1. Introduction

It is a little sobering to remember that Long COVID began not in a laboratory or a clinical trial, but in patient support groups and citizen-science collectives who noticed, in the spring of 2020, that their symptoms simply would not go away (Maltezou et al., 2021; Yong, 2021). What started as an anecdotal phenomenon-dismissed, for a time, by clinicians who had little precedent to draw on-has since hardened into a widely acknowledged, chronic, systemic disease, officially termed post-acute sequelae of SARS-CoV-2 infection, or PASC (Maltezou et al., 2021; Yong, 2021). The scale is difficult to overstate: more than 400 million people worldwide are now thought to be affected, with annual economic costs exceeding $1 trillion (Faghy et al., 2026). And yet, despite this enormous burden, defining and diagnosing the condition remains, even now, something of an unresolved puzzle for the medical community (Struhal & Almamoori, 2025; Captur et al., 2022; Giron et al., 2021, Patel et al., 2022,).

Part of the difficulty lies in simple terminological disagreement. The UK's National Institute for Health and Care Excellence (NICE) splits the condition into 'ongoing symptomatic COVID-19' (symptoms persisting four to twelve weeks post-infection) and 'post-COVID-19 syndrome' (symptoms beyond twelve weeks) (Struhal & Almamoori, 2025). The World Health Organization, meanwhile, defines 'post-COVID-19 condition' somewhat differently-as symptoms persisting or newly emerging three months after infection, lasting at least two months, and not otherwise explained (Castanares-Zapatero et al., 2022; Tsilingiris et al., 2023). These competing definitions are not merely bureaucratic quibbling; they reflect, we think, the genuinely fluctuating and heterogeneous nature of the illness itself, which spans more than 200 reported symptoms across nearly every physiological system-severe fatigue, cognitive impairment ("brain fog"), dyspnea, chronic headache, sleep disturbance, cardiovascular complications, and gastrointestinal disturbance, among others (Faghy et al., 2026; Gheorghita et al., 2024; Yong, 2021). While the risk of developing Long COVID rises sharply with acute illness severity-nearly tenfold in those who were critically ill-the condition is now understood to occur across the full spectrum of initial severity, including in young, previously healthy, non-hospitalized individuals, and even after asymptomatic infection (Faghy et al., 2026; Ozanic et al., 2025; Yong, 2021).

To make sense of the underlying biology, it helps-perhaps more than one might initially expect-to situate Long COVID within the broader, historically underappreciated category of post-acute infection syndromes (PAIS) (Wendt et al., 2026; Patel et al., 2023; Patrascu & Dumitru, 2025), Seco-González et al. (2024), Su et al. (2022)). Persistent illness following viral, bacterial, or parasitic infection is not a new observation; it has simply been poorly studied, owing largely to how heterogeneous these presentations tend to be (Struhal & Almamoori, 2025; Wendt et al., 2026). Q fever fatigue syndrome, post-Ebola syndrome, and the chronic fatigue states documented after the Russian influenza, SARS-CoV-1, and MERS all belong to this same family (Struhal & Almamoori, 2025; Tsilingiris et al., 2023). Following the original SARS-CoV-1 outbreak, roughly 40% of survivors reported chronic fatigue as far out as 41 months post-recovery, while up to 48% of MERS survivors experienced lasting neuropsychiatric and physical impairment at 18 months (Tsilingiris et al., 2023)-numbers that, in retrospect, probably should have prepared us better for what followed SARS-CoV-2.

This comparative lens becomes even more useful once one notices how closely Long COVID resembles myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS) and postural orthostatic tachycardia syndrome (POTS) (Faghy et al., 2026; Yong, 2021). Up to half of Long COVID patients meet formal ME/CFS diagnostic criteria, sharing hallmark features such as post-exertional malaise (PEM)-a severe symptom exacerbation following even minimal physical or cognitive exertion-alongside unrefreshing sleep, orthostatic intolerance, and brain fog (Faghy et al., 2026; Wendt et al., 2026). This overlap is, we would argue, too extensive to be coincidental; it points toward shared, pathogen-independent upstream mechanisms across post-viral states, in which the initial infection functions less as an ongoing cause and more as a trigger for sustained, multi-systemic dysregulation (Faghy et al., 2026; Groysman, 2026; Wendt et al., 2026).

Contemporary biomedical research increasingly frames Long COVID as a multi-systemic network disorder, driven by several interconnected, self-reinforcing pathophysiological domains rather than any single causal lesion (Faghy et al., 2026; Groysman, 2026). These mechanisms are not, importantly, mutually exclusive-they appear to interact within biologically enriched patient subsets, producing the diverse clinical pictures clinicians actually encounter (Groysman, 2026).

One prominent hypothesis concerns occult viral persistence-the formation of anatomical reservoirs of SARS-CoV-2 that evade full clearance (Faghy et al., 2026; Wang et al., 2023; Woodruff et al., 2023). Viral RNA and structural proteins, notably the spike (S1) and nucleocapsid (N) proteins, have been detected in deep tissues-gut enterocytes, hepatic tissue, lung parenchyma, and along the skull-meninges-brain axis-long after the virus has cleared the upper respiratory tract (Faghy et al., 2026; Gupta et al., 2025). Circulating S1 spike protein has even been measured in patient plasma up to 14 months post-infection, hinting at ongoing antigen leakage (Faghy et al., 2026), and dose-response modeling shows that higher circulating spike concentrations correlate directly with symptom burden and stimulate proinflammatory cytokine release-CXCL8, IL-6, IL-1β, and TNF-α-from human lung macrophages, sustaining a state of chronic, low-grade inflammation (Yang et al., 2025).

This same chronic antigen exposure appears to disrupt immune regulation more broadly, driving persistent cytotoxic activation, T-cell exhaustion, and reactivation of latent herpesviruses such as Epstein-Barr virus (EBV) and human herpesvirus 6 (HHV-6)-both strongly correlated with fatigue and cognitive deficits in Long COVID cohorts (Gupta et al., 2025; Wendt et al., 2026; Tsilingiris et al., 2023). The same dysregulated environment also seems to promote autoimmunity via molecular mimicry, with functional autoantibodies against G-protein coupled receptors (GPCRs, which govern vascular tone and heart rate) and ACE2 (which impairs the renin-angiotensin system) now reported with some consistency across cohorts (Castanares-Zapatero et al., 2022; Faghy et al., 2026; Che Ramli et al., 2025).

Vascular pathology compounds all of this. Direct endothelial injury and sustained thromboinflammation produce measurable endotheliopathy-capillary rarefaction, impaired microvascular reactivity, hyperactivated platelets (Faghy et al., 2026; Castanares-Zapatero et al., 2022)-and, perhaps most strikingly, anomalous amyloid-containing fibrin microclots that resist normal fibrinolysis and physically obstruct the microcirculation (Faghy et al., 2026; Pretorius et al., 2021). These microclots entrap numerous pro-inflammatory molecules, impair systemic oxygen extraction, and drive tissue hypoxia-a mechanism that may well explain the severe post-exertional malaise so many patients describe (Faghy et al., 2026; Pretorius et al., 2021; Kruger et al., 2022). At the cellular level, this hypoxia combines with direct mitochondrial injury to produce measurable bioenergetic depletion: reduced oxidative phosphorylation capacity, impaired fatty acid oxidation, and persistent oxidative stress (Faghy et al., 2026; Al-Hakeim et al., 2022), with lower acute-phase oxygen saturation and higher peak body temperature both predicting the severity of chronic fatigue and neuropsychiatric symptoms months later (Al-Hakeim et al., 2022).

Given this biological heterogeneity, the case for standardized, objective biomarkers seems, to us, fairly compelling-not as an academic nicety, but as a precondition for accurate diagnosis, prognosis, and meaningful clinical trial design (Faghy et al., 2026). Multi-omics approaches spanning genomics, epigenomics, transcriptomics, proteomics, and metabolomics have begun mapping molecular signatures onto distinct patient subgroups (Pinero et al., 2025). Genome-wide association studies have implicated susceptibility loci such as FOXP4, epigenome-wide studies show persistent DNA methylation changes in immune genes like IFI44L, and proteomic and metabolomic profiling reveal complement and coagulation dysregulation alongside altered amino acid and lipid pathways, including kynurenine pathway activation linked to cognitive symptoms (Pinero et al., 2025; Wendt et al., 2026; Tsilingiris et al., 2023). Circulating microRNAs-miR-200c-3p and miR-142-3p among them-appear to regulate inflammatory and immune pathways and show promise as diagnostic markers (Paval et al., 2025), while neuro-injury markers such as neurofilament light chain (NfL) and glial fibrillary acidic protein (GFAP) offer an objective window onto central nervous system damage in patients with prominent neurological symptoms (Che Ramli et al., 2025).

Translating any of this into clinical practice, however, remains hampered by real and persistent uncertainty (Faghy et al., 2026). No curative pharmacological treatment currently exists, forcing clinicians toward off-label symptom management (Faghy et al., 2026; Ozanic et al., 2025). Worse, the absence of standardized diagnostic biomarkers has historically driven a kind of 'psychologization' of Long COVID, in which patients are misdiagnosed with somatic symptom disorder or anxiety-a pattern that Spanoghe et al. (2026) describe, not unreasonably, as capable of causing serious harm. Graded exercise therapy (GET), once recommended on the assumption that post-viral fatigue reflected simple deconditioning, has been shown instead to trigger physiological relapse and worsen underlying vascular and mitochondrial pathology in patients with PEM-prompting international guidelines to recommend pacing and energy conservation instead (Faghy et al., 2026).

Given all of this, we think the field urgently needs to move from broad, symptom-defined enrollment toward biologically stratified, mechanism-anchored trial design (Groysman, 2026), identifying treatable biological traits within more homogeneous patient subgroups (Faghy et al., 2026; Groysman, 2026). This manuscript accordingly pursues four guiding research questions: first, to what extent circulating SARS-CoV-2 spike protein concentration correlates with distinct clinical endotypes and predicts differential treatment response; second, how circulating microRNA and DNA methylation profiles regulate chronic immune-inflammatory cascades in Long COVID relative to ME/CFS; third, what diagnostic and prognostic value combining microvascular biomarkers (amyloid microclots, endothelin-1) with neurological injury markers (NfL, GFAP) might offer for predicting cognitive decline and exertional intolerance; and fourth, whether targeted pharmacological interventions-low-dose naltrexone, rovunaptabin, therapeutic apheresis-outperform non-pharmacological pacing strategies in reversing mitochondrial and bioenergetic impairment. The corresponding research objectives aim to establish a standardized multi-systemic biomarker panel, quantitatively characterize spike-protein dose-response relationships over time, compare Long COVID with ME/CFS and POTS to isolate disease-specific versus pathogen-independent signatures, and evaluate candidate therapeutics in biologically stratified cohorts.

2. Mechanisms, Biomarkers, and Management Uncertainties in Long COVID and Post-Viral Syndromes

Beneath the enormous diversity of Long COVID symptoms lies, or so the evidence increasingly suggests, a smaller set of interconnected biological processes-fewer, certainly, than the 200-plus symptoms would imply. This section works through those processes stage by stage, from the molecular hub that seems to unify them, down through metabolic, vascular, epigenetic, and sex-specific layers, before turning to the therapeutic and management controversies that follow directly from this biology.

2.1 The Unified Pathophysiological Hub: NF-κB and Bidirectional Cellular Crosstalk

At the center of the mechanistic landscape shared by post-acute infection syndromes-Long COVID and ME/CFS alike-sits, somewhat surprisingly, not an active ongoing infection but a complex, interconnected web of host response systems (Wendt et al., 2026). Extensive multi-omics and systematic review evidence points to the nuclear factor kappa B (NF-κB) transcription factor pathway as the primary unifying molecular hub bridging these clinical phenotypes (Wendt et al., 2026; Figure 1). Rather than acting in isolation, NF-κB appears to be continuously fed by bidirectional feedback loops spanning multiple physiological domains, sustaining a state of chronic, low-grade systemic inflammation (Groysman, 2026; Wendt et al., 2026).

One of the principal upstream triggers of this chronic activation is gut dysbiosis and mucosal barrier disruption (Gheorghita et al., 2024; Groysman, 2026). Under healthy conditions, a balanced microbiome produces short-chain fatty acids such as butyrate, which maintain epithelial barrier integrity and suppress inflammation by downregulating the IκB kinase (IKK) complex (Gheorghita et al., 2024; Wendt et al., 2026). In PASC and ME/CFS, however, a marked depletion of butyrate-producing Firmicutes-Faecalibacterium and Ruminococcus in particular-appears to compromise this protective mechanism (Gheorghita et al., 2024; Wendt et al., 2026). The resulting "leaky gut" allows bacterial lipopolysaccharide (LPS) to translocate into systemic circulation, where it engages Toll-like receptors 4 and 2 (TLR4/TLR2), triggering IKK-mediated degradation of IκBα, release of the p65/p50 heterodimer, and transcription of IL-6, TNF-α, and IL-1β (Wendt et al., 2026).

This inflammatory cascade, in turn, engages in immediate, self-reinforcing crosstalk with mitochondrial stress pathways (Gheorghita et al., 2024; Groysman, 2026). Chronic inflammation generates reactive oxygen species that damage mitochondrial structures, prompting the leakage of mitochondrial DNA (mtDNA) fragments into the cytosol, where the cGAS-STING pathway detects them and recruits TBK1 and IKKε to further amplify NF-κB translocation via IRF3 phosphorylation (Wendt et al., 2026). The result is a genuinely vicious cycle: gut barrier compromise and mtDNA leakage drive NF-κB transcription, which fuels the very inflammatory and oxidative environment that continues to degrade both mucosal and mitochondrial integrity (Gheorghita et al., 2024; Groysman, 2026; Wendt et al., 2026).

2.2 The Tryptophan-Kynurenine Pathway and the Bioenergetic Bottleneck

Downstream of this chronic inflammatory milieu lies a substantial reprogramming of host metabolism, expressed most visibly through the tryptophan-kynurenine pathway (TKP) and skeletal muscle bioenergetics (Gupta et al., 2025; Wendt et al., 2026). Under normal conditions, tryptophan feeds serotonin synthesis (mood, gut motility, vascular tone) and melatonin production (sleep-wake regulation) (Wendt et al., 2026). Pro-inflammatory cytokines-interferon-gamma and TNF-α especially-strongly upregulate indoleamine 2,3-dioxygenase 1 (IDO1) and tryptophan 2,3-dioxygenase (TDO2), diverting tryptophan away from serotonin and toward kynurenine metabolism, depleting serotonin and skewing the tryptophan-to-kynurenine ratio (Wendt et al., 2026). Kynurenine is then metabolized into quinolinic acid (QUIN), a neurotoxic NMDA receptor agonist, while the neuroprotective kynurenic acid (KYNA) branch is concurrently suppressed-leaving the brain exposed to NMDA-mediated excitotoxicity and diffuse neuroinflammation (Tsilingiris et al., 2023; Wendt et al., 2026). This "tryptophan drain" links fairly directly to the brain fog, sleep disturbance, and neuropsychiatric symptoms reported across Long COVID and ME/CFS cohorts (Tsilingiris et al., 2023; Wendt et al., 2026).

Systemic bioenergetics are, simultaneously, throttled by an imbalance between oxidative toxicity (OSTOX) and antioxidant defenses (ANTIOX) (Al-Hakeim et al., 2022). Rigorous cohort analysis shows that roughly 31.7% of Long COVID patients present a distinct, severe endophenotype marked by elevated malondialdehyde (MDA) and protein carbonyls alongside dramatic depletion of glutathione peroxidase (Gpx) and zinc (Al-Hakeim et al., 2022; Table 1). This OSTOX/ANTIOX imbalance correlates directly with fatigue and fibromyalgia-like symptom severity (Al-Hakeim et al., 2022). At the cellular level, chronic oxidative stress impairs the electron transport chain, reducing oxidative phosphorylation and ATP output; deprived of aerobic capacity, skeletal muscle shifts prematurely to anaerobic glycolysis even under minimal exertion, accumulating lactic acid and amyloid deposits (Al-Hakeim et al., 2022; Gupta et al., 2025; Faghy et al., 2026). This mitochondrial bottleneck offers, we think, a fairly clear and objective biological account of post-exertional malaise-the hallmark feature in which minor activity triggers a prolonged, debilitating relapse (Gupta et al., 2025; Faghy et al., 2026).

2.3 Endothelial Injury, Microvascular Rarefaction, and Immunothrombosis

Even where cellular energy production is intact, oxygen and nutrient delivery can still be compromised by systemic endothelial injury, microvascular rarefaction, and immunothrombotic pathology (Gheorghita et al., 2024; Groysman, 2026). Direct endothelial viral invasion via ACE2 receptors, combined with persistent exposure to circulating spike protein and inflammatory cytokines, produces widespread vascular endotheliitis and capillary rarefaction (Ozanic et al., 2025; Faghy et al., 2026), reflected in elevated Angiopoietin-1 (ANG-1) and soluble vascular cell adhesion molecule-1 (sVCAM-1)-the former showing notably high sensitivity and specificity for distinguishing Long COVID patients from controls (Tsilingiris et al., 2023; Wendt et al., 2026).

This endotheliitis travels alongside persistent platelet and leukocyte hyperactivation (Faghy et al., 2026), and under the influence of chronic inflammation and elevated alpha-2 antiplasmin and CRP, the coagulation cascade is diverted toward anomalous, amyloid-containing fibrin microclots (Tsilingiris et al., 2023; Faghy et al., 2026). These microclots are structurally distinct from ordinary clots-dense, packed with entrapped pro-inflammatory molecules, and highly resistant to normal fibrinolysis (Wendt et al., 2026; Faghy et al., 2026)-and as they circulate they physically lodge within the microvasculature, producing focal capillary blockages, chronic hypoperfusion, and widespread hypoxia (Wendt et al., 2026; Faghy et al., 2026). During physical exertion, the combination of impaired oxygen delivery (microvascular blockage) and impaired oxygen utilization (mitochondrial dysfunction) forces skeletal muscle into a genuine bioenergetic crisis, further driving the clinical picture of PEM (Gupta et al., 2025; Faghy et al., 2026).

2.4 The Epigenetic and Regulatory Landscape: The miRNA Target Network

This vascular, metabolic, and immune dysregulation is maintained at the transcriptional level by persistent epigenetic modification and a substantially dysregulated microRNA (miRNA) network (Wendt et al., 2026; Paval et al., 2025; Figure 1). High-throughput small RNA sequencing shows that miR-155, miR-146a, and miR-21 are chronically upregulated across both PASC and ME/CFS cohorts, functioning as master regulators of innate and adaptive immunity (Wendt et al., 2026; Paval et al., 2025). miR-155 promotes pro-inflammatory cytokine transcription and elevates Serum Amyloid A (SAA), an acute-phase protein that further drives endotheliitis and clot pathology, while miR-146a and miR-21 modulate systemic and cardiac remodeling but, in their chronically dysregulated state, contribute to persistent T-cell exhaustion and upregulation of Galectin-9 (GAL-9), a marker linked to immune dysfunction and viral persistence (Wendt et al., 2026; Paval et al., 2025).

In contrast, miR-124-a crucial neuroprotective regulator-is frequently depleted in post-viral neurological sequelae (Paval et al., 2025). Under normal conditions, miR-124 promotes neuronal differentiation and keeps microglia quiescent by suppressing the REST/SCP1 pathway and downregulating STAT3 and CEBP-α; when miR-124 expression falls, this inhibitory brake is lost, and unchecked STAT3/CEBP-α signaling promotes astrocyte

Figure 1. This figure depicts a self-reinforcing pathological loop linking gut barrier breakdown, systemic inflammation, and mitochondrial dysfunction in PASC and ME/CFS, with NF-κB signaling acting as the central unifying hub. A leaky gut barrier permits bacterial LPS translocation, activating TLR4/TLR2 receptors and the IKK complex, which drives IκBα degradation and release of the p65/p50 NF-κB subunits into the nucleus to switch on inflammatory gene expression. This NF-κB activation fuels a systemic surge in pro-inflammatory cytokines (IL-6, TNF-α, IL-1β) and reactive oxygen species, producing chronic low-grade inflammation that in turn damages mitochondria, triggering further ROS generation and mtDNA leakage. Leaked mtDNA is sensed by the cGAS-STING/TBN1-IKKε pathway, which further activates NF-κB and IRF3 phosphorylation, closing the vicious cycle and perpetuating gut, immune, and mitochondrial dysfunction in a mutually reinforcing manner.

Figure 2. Diagnostic Performance of Machine Learning Classifiers Across Omics Layers in Long COVID Biomarker Studies. This bar chart compares the diagnostic or classification area under the curve (AUC) achieved by machine learning models built on transcriptomic, proteomic, metabolomic, and integrated multi-omics data, as summarized in Table 4. Transcriptomic and metabolomic classifiers generally achieved the highest accuracy (AUC 0.86-0.95), while proteomic complement-based models showed comparatively lower but still clinically meaningful performance (AUC 0.785-0.857), illustrating the current state of biomarker-based patient stratification (Hou et al., 2025; Lin et al., 2024; Ai et al., 2024).

reactivity and diffuse neuroinflammation, contributing to cognitive dysfunction, dysautonomia, and persistent anosmia (Che Ramli et al., 2025; Paval et al., 2025).

2.5 Sex-Specific Phenotypes and Stratification Dilemmas

One of the more striking epidemiological features of both Long COVID and ME/CFS is a pronounced sex disparity, with females at significantly higher risk (Struhal & Almamoori, 2025; Paval et al., 2025). Systematic meta-analytic estimates put the relative risk at 1.40 for persistent post-viral fatigue and 1.38 for chronic cognitive disturbance in women relative to men (Struhal & Almamoori, 2025), a disparity rooted in distinct sex-specific immunological and transcriptional signatures (Wendt et al., 2026; Paval et al., 2025). During recovery, women show increased frequencies of exhausted CD4+/CD8+ T cells, heightened antibody responses to reactivated EBV and HHV-6, and elevated expression of the long non-coding RNA XIST, implicated in female-biased autoimmune susceptibility (Paval et al., 2025). Men, by contrast, tend to show elevated NK cells, pro-inflammatory IL-8, and enhanced TGF-β fibrotic signaling, with male phenotypes more often characterized clinically by neuro-cardiovascular complications and innate immune-mediated mitochondrial stress, while female phenotypes trend toward more diverse, multi-system immune-inflammatory and respiratory presentations (Wendt et al., 2026; Paval et al., 2025).

These biological differences expose, fairly starkly, the limits of symptom-based diagnostic frameworks (Groysman, 2026). Grouping biologically distinct patients under broad labels such as "fatigue" or "brain fog" inevitably combines mechanistically different conditions into a single cohort, diluting therapeutic signal and likely contributing to repeated clinical trial failures (Groysman, 2026). Groysman (2026) proposes instead a mechanism-anchored stratification framework, sorting patients into biologically enriched subsets defined by objective markers-kynurenine pathway metabolites, amyloid microclot burden, autonomic testing, specific miRNA profiles-arguing that only mechanism-defined cohorts will allow future randomized trials the statistical precision needed to evaluate genuinely targeted, disease-modifying therapies.

2.6 Therapeutic Trajectories and Management Controversies

The clinical management history of post-viral syndromes has been marked by a fairly consequential paradigm shift, particularly regarding the harm caused by functional or psychosomatic misdiagnosis (Spanoghe et al., 2026). In the absence of validated routine laboratory tests, clinicians have often defaulted to the construct of Functional Somatic Disorder (FSD)-previously termed somatoform disorder-to classify Long COVID and ME/CFS (Spanoghe et al., 2026). As Spanoghe et al. (2026) document, this "psychologization" of what is, on the growing evidence, a primarily biological disease ignores substantial objective pathophysiology-persistent tissue viral reservoirs, endotheliitis, mitochondrial respiratory defects-and causes real, documented harm: medical gaslighting, diagnostic delay, psychiatric stigma, and erosion of the therapeutic alliance.

This conceptual error translated directly into clinical guidelines endorsing Graded Exercise Therapy (GET), a regimen of incrementally increasing activity intended to reverse presumed deconditioning (Spanoghe et al., 2026; Faghy et al., 2026). GET is now widely recognized as contraindicated and potentially dangerous for patients experiencing PEM (Faghy et al., 2026): because these patients suffer genuine mitochondrial bioenergetic failure and microvascular hypoperfusion, forcing activity past the anaerobic threshold causes measurable muscle fiber damage, worsens ischemia-reperfusion injury, and triggers long-lasting systemic relapse (Gupta et al., 2025; Faghy et al., 2026). International consensus has shifted instead toward Pacing-a structured energy-management strategy in which patients monitor and restrict physical, cognitive, and emotional activity to remain within an individual "energy envelope," often using wearables to track heart rate variability and avoid triggering a crash (Gheorghita et al., 2024; Faghy et al., 2026).

Meanwhile, the pharmacological pipeline is slowly-if unevenly-pivoting toward mechanism-based intervention (Groysman, 2026; Faghy et al., 2026). Extended-course antiviral regimens (such as nirmatrelvir/ritonavir) are being trialed to clear persistent tissue viral reservoirs (Gupta et al., 2025; Groysman, 2026); low-dose naltrexone (LDN) is used to suppress microglial activation and neuroinflammation, while intravenous immunoglobulin (IVIg) neutralizes GPCR and ACE2 autoantibodies (Groysman, 2026; Faghy et al., 2026); and advanced apheresis techniques-double filtration plasmapheresis, heparin-mediated extracorporeal LDL precipitation-are under active investigation to physically filter circulating autoantibodies, cytokines, and amyloid microclots, aiming to restore microvascular flow to hypoperfused tissue (Gupta et al., 2025; Faghy et al., 2026). As these trials progress, pairing objective biomarker-driven stratification with rigorous pacing appears, at least for now, to remain the safest and most defensible cornerstone of care (Groysman, 2026; Faghy et al., 2026).

3. Methods

This manuscript synthesizes mechanistic, biomarker, and machine-learning literature on Long COVID and related post-acute infection syndromes into a single narrative framework. Because our purpose was integrative rather than confirmatory-mapping a genuinely fragmented evidence base onto a coherent set of pathophysiological domains-we adopted a structured scoping-review methodology, described below in enough detail, we hope, that another group working from the same source material could reproduce both the synthesis and the comparative figures.

3.1 Study Design

We used a narrative, scoping-review design consistent with PRISMA-ScR reporting conventions, an approach well suited to heterogeneous mechanistic and biomarker literature where pooled effect-size meta-analysis is not really appropriate (Pinero et al., 2025). This choice reflects the nature of the underlying evidence: mechanistic pathway studies, cross-sectional biomarker cohorts, and machine-learning classification papers report fundamentally different outcome types (correlation coefficients, regression betas, AUCs) that cannot be meaningfully pooled into a single summary statistic.

3.2 Information Sources and Search Strategy

Evidence was drawn from peer-reviewed journal articles, systematic reviews, and preprint sources addressing three pillars: (a) the molecular and cellular pathophysiology of Long COVID and post-acute infection syndromes more broadly (Wendt et al., 2026; Groysman, 2026; Gupta et al., 2025); (b) cross-sectional and longitudinal biomarker studies quantifying nitro-oxidative, inflammatory, and vascular markers (Al-Hakeim et al., 2022; Tsilingiris et al., 2023); and (c) machine-learning and multi-omics classification studies reporting diagnostic or prognostic performance metrics (Ai et al., 2024; Lin et al., 2024; Baillie et al., 2024). Sources were organized around five interconnected pathophysiological domains-immune/inflammatory signaling, metabolic/bioenergetic reprogramming, vascular/thromboinflammatory injury, epigenetic/miRNA regulation, and sex-specific phenotyping-cross-checked against the network-disorder framework proposed by Groysman (2026) to ensure balanced coverage rather than over-representation of any single mechanism.

3.3 Eligibility Criteria

Sources were retained if they (a) addressed Long COVID, PASC, ME/CFS, POTS, or a comparable post-acute infection syndrome; (b) reported empirical biomarker data, mechanistic pathway evidence, or machine-learning classification performance; and (c) appeared in indexed journals or recognized preprint servers. Case reports without quantitative biomarker or mechanistic data were excluded, as were sources addressing only acute-phase (rather than post-acute or chronic) COVID-19 pathology, unless they offered directly transferable prognostic value for chronic sequelae (e.g., Al-Hakeim et al., 2022, linking acute-phase SpO2 and temperature to chronic outcomes).

3.4 Data Extraction

For mechanistic sources, we extracted the implicated pathway, its principal molecular effectors, and its point of intersection with other domains (for example, the cGAS-STING-to-NF-κB link connecting mitochondrial stress to inflammatory signaling). For biomarker cohort studies, we extracted sample size, cluster or subgroup definitions, effect sizes (F-statistics, standardized betas, z-scores), and reported significance levels, forming the basis of Tables 1 through 3. For machine-learning studies, we extracted the omics data layer, algorithm family, target prediction task, key predictive features, and the primary classification accuracy metric (typically AUC), forming the basis of Table 4 and the comparative visualization in Figure 2.

3.5 Quality Considerations

Given the heterogeneity of study designs represented here-mechanistic reviews, cross-sectional biomarker cohorts, and computational classification studies-we did not apply a single formal risk-of-bias instrument, which would not meaningfully discriminate across such different evidence types. Instead, we prioritized three recurring quality markers consistent with PubMed-indexed methodological reviews in this field: (a) whether biomarker associations were reported with explicit effect sizes and significance values rather than qualitative description alone; (b) whether machine-learning models reported validation strategy (internal cross-validation versus independent cohort testing); and (c) whether mechanistic claims were supported by more than one independent source, which we required before treating a pathway link (e.g., gut dysbiosis to NF-κB activation) as sufficiently well-established to include in Figure 1.

3.6 Data Synthesis and Figure Construction

Mechanistic findings were synthesized thematically into the five-domain framework described above, with cross-domain feedback loops (for example, the mitochondrial-inflammatory cycle described in Section 2.1) mapped explicitly to support the network diagram presented in Figure 1. Machine-learning performance metrics extracted from Table 4 were synthesized into the comparative bar chart in Figure 2, grouping studies by omics layer (transcriptomic, proteomic, metabolomic, multi-omics) to allow visual comparison of diagnostic accuracy across modalities. Where a single study reported multiple models (e.g., Al-Hakeim et al., 2022, reporting eight regression models), all reported outcomes were retained and are presented in full in Table 3 rather than being condensed to a single summary value, in the interest of reproducibility.

3.7 Reporting

This synthesis is reported in a manner intended to allow independent verification against the primary sources cited throughout: all extracted effect sizes, AUC values, and sample sizes in Tables 1 through 4 are traceable to a single named primary study, and the two figures are described in their captions with sufficient specificity (data source, modeling approach, and metric) that another reviewer could reconstruct them from the same underlying literature.

4. Characterizing the Clinical-Biological Interface of Long COVID

Bringing together the clustering, regression, and machine-learning evidence reviewed here, a fairly consistent picture emerges: what once looked like a chaotic, symptom-driven spectrum resolves, once objective biomarkers are applied, into distinct and reproducible biological endophenotypes (Table 1; Table 2; Table 3; Table 4).

4.1 Nitro-Oxidative Endophenotyping: From Symptom Checklist to Biological Cluster

To untangle the dense web of symptoms characterizing PASC, researchers have increasingly moved away from broad, subjective symptom checklists and toward objective biochemical profiling. Unsupervised two-step cluster analysis has successfully resolved what was once considered a chaotic clinical spectrum into distinct, biologically defined endophenotypes (Al-Hakeim et al., 2022; Table 1). Pairing clinical measurements with a panel of inflammatory and nitro-oxidative biomarkers separates patients into three clearly divergent groups: healthy controls, a mild-to-moderate cohort (Cluster 1), and a severe cohort with pronounced nitro-oxidative stress, termed LC+O&NS (Cluster 2) (Al-Hakeim et al., 2022; Table 1).

Patients in Cluster 2 are distinguished, first, by an acute-phase history of substantial respiratory compromise: their lowest recorded peripheral oxygen saturation averaged 87.5% ± 0.5%, compared with 92.3% ± 0.4% in Cluster 1 and 96.6% ± 0.6% in controls (Al-Hakeim et al., 2022; Table 1), alongside a markedly elevated peak body temperature (39.3 ± 0.1°C) that together produced a sharply escalated composite zTO2 index (0.862 ± 0.079) (Al-Hakeim et al., 2022). By the chronic phase-roughly 13 to 16 weeks post-infection-this acute insult had translated into a system-wide collapse in antioxidant capacity: glutathione peroxidase fell to 17.97 ± 0.84 ng/ml (versus 23.46 ± 0.97 ng/ml in controls), and zinc collapsed to 0.657 ± 0.030 ppm, driving a strongly negative composite ANTIOX z-score of -0.580 ± 0.130 (Al-Hakeim et al., 2022; Table 1). Deprived of these defenses, Cluster 2 patients showed correspondingly elevated malondialdehyde, protein carbonyls, and myeloperoxidase-markers of lipid peroxidation, protein oxidation, and neutrophil activation, respectively-culminating in a skewed OSTOX/ANTIOX ratio of 0.957 ± 0.113 (Al-Hakeim et al., 2022; Table 1).

This biological signature correlates, unsurprisingly but still strikingly, with substantial clinical and neuropsychiatric burden (Table 2). Hamilton Depression Rating Scale (HAMD) scores rose to 20.17 ± 0.70 in Cluster 2, versus 15.24 ± 0.61 in Cluster 1 and 6.01 ± 0.82 in controls, driven largely by somatic presentations of depression, fatigue, and pain (Al-Hakeim et al., 2022; Table 2). Hamilton Anxiety Scale and Fibro-Fatigue scores followed a matching upward trajectory, and composite measures of autonomic dysfunction (0.623 ± 0.104), sleep disturbance (0.591 ± 0.121), and fatigue (0.519 ± 0.117) together constitute what looks very much like the core

Table 1: Biomarker and Demographic Profile of Healthy Controls versus Long COVID Endophenotypes. This table compares socio-demographic characteristics and nitro-oxidative stress biomarkers across healthy controls and two clustered Long COVID subgroups identified by Al-Hakeim et al. (2022): Cluster 1 (mild-to-moderate Long COVID) and Cluster 2 (severe Long COVID with nitro-oxidative stress, LC+O&NS). Cluster 2 is distinguished by markedly elevated peak body temperature, reduced oxygen saturation during acute illness, and pronounced depletion of antioxidant defenses (glutathione peroxidase, zinc). Statistical significance (F/χ², df, p) is reported for each variable to indicate the strength of cluster separation.

Table 2: Clinical and Neuropsychiatric Rating Scale Outcomes across Long COVID Endophenotypes. This table reports symptom severity scores from validated psychiatric and neuro-somatic instruments-the Hamilton Depression Rating Scale (HAMD), Hamilton Anxiety Rating Scale (HAMA), and Fibro-Fatigue (FF) scale-across the same three clustered cohorts described in Table 1 (Al-Hakeim et al., 2022). Cluster 2 (LC+O&NS) patients show markedly elevated scores across all domains, with the largest effect sizes observed for somatic depression, fatigue, and autonomic dysfunction. These findings link the biochemical endophenotype in Table 1 directly to observable clinical symptom burden.

neuropsychiatric phenome of severe PASC (Al-Hakeim et al., 2022; Table 2).

4.2 Upstream Clinical Triggers and the Adverse Outcome Pathway Model

To trace how acute-phase respiratory and immunologic insult initiates this chronic, self-sustaining pathology, multivariate regression models have mapped the pathway from initial infection to long-term sequelae with some precision (Al-Hakeim et al., 2022; Table 3). These models confirm that acute physical parameters are potent predictors of somatic and affective severity three to four months later. Depressive symptom severity, for instance, is directly predicted by peak acute-phase body temperature (β = 0.390, t = 5.12, p < 0.001), chronic oxidative toxicity (β = 2.37, t = 3.70, p < 0.001), and acute-phase oxygen desaturation (β = -0.195, t = -2.49, p = 0.014) (Al-Hakeim et al., 2022; Table 3). The predictive structure becomes, if anything, more robust for somatic outcomes: roughly 58.7% of the variance in somatic Fibro-Fatigue scores is explained by a model combining acute-phase oxygen desaturation (β = -0.457, p < 0.001), peak body temperature (β = 0.322, p < 0.001), nitric oxide (β = 0.151), and lipid peroxidation (β = 0.132) (Al-Hakeim et al., 2022; Table 3).

Crucially, this mathematical mapping suggests a clear causal ordering: lowered SpO2 during acute infection appears to act as the primary driver of chronic systemic redox imbalance, directly predicting the chronic OSTOX/ANTIOX ratio (β = -0.405, p < 0.001) via glutathione peroxidase depletion and nitric oxide overproduction (Al-Hakeim et al., 2022; Table 3). Under this adverse outcome pathway framework, overall neuropsychiatric symptom burden is substantially explained (R² = 0.613) by the combined, direct effects of acute respiratory hypoxia, acute systemic inflammation, and the resulting chronic redox imbalance (Al-Hakeim et al., 2022; Table 3). Taken together, these regression models argue fairly persuasively that chronic neuropsychiatric and fibromyalgia-like fatigue in PASC are not isolated psychosomatic entities, but rather direct downstream consequences of acute hypoxia-driven redox collapse and chronic neuro-oxidative injury (Al-Hakeim et al., 2022).

4.3 Resolving Heterogeneity: Computational Stratification and Multi-Omics Machine Learning

While the clinical regression models above establish clear pathways of disease progression, advanced computational approaches applying machine learning to high-dimensional multi-omics data have proven instrumental in segmenting patient cohorts, selecting key molecular features, and uncovering diagnostic panels with genuinely impressive accuracy (Pinero et al., 2025; Table 4; Figure 2). In the transcriptomic domain, supervised feature selection-SVM-RFE, Random Forest, and XGBoost-isolated a four-gene panel (CEP55, CDCA2, MELK, DEPDC1B) predicting overall Long COVID risk with an AUC of 0.876, implicating dysregulated cell-cycle and cell-division pathways in persistent immune dysfunction (Hou et al., 2025; Table 4; Figure 2). Unsupervised K-medoids clustering, separately, identified three transcriptomic endotypes-resolved, suppressive, and unresolved-each with distinct inflammatory trajectories (An et al., 2023), while Lin et al. (2024) used unsupervised clustering and Random Forest modeling to define neutrophil-function upregulated (NU-LC) and downregulated (ND-LC) subtypes, with a five-gene classifier (ABCA13, CEACAM6, CRISP3, CTSG, BPI) reaching a diagnostic AUC of 0.95 for the severe NU-LC endotype (Lin et al., 2024; Table 4; Figure 2)-among the strongest classification results reported in this entire literature.

Proteomic analyses, meanwhile, have focused heavily on thrombo-inflammatory and complement pathology. Baillie et al. (2024) developed a generalized linear model incorporating complement activation proteins (iC3b, terminal complement complex, Ba, C5a) that predicted Long COVID severity with an AUC of 0.785, offering fairly direct biological evidence of active complement dysregulation (Baillie et al., 2024; Table 4; Figure 2). Cervia-Hasler et al. (2024) extended this longitudinally, training a Random Forest classifier with SHAP interpretability that identified complement and thrombo-inflammatory markers as the most potent predictors of active PASC at both 6 and 12 months (AUC = 0.81 ± 0.08) (Cervia-Hasler et al., 2024; Table 4; Figure 2), while Wolday et al. (2024) used PCA and Cox proportional hazards modeling to identify inflammatory risk markers (SLAMF1, TNF, IL15RA, IL18, CXCL10) forecasting chronic multi-organ symptoms with an AUC of 0.857 (Wolday et al., 2024; Table 4).

Metabolomic modeling has, for its part, concentrated on capturing systemic bioenergetic collapse over time. Holmes et al. (2021) applied PCA and O-PLS-DA to NMR-based metabolomic and lipoprotein profiles, achieving an

Table 3: Multiple Regression Models Predicting Clinical and Biochemical Severity in Long COVID. This table summarizes multiple linear regression models built by Al-Hakeim et al. (2022) evaluating how acute-phase physiological parameters (SpO2, body temperature) and chronic biomarker indices (OSTOX/ANTIOX ratio) predict downstream somatic and neuropsychiatric severity. Standardized beta coefficients, t-values, and model R² are reported for each dependent variable, illustrating a coherent adverse outcome pathway from acute hypoxia to chronic symptom burden. Higher R² values indicate stronger explanatory power of the acute-phase predictors.

Table 4: Machine Learning and Multi-Omics Classification Models in Long COVID Subtyping and Diagnosis. This table catalogs computational modeling approaches applied across transcriptomic, proteomic, metabolomic, and integrated multi-omics datasets to stratify patients, select predictive features, and discover diagnostic biomarker panels. For each study, the omics layer, algorithm, key predictive biomarkers, and classification accuracy (AUC) are reported. These results, visualized comparatively in Figure 2, demonstrate that objective molecular classification of Long COVID subtypes is achievable with diagnostic accuracy comparable to established clinical tests.

AUC of 0.86 for metabolic data and 0.81 for lipoprotein profiles in distinguishing PASC from healthy controls-characterizing a state the authors term "metabolic phenoreversion" (Holmes et al., 2021; Table 4; Figure 2). López-Hernández et al. (2023) used multivariate logistic regression to differentiate high-symptom-burden patients (more than five persistent symptoms) with an AUC of 0.94, selecting altered plasma metabolites and amino acid ratios as key predictive features (López-Hernández et al., 2023; Table 4; Figure 2).

Finally, multi-omics integration frameworks have begun to overcome the limits of single-layer analysis. Ai et al. (2024) applied Similarity Network Fusion to combine proteomic, transcriptomic, and metabolomic data, then used Random Forest and convolutional neural network classifiers to identify nine proteogenomic signatures segmenting patients into five molecular endotypes, with subtype-classification AUCs of 0.92-0.93 and an overall diagnostic AUC of 0.90 (Ai et al., 2024; Table 4; Figure 2). Chen et al. (2024) integrated cerebrospinal fluid and plasma proteomics/metabolomics using LASSO regression, producing a diagnostic panel (ST1A1, sphinganine, 7,8-dihydroneopterin) that identified neurological PASC with an AUC of 0.86-pointing toward objective, measurable neuroinflammation and blood-brain barrier dysfunction (Chen et al., 2024; Table 4; Figure 2).

4.4 Clinical and Translational Synthesis: Moving Toward Biomarker-Guided Care

Taken as a whole, the integration of clinical regression modeling with high-dimensional machine-learning classification demonstrates, we think fairly convincingly, that Long COVID is a multi-system network disorder driven by objective, reproducible pathophysiological pathways rather than a diffuse, symptom-based construct (Groysman, 2026). These findings sit in some tension with the historical, subjective-consensus frameworks used to classify post-viral syndromes (Groysman, 2026; Spanoghe et al., 2026): as Spanoghe et al. (2026) argue, classifying these patients under "Functional Somatic Disorder" or somatoform labels overlooks a large and growing body of objective, replicated biomedical evidence. The clinical regression models summarized in Table 3 show that chronic fatigue and neuropsychiatric symptoms are directly predicted by acute-phase tissue hypoxia and peak inflammatory response (Al-Hakeim et al., 2022), while the machine-learning architectures in Table 4 and Figure 2 show that patients can be classified with high precision-AUCs reaching 0.95-0.96-using objective proteogenomic, complement, and metabolomic signatures (Ai et al., 2024; Lin et al., 2024).

Consequently, this synthesis argues for a fairly fundamental restructuring of clinical trial and management pathways (Groysman, 2026). Rather than enrolling heterogeneous, symptom-defined cohorts into single therapeutic trials-an approach that, almost by design, dilutes therapeutic signal and likely contributes to repeated trial failure-future designs should employ mechanism-anchored stratification (Groysman, 2026). Using the machine-learning classifiers and biomarker panels validated here (Table 4; Figure 2), patients could plausibly be pre-screened and matched to therapies targeting their dominant pathophysiological endotype: complement inhibitors for those with active complement signatures, therapeutic apheresis for those with high amyloid microclot burden, or metabolic and mitochondrial support for those with pronounced bioenergetic and glutathione peroxidase depletion (Groysman, 2026; Faghy et al., 2026).

5. Discussion

5.1 From Symptom Cluster to Biological Network

Read together, the evidence assembled here suggests something a little more hopeful than the "200 symptoms, no answers" framing that has dogged Long COVID coverage since 2020. What looks, symptomatically, like chaos resolves-once nitro-oxidative, inflammatory, and multi-omics markers are applied-into a fairly small number of recognizable biological endophenotypes (Table 1; Table 2). The convergence of gut-derived LPS signaling, mitochondrial mtDNA leakage, and cytokine amplification onto a single NF-κB hub (Figure 1) is, we think, the single most important organizing finding in this literature: it offers a plausible mechanistic explanation for why so many ostensibly unrelated symptoms-fatigue, brain fog, dysautonomia, gastrointestinal upset-tend to travel together in the same patients.

5.2 Post-Exertional Malaise as a Convergent Endpoint

Perhaps the most clinically consequential implication concerns post-exertional malaise. PEM is not, on this evidence, a psychological phenomenon or a matter of deconditioning; it appears to be the predictable output of at least three converging mechanisms documented in Tables 1 through 3-mitochondrial bioenergetic failure, amyloid microclot-driven microvascular hypoperfusion, and the TKP-mediated neurotoxic bottleneck described in Section 2.2. The regression models in Table 3 make this causal chain unusually explicit: acute-phase hypoxemia predicts chronic redox imbalance, which predicts fatigue and neuropsychiatric severity, in a pathway that leaves little room for a purely psychosomatic account (Al-Hakeim et al., 2022). This matters enormously for management: as Section 2.6 discusses, Graded Exercise Therapy-built on exactly the deconditioning assumption this evidence contradicts-appears capable of directly worsening the biological lesion it was meant to treat (Faghy et al., 2026).

5.3 What the Machine Learning Evidence Actually Adds

It would be easy to treat the AUC values in Table 4 and Figure 2 as a technical curiosity, but we think they carry a more direct clinical implication: objective, biomarker-based diagnosis is not a distant aspiration-it is already achievable with existing methods. AUCs of 0.90-0.96 for multi-omics and neutrophil-subtype classifiers (Ai et al., 2024; Lin et al., 2024; Figure 2) compare favorably with diagnostic thresholds accepted in many other areas of medicine. What stands out across Figure 2, though, is the variability by omics layer: proteomic complement-based models cluster somewhat lower (AUC 0.785-0.857) than transcriptomic and metabolomic models (0.86-0.95), which may reflect either genuine differences in how strongly each biological layer encodes disease state, or simply differences in cohort size and modeling maturity across these still-young research programs. We would be cautious about over-interpreting this gap without larger, harmonized external validation cohorts.

5.4 Reconciling Mechanism-Anchored Subtypes with a Unified Hub

One tension worth naming directly: Section 2.1 argues for a single unifying NF-κB hub (Figure 1), while Section 4.3 and Table 4 describe five or more distinct molecular endotypes (Ai et al., 2024). These are not, we think, actually contradictory. A shared downstream inflammatory pathway is entirely compatible with different upstream drivers-viral persistence in one patient, gut dysbiosis in another, complement dysregulation in a third-converging on the same final common pathway while producing measurably different upstream biomarker profiles. If anything, this dual structure strengthens the case for mechanism-anchored stratification (Groysman, 2026): therapies aimed at the shared downstream hub (e.g., broad anti-inflammatory approaches) might help many patients somewhat, while therapies aimed at the specific upstream driver identified in Table 4 (complement inhibitors, apheresis, antivirals) could plausibly help a well-matched subgroup much more.

5.5 Sex-Specific Biology and Trial Design

The sex-specific signatures described in Section 2.5-heightened T-cell exhaustion and herpesvirus reactivation in women, versus NK-cell and TGF-β fibrotic signaling in men (Paval et al., 2025)-are, we suspect, still under-accounted for in current trial designs. Given that women carry a 1.4-fold higher risk of persistent post-viral fatigue (Struhal & Almamoori, 2025), any mechanism-anchored stratification framework that ignores sex as a stratifying variable risks reproducing exactly the signal-dilution problem it is meant to solve. This seems, to us, a fairly straightforward and low-cost improvement to build into the trial designs proposed in Section 4.4.

5.6 Toward Biomarker-Guided Clinical Trials

The translational argument that emerges from this synthesis is, at bottom, fairly simple: heterogeneous, symptom-defined trial cohorts dilute therapeutic signal, and objective biomarkers (Tables 1-4) now exist to prevent this dilution (Groysman, 2026). Complement inhibitors could reasonably be trialed in patients selected via the complement-dysregulation signatures in Table 4 (Baillie et al., 2024); therapeutic apheresis could be reserved for patients with confirmed amyloid microclot burden (Kruger et al., 2022; Pretorius et al., 2021); and mitochondrial or metabolic support could be prioritized for patients matching the OSTOX/ANTIOX profile in Table 1 (Al-Hakeim et al., 2022). None of this is speculative machinery-each biomarker panel discussed here has already been validated, at least in single cohorts, and the remaining work is largely one of trial design and cross-institutional replication rather than fundamental discovery.

5.7 Limitations

This synthesis has real limitations that are worth stating plainly. Much of the quantitative evidence in Tables 1 through 3 derives from a single, well-characterized cohort (Al-Hakeim et al., 2022), and while the pattern it describes is internally consistent, broader external replication across more diverse populations remains, as far as we can tell, limited. The machine-learning AUCs summarized in Figure 2 come from studies of varying sample size and validation rigor-some using independent external cohorts, others relying on internal cross-validation alone-and we did not have access to a uniform risk-of-bias assessment across all of them. Finally, this review draws on a curated rather than fully exhaustive literature search; a formal systematic review with dual-reviewer screening would be a valuable next step, particularly to quantify how consistently the NF-κB-centered mechanistic model in Figure 1 holds across independent research groups.

6. Conclusion

Pulling the threads of this synthesis together, Long COVID looks less like one disease than a family of related, biologically distinguishable conditions that happen to share a final common inflammatory pathway. Persistent viral antigen, gut-barrier compromise, mitochondrial injury, and microclot-driven hypoxia all appear to converge on NF-κB signaling, yet the relative weight each mechanism carries seems to differ meaningfully from patient to patient (Figure 1). Machine learning models built across transcriptomic, proteomic, and metabolomic layers can already tell these subgroups apart with respectable accuracy (Figure 2; Table 4), which is arguably the strongest argument yet for abandoning purely symptom-based trial enrollment. If the field can commit to mechanism-anchored stratification-matching complement inhibitors, apheresis, or mitochondrial support to the patients whose biology actually predicts a response-both the reproducibility of future trials and, more importantly, patient outcomes stand to improve considerably.

Author Contributions

S.F. contributed to the conception and design of the review, literature search, analysis and synthesis of the relevant evidence, and drafting of the manuscript. R.P. contributed to the literature search, interpretation of the mechanistic and biomarker evidence, and critical revision of the manuscript. U.J. contributed to the analysis and interpretation of clinical-translational and multi-omics evidence and critically revised the manuscript for important intellectual content. All authors reviewed and approved the final version of the manuscript and agreed to be accountable for all aspects of the work.

Acknowledgements

The authors would like to acknowledge the Department of Clinical Pharmacy, Faculty of Medical Sciences, Universitas Baiturrahmah, Padang, Indonesia, and the Department of Pharmacy, Kalinga University, Naya Raipur, Chhattisgarh, India, for their academic and institutional support. The authors also acknowledge the researchers whose published studies contributed to the scientific foundation of this review.

References