature selection stability for the core biomarkers, supporting the reproducibility of the identified metabolic signature. LASSO-regularized regression independently retained the three clinical risk factors, confirming their discriminative value beyond univariate associations. Decision curve analysis demonstrated positive net benefit across clinically relevant threshold probabilities, providing preliminary evidence of potential clinical utility pending external validation.

This study has several limitations that need to be acknowledged. First, this study represents a single-center, proof-of-concept investigation with a modest sample size (n=93). While the nested cross-validation framework provides the most rigorous internal validation possible with available data, we explicitly acknowledge that external validation in independent multi-center cohorts is mandatory before any clinical deployment. Second, the TRD definition based on HAMD-17 criteria does not capture functional recovery and medication adherence, potentially introducing diagnostic bias. Third, the cross-sectional design precludes determining the causal relationship between metabolic alterations and TRD, and some metabolites require further structural validation. Future research should prioritize longitudinal investigations to develop TRD treatment response prediction models, mechanistic studies using cellular/animal models to verify ECS hypofunction and oleamide deficiency, and integration of metabolomic data with pharmacogenomic and microbiome profiling. Most importantly, intervention studies targeting the identified metabolic pathways are needed to develop novel mechanism-based treatment strategies for TRD.