Toward a Precision-Medicine Framework for Post-Acute Sequelae of SARS-CoV-2 Infection<\/p>\n\n\n\n
John Murphy, CEO, The COV-19 Long haul Foundation<\/p>\n\n\n\n
Abstract<\/p>\n\n\n\n
Background: Long COVID, or post-acute sequelae of SARS-CoV-2 infection (PASC), encompasses a heterogeneous collection of persistent, relapsing, and sometimes newly emerging clinical manifestations. Symptom-based classification systems and statistical clustering have identified reproducible patterns of illness, but whether these patterns correspond to biologically distinct disease processes and predict differential treatment response remains uncertain.<\/p>\n\n\n\n
Methods: This perspective synthesizes major developments in clinical phenotyping, symptom-based classification, biomarker research, and therapeutic trials, with particular attention to the distinction between descriptive subtypes, mechanistically defined endotypes, and treatment-predictive classifications.<\/p>\n\n\n\n
Evidence synthesis: Large prospective cohorts, including the National Institutes of Health RECOVER-Adult study, have established that Long COVID can be divided into symptom clusters associated with differences in functional impairment and quality of life. A systematic review encompassing 64 studies and approximately 2.43 million participants documented substantial variation in the methods used to classify patients. Nevertheless, symptom clusters are not necessarily biologically discrete entities, and associations between phenotypes and demographic or clinical characteristics do not establish that a particular treatment will be effective in a given subgroup. Recent randomized trials, including the RECOVER-VITAL trial of extended nirmatrelvir\u2013ritonavir, underscore the importance of testing therapeutic hypotheses prospectively rather than inferring efficacy from proposed mechanisms alone.<\/p>\n\n\n\n
Conclusions: Clinically actionable subtyping of Long COVID is a plausible but unproven objective. Its achievement will require standardized phenotyping, longitudinal validation, reproducible biomarkers, independent replication, and randomized trials designed to detect treatment-effect heterogeneity. The central test of a precision-medicine classification should be whether using it to select treatment improves patient outcomes compared with well-conducted, symptom-guided usual care.<\/p>\n\n\n\n
Introduction<\/p>\n\n\n\n
The clinical heterogeneity of Long COVID presents a fundamental challenge to medicine. Patients with the same diagnostic designation may experience profound exertional intolerance, cognitive dysfunction, autonomic instability, persistent respiratory symptoms, olfactory impairment, or combinations of manifestations involving several organ systems. Some improve gradually; others experience relapses, delayed symptom exacerbation after exertion, or persistent functional impairment. These differences may reflect variation in disease mechanisms, host susceptibility, the consequences of acute infection, coexisting conditions, or several interacting processes.\u00b9<\/p>\n\n\n\n
The practical question is whether this heterogeneity can be organized into clinically meaningful subtypes that do more than describe the distribution of symptoms. Can patients be classified into reproducible groups that identify the dominant mechanisms of disease, predict subsequent clinical trajectories, and, most importantly, determine which treatments are likely to confer benefit?<\/p>\n\n\n\n
The distinction is consequential. A classification that describes symptom burden may improve epidemiological research and communication. A prognostic classification may identify patients at increased risk of persistent disability. A mechanistic classification may identify pathways suitable for investigation. A predictive classification, however, must identify differences in treatment effect: patients assigned to one subtype must derive a different relative or absolute benefit from a treatment than patients assigned to another subtype, under appropriately controlled conditions.<\/p>\n\n\n\n
These objectives are related but not interchangeable. The existence of recognizable symptom clusters does not establish the existence of discrete biological diseases, and evidence of a plausible mechanism does not demonstrate that targeting it will improve clinical outcomes. The development of precision medicine for Long COVID therefore requires a progression from descriptive classification to mechanistic validation and, ultimately, prospective demonstration of clinical utility.<\/p>\n\n\n\n
Existing evidence for clinical subtypes<\/p>\n\n\n\n
Symptom-based classification<\/p>\n\n\n\n
One of the most substantial advances in Long COVID research has been the development of standardized research definitions and symptom-based indices. In 2023, Thaweethai and colleagues analyzed data from 9764 participants in the RECOVER-Adult cohort and identified 37 symptoms occurring more frequently among participants with prior SARS-CoV-2 infection at least six months after infection. A preliminary symptom score was developed to identify participants with PASC, and cluster analysis identified four symptom-defined subgroups.\u00b2<\/p>\n\n\n\n
These groups included participants characterized predominantly by altered smell or taste; fatigue and post-exertional malaise; cognitive symptoms accompanied by fatigue and post-exertional malaise; and a broader multisystem presentation involving fatigue, exertional symptoms, dizziness, cognitive difficulties, gastrointestinal symptoms, and palpitations. Functional impairment differed across groups, suggesting that symptom combinations carry information beyond a simple count of complaints.\u00b2<\/p>\n\n\n\n
The subsequent 2024 RECOVER-Adult research index, reported by Geng and colleagues in 2025, incorporated an expanded symptom questionnaire and data from 13,647 participants. The investigators identified five symptom subtypes associated with differences in quality of life and functional status.\u00b3 This larger analysis strengthened the evidence that Long COVID is not adequately represented by a single, uniform symptom profile.<\/p>\n\n\n\n
Nevertheless, these indices were developed primarily to classify symptomatic participants for research. They were not designed to establish definitive clinical diagnoses, identify exclusive pathophysiological categories, or determine treatment selection. A symptom score may be useful for identifying research participants without demonstrating that the underlying illness can be partitioned into biologically discrete entities.<\/p>\n\n\n\n
The problem of non-equivalent classifications<\/p>\n\n\n\n
A systematic review published in eClinicalMedicine in 2026 synthesized 64 studies, including 47 cohort studies and 17 cross-sectional studies, encompassing approximately 2.43 million participants across 20 countries. The investigators identified several broad approaches to subtyping: symptom co-occurrence, affected organ system, disease severity, and clinical indicators. Symptom co-occurrence and organ-system classifications predominated.\u2074<\/p>\n\n\n\n
The breadth of this literature establishes that heterogeneity is a consistent finding. It does not, however, establish a universally applicable taxonomy. Studies differ in diagnostic definitions, recruitment strategies, follow-up intervals, symptom instruments, statistical methods, and populations. The resulting classifications may therefore describe overlapping but non-identical dimensions of illness.<\/p>\n\n\n\n
A patient classified within a cognitive symptom cluster may simultaneously have autonomic dysfunction, exertional intolerance, sleep disturbance, or evidence of another organ-system abnormality. Conversely, patients assigned to the same symptom cluster may have different underlying biological processes. A classification that assigns each patient to a single, mutually exclusive category may consequently misrepresent a disease characterized by overlapping mechanisms and changing clinical states.<\/p>\n\n\n\n
The appropriate objective may not be to identify a small number of immutable Long COVID subtypes. It may instead be to develop a multidimensional classification that captures clinical manifestations, underlying mechanisms, disease activity, comorbidity, and temporal evolution.<\/p>\n\n\n\n
From clinical phenotype to biological endotype<\/p>\n\n\n\n
A clinically useful framework should distinguish three concepts.<\/p>\n\n\n\n
A phenotype is the observable pattern of disease: symptoms, examination findings, physiological abnormalities, and functional limitations.<\/p>\n\n\n\n
An endotype is a subgroup defined by a sufficiently supported underlying biological mechanism. Examples proposed for Long COVID include persistent viral antigen or viral material, altered immune regulation, endothelial or microvascular abnormalities, autonomic dysfunction, and disturbances of energy metabolism. These mechanisms may coexist, and their relative contributions may vary among patients.\u00b9<\/p>\n\n\n\n
A treatment-predictive subtype is a classification that identifies differential response to a particular intervention. It requires evidence that the treatment effect differs between subgroups, rather than simply evidence that one subgroup has more severe disease or a different biomarker profile.<\/p>\n\n\n\n
This distinction provides a framework for interpreting emerging research. A phenotype can be reproducible without representing a unique endotype. An endotype can be biologically credible without being therapeutically actionable. A biomarker can correlate with disease severity without predicting response to treatment.<\/p>\n\n\n\n
For example, the detection of a marker associated with viral persistence would not, by itself, establish that an antiviral intervention will improve symptoms. The marker could reflect residual antigen, non-replicating viral material, an incidental finding, or a process that is no longer driving disease. Similarly, an inflammatory signature would not establish that immunomodulation is beneficial; the signature might be a consequence rather than a cause of illness, or it might identify a subgroup in which treatment carries substantial risk.<\/p>\n\n\n\n
The distinction is not semantic. It determines the evidentiary standard required before a classification can be used to guide treatment.<\/p>\n\n\n\n
Biological hypotheses and their evidentiary requirements<\/p>\n\n\n\n
Several candidate mechanisms merit investigation as potential components of a multidimensional classification.<\/p>\n\n\n\n
Viral persistence. Persistent viral antigen or viral material has been reported in some research settings, and viral persistence remains a plausible mechanism in a subset of patients.\u00b9 The critical questions are whether the findings are reproducible, whether they distinguish clinically relevant subgroups, whether the detected material is causally related to symptoms, and whether an assay can identify patients likely to benefit from a mechanism-directed intervention.<\/p>\n\n\n\n
Immune dysregulation. Studies have reported altered immune-cell profiles, inflammatory signaling, and other immune abnormalities in subsets of patients.\u00b9 Such findings may eventually support mechanistic classification. At present, a clinically validated immune endotype requires replicated assays, prespecified thresholds, temporal stability data, and evidence that the result predicts a meaningful clinical outcome or differential treatment benefit.<\/p>\n\n\n\n
Autonomic and cardiovascular dysfunction. Orthostatic intolerance, palpitations, and related manifestations can be assessed through established clinical and physiological methods. However, similar symptoms may arise from different mechanisms, and autonomic dysfunction may coexist with other disease processes. Objective testing can refine a clinical phenotype, but the extent to which a particular test identifies a treatment-responsive Long COVID endotype must be established separately.<\/p>\n\n\n\n
Metabolic and exertional abnormalities. Post-exertional malaise and exercise intolerance are important features of Long COVID, but they should not be assumed to represent one mechanism. Research should distinguish delayed symptom exacerbation from deconditioning, cardiopulmonary limitation, muscle dysfunction, sleep disorders, and other contributors. Physiological investigations must also account for the possibility that exertional testing itself may exacerbate symptoms in susceptible patients.<\/p>\n\n\n\n
These hypotheses should not be forced into a single-mechanism model. The more plausible working assumption is that some patients have one dominant mechanism, others have several interacting mechanisms, and still others have symptoms that cannot yet be assigned confidently to a defined biological process.<\/p>\n\n\n\n
Treatment trials: the decisive test of clinical subtyping<\/p>\n\n\n\n
The purpose of precision classification is not simply to produce more elaborate descriptions of illness. It is to improve decisions.<\/p>\n\n\n\n
The RECOVER-VITAL randomized, double-blind, placebo-controlled phase 2 trial provides an important example. The study enrolled 959 participants across cognitive, autonomic, and exercise-related symptom phenotypes and compared 15-day or 25-day nirmatrelvir\u2013ritonavir regimens with placebo-controlled treatment. The investigators found no statistically significant benefit for the primary endpoints in any of the three phenotypes studied.\u2075<\/p>\n\n\n\n
These results should be interpreted narrowly. They do not establish that viral persistence is irrelevant to all Long COVID, nor do they exclude potential benefit from different interventions in biologically defined subgroups. They do demonstrate that selecting participants by broad symptom phenotype and applying an antiviral regimen did not produce the expected treatment benefit in this trial.<\/p>\n\n\n\n
The implication is that symptom-defined phenotypes cannot automatically be treated as mechanistic categories. If a therapeutic hypothesis depends on a specific biological process, trial enrollment may need to incorporate a validated marker of that process rather than relying on a symptom cluster alone.<\/p>\n\n\n\n
Even that refinement requires caution. Selecting participants on the basis of a biomarker does not prove that the biomarker predicts treatment response. A biomarker may be prognostic, indicating a worse outcome regardless of treatment, without being predictive of differential benefit. To establish predictive utility, investigators must compare treatment effects across biomarker-defined groups, ideally using a prespecified treatment-by-biomarker interaction and adequate statistical power.<\/p>\n\n\n\n
A negative trial should therefore prompt examination of several possibilities: the target mechanism may not be causal; the intervention may not adequately engage the target; treatment may have been initiated at an inappropriate disease stage; the selected phenotype may have included biologically heterogeneous participants; or the proposed mechanism may matter only in a smaller subgroup. These explanations are hypotheses to be tested, not reasons to disregard negative results.<\/p>\n\n\n\n
The same principle applies to positive trials. An overall treatment benefit does not establish the necessity of subtype-based selection unless the classification improves the benefit\u2013risk balance, identifies clinically meaningful effect heterogeneity, or reduces unnecessary treatment.<\/p>\n\n\n\n
A proposed research framework<\/p>\n\n\n\n
A credible program to develop clinically validated Long COVID subtypes should proceed through sequential, prespecified stages.<\/p>\n\n\n\n
First, standardize clinical phenotyping. Multicenter cohorts should use common definitions of symptom duration, post-exertional malaise, cognitive impairment, autonomic dysfunction, respiratory limitations, and functional outcomes. Baseline assessments should capture prior infections, vaccination, acute disease severity, relevant comorbidities, medications, and important alternative diagnoses. Symptom severity and functional impairment should be measured separately.<\/p>\n\n\n\n
Second, determine longitudinal stability. Patients should be assessed repeatedly to establish whether assigned phenotypes persist, transition, or resolve. A classification that changes as symptoms fluctuate may still be clinically useful, but it should be designed to accommodate transitions rather than assume permanent membership in a single group.<\/p>\n\n\n\n
Third, integrate objective measures. Candidate biomarkers should be assessed alongside clinical phenotypes, not in isolation. These may include standardized autonomic measurements, pulmonary function, neurological assessments, immune profiling, and carefully selected measures of viral material or metabolic function. Testing should be hypothesis-driven, analytically validated, and proportionate to participant burden.<\/p>\n\n\n\n
Fourth, derive and independently validate the classification. Investigators should prespecify candidate variables, control overfitting, assess missing data and selection bias, and evaluate reproducibility across sites and populations. The model should be compared with simpler alternatives, including symptom-based classification and conventional clinical assessment. Independent validation is essential because apparent separation of clusters in the derivation cohort may not generalize.<\/p>\n\n\n\n
Fifth, test treatment prediction prospectively. Randomized trials should enroll participants across relevant phenotypes and, where feasible, stratify randomization by prespecified subtype. Analyses should test treatment-by-subtype interactions and report absolute benefits, harms, confidence intervals, and clinically meaningful functional outcomes. Subgroup findings discovered after a trial has concluded should be treated as exploratory until replicated.<\/p>\n\n\n\n
Sixth, demonstrate clinical utility. Even a reproducible treatment-predictive classifier may not improve care if it is expensive, inaccessible, unstable, or no better than routine clinical evaluation. A pragmatic trial should compare phenotype-guided treatment with high-quality usual care to determine whether classification improves outcomes, reduces adverse effects, or shortens the time to effective management.<\/p>\n\n\n\n
This framework should incorporate patient participation from the outset. Patients can identify meaningful outcomes that are missed by conventional symptom scales, including delayed exertional worsening, loss of independence, cognitive endurance, and the ability to sustain everyday activities. Clinical validation must ultimately be judged by outcomes that matter to patients, not by biomarker normalization alone.<\/p>\n\n\n\n
Statistical and methodological considerations<\/p>\n\n\n\n
Subtype discovery presents substantial risks of overinterpretation. Clustering algorithms will generally partition data when instructed to do so; the existence of clusters is not proof that nature contains discrete disease categories. Investigators should report the stability of the solution, sensitivity to variable selection and preprocessing, and performance in external cohorts.<\/p>\n\n\n\n
Continuous biological variation should not be forced into categorical labels unless the resulting categories provide a meaningful clinical advantage. Multidimensional or probabilistic classifications may better represent patients with overlapping mechanisms. A patient could, for example, have a high probability of autonomic dysfunction and a separate probability of immune dysregulation without belonging exclusively to either category.<\/p>\n\n\n\n
Prediction models must also be distinguished from causal models. A feature that predicts treatment response in one population may not identify the mechanism responsible for the response. Differences in baseline risk, concomitant disease, treatment exposure, adherence, or access to care can create apparent subgroup effects. Prespecified interaction testing, adjustment for appropriate confounding in observational studies, and independent replication are therefore essential.<\/p>\n\n\n\n
Finally, studies should account for temporal changes in SARS-CoV-2 variants, population immunity, reinfection, and prior treatment. A classification developed in an early-pandemic cohort may not retain the same calibration or clinical relevance in a contemporary population. Revalidation should be an explicit requirement rather than an assumption.<\/p>\n\n\n\n
Implications for clinical practice<\/p>\n\n\n\n
At present, symptom-based subtyping can support research recruitment, structured assessment, and communication about the diversity of Long COVID. It should not be presented as a validated method for selecting disease-modifying therapy unless the specific classification and its treatment-predictive performance have been demonstrated.<\/p>\n\n\n\n
Clinicians should continue to evaluate individual symptoms, functional impairment, comorbidities, and alternative or coexisting diagnoses. Objective findings may guide established symptom-directed care even when they do not establish a distinct Long COVID endotype. Conversely, the absence of a validated biomarker should not invalidate a patient’s illness or prevent appropriate evaluation and management.<\/p>\n\n\n\n
A precision-medicine framework should complement, not replace, clinical judgment. Its purpose is to improve the match between the patient, the suspected mechanism, and the intervention while maintaining appropriate attention to safety and uncertainty.<\/p>\n\n\n\n
Conclusions<\/p>\n\n\n\n
Long COVID can already be divided into reproducible symptom-defined subgroups for research purposes. The evidence is substantially less mature for the proposition that these groups represent distinct biological endotypes, and more limited still for the claim that a classification can reliably predict differential treatment response.<\/p>\n\n\n\n
The next phase of research should move beyond the enumeration of clusters. It should establish which clinical patterns are stable, which correspond to measurable biological processes, which biomarkers are analytically and clinically valid, and which classifications improve treatment decisions in prospective trials.<\/p>\n\n\n\n
The most important test is not whether researchers can identify subtypes. It is whether using those subtypes changes outcomes for the better.<\/p>\n\n\n\n
Precision medicine for Long COVID should therefore be regarded as a testable research program rather than an established clinical capability. Its success will depend on methodological rigor, transparent reporting of negative findings, independent replication, and a sustained commitment to outcomes that reflect patients’ daily lives.<\/p>\n\n\n\n
Footnotes<\/p>\n\n\n\n
References<\/p>\n\n\n\n
Al-Aly Z, Davis H, McCorkell L, et al. Long COVID science, research and policy. Nat Med. 2024;30:2148-2164. doi:10.1038\/s41591-024-03173-6.<\/p>\n\n\n\n
Thaweethai T, Jolley SE, Karlson EW, et al. Development of a definition of postacute sequelae of SARS-CoV-2 infection. JAMA. 2023;329:1934-1946. doi:10.1001\/jama.2023.8823.<\/p>\n\n\n\n
Geng LN, Erlandson KM, Hornig M, et al. 2024 update of the RECOVER-Adult Long COVID Research Index. JAMA. 2025;333:694-700. doi:10.1001\/jama.2024.24184.<\/p>\n\n\n\n
Wang B, Luo X, Wu M, et al. Identifying subtypes of Long COVID: a systematic review. eClinicalMedicine. 2026;91:103705. doi:10.1016\/j.eclinm.2025.103705.<\/p>\n\n\n\n
RECOVER-VITAL investigators. Nirmatrelvir\u2013ritonavir targeting viral persistence in post-COVID-19 condition (long COVID) in the USA (RECOVER-VITAL): a randomised, double-blind, placebo-controlled, phase 2 trial. Lancet Infect Dis. Published online September 2026. doi: [Verify final bibliographic record before submission.]<\/p>\n\n\n\n
<\/p>\n","protected":false},"excerpt":{"rendered":"
Toward a Precision-Medicine Framework for Post-Acute Sequelae of SARS-CoV-2 Infection John Murphy, CEO, The COV-19 Long haul Foundation Abstract Background: Long COVID, or post-acute sequelae of SARS-CoV-2 infection (PASC), encompasses […]<\/p>\n","protected":false},"author":2,"featured_media":16790,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[101,1379],"tags":[],"class_list":["post-16746","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-covid-19","category-subtypes"],"_links":{"self":[{"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=\/wp\/v2\/posts\/16746","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=16746"}],"version-history":[{"count":5,"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=\/wp\/v2\/posts\/16746\/revisions"}],"predecessor-version":[{"id":16791,"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=\/wp\/v2\/posts\/16746\/revisions\/16791"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=\/wp\/v2\/media\/16790"}],"wp:attachment":[{"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=16746"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=16746"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/cov19longhaulfoundation.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=16746"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}