Long COVID: What Research Says About Causes and Treatments

Picture a section of small intestine taken at endoscopy months after a mild SARS-CoV-2 infection. Under the microscope, scattered among otherwise unremarkable epithelial cells, a few stain positive for viral nucleocapsid or spike protein. There's no florid inflammation and no replicating virus you could culture. A small amount of antigen is simply still there, long after the nasopharyngeal swab turned negative.

That image, which several groups have now produced from gut, lymphoid and other tissues, is the most useful starting point I know for thinking about long COVID. It doesn't explain the condition. It does show that "the infection is over" and "the body is back to baseline" are two different claims, and much of the field's work since 2021 has gone into the gap between them.

Here's where I think the evidence actually stands, sorted by how much weight it can bear.

What we're actually trying to explain

Long COVID is a clinical umbrella, not a diagnosis with a biomarker. The World Health Organization published a consensus definition in 2021, and the US National Academies proposed a broader one in 2024 that deliberately doesn't require laboratory confirmation of the original infection. Both describe symptoms persisting or emerging after SARS-CoV-2 infection that can affect almost any organ system. The most commonly reported are fatigue, post-exertional malaise, cognitive difficulties, dysautonomia and breathlessness.

That breadth is a methodological problem before it's a biological one. A cohort defined as "any symptom at 12 weeks" will mix people with post-ICU syndrome, people with new organ damage (myocarditis, pulmonary fibrosis, new-onset diabetes), and people with an ME/CFS-like picture and no structural lesion. My strong view is that much of the confusion in the treatment literature comes from testing mechanism-specific drugs in mechanism-agnostic populations. Keep that in mind as you read the trial section.

Established: what large cohorts and meta-analyses support

Long COVID is real, common and measurable at the population level

Large electronic health record cohorts, including the US Veterans Affairs analyses, along with national surveys and prospective community studies, consistently show excess risk of new cardiovascular, metabolic, neurological and psychiatric diagnoses after infection compared with uninfected controls. Estimates of prevalence vary enormously with definition, comparator and era. That's expected, and you shouldn't anchor on any single number. What's robust is the direction of the effect and the fact that it shows up across different health systems and study designs.

Risk has fallen, and vaccination helps

Meta-analyses of observational data agree that vaccination before infection is associated with lower risk of long COVID, and risk appears lower with Omicron-era infections than with earlier variants. Reinfection carries additional risk. These remain observational findings with the usual problems of confounding (vaccinated people differ from unvaccinated ones in many ways), but the consistency across settings is reassuring. The PubMed literature on vaccination and long COVID risk is large enough now that pooled estimates are reasonably stable.

Cognitive effects are detectable, even after mild illness

The UK Biobank imaging study, which scanned people before and after infection, found small but measurable structural brain changes on average. Large online cognitive testing studies have found deficits that scale with illness severity and persistence of symptoms. The effect sizes at population level are modest. For an individual with persistent symptoms, they may not be. Population averages and individual experience are different questions, and both deserve answers.

Emerging: mechanistic studies with real signal

This is where most of the excitement lives, and where I'd urge the most care. These studies are usually small, cross-sectional and case-control, and they often draw on cohorts recruited through specialist clinics. They're hypothesis-generating. Several are also very good.

Viral persistence

Beyond the tissue findings above, several groups have detected circulating spike or other viral antigens in a subset of people months after infection, more often in those with persistent symptoms. Autopsy work has found viral RNA widely distributed, including in the brain, well beyond the acute phase. Studies on persistent SARS-CoV-2 reservoirs now form a coherent body of work.

What remains unsettled is whether persistence is a cause or a marker. Antigen in tissue could drive chronic immune activation. It could also be a bystander finding that's common in people who recover fully. The clearest test is therapeutic: if persistence matters, clearing it should help someone. So far, that test hasn't come back clearly positive (see below).

Immune dysregulation

Deep immune profiling studies, including a widely cited 2023 Nature paper from the Yale and Mount Sinai groups, have reported differences between people with long COVID and recovered controls. These include altered T cell and monocyte populations, signs of reactivation of latent herpesviruses such as Epstein-Barr virus, and, strikingly, lower morning cortisol. Machine-learning classifiers built on these features separated groups reasonably well.

I take these findings seriously, but classifiers trained on a few hundred people tend to shrink on external validation. The low-cortisol result is interesting precisely because it's cheap to measure and easy to replicate, and it's one I'd like to see tested in a large population-based cohort.

Autoimmunity

Autoantibodies against a range of targets, including G-protein-coupled receptors, have been reported in long COVID. The more persuasive recent work uses passive transfer. IgG purified from patients and injected into mice has reproduced aspects of the phenotype, such as pain sensitivity or reduced activity, in some experiments. That's a meaningful step up from correlation. It's still animal work, the effects are partial, and not every patient's IgG does it. My reading is that autoimmunity probably explains a subset of cases, not the syndrome as a whole.

Thromboinflammation, complement and fibrin

Proteomic studies have reported complement dysregulation and markers of thromboinflammation in people with persistent symptoms. Separate work has shown that the spike protein can bind fibrinogen and promote abnormal fibrin formation that is pro-inflammatory, including in the brain in animal models. The "microclot" hypothesis, which rests largely on fluorescence microscopy of platelet-poor plasma, has attracted attention and criticism in roughly equal measure. The method needs standardization and blinded replication before it can carry clinical weight.

Serotonin, the gut and the vagus nerve

A 2023 Cell paper proposed a chain: persistent viral material drives type I interferon signaling, which reduces intestinal tryptophan uptake and therefore circulating serotonin, which in turn impairs vagal signaling and cognition. It's an elegant model with supporting mouse data. It's also a long causal chain, and each link needs independent confirmation in humans.

Muscle and post-exertional malaise

A study that took muscle biopsies before and after an exercise challenge found mitochondrial dysfunction, amyloid-containing deposits and tissue damage in people with long COVID and post-exertional malaise. For me, this is one of the more clinically important mechanistic findings. It gives a biological basis for something patients have reported for years, and it has direct implications for how rehabilitation should be designed.

Early: where the ground is still soft

Several hypotheses rest mainly on animal models, in vitro work or very small human samples. These include brainstem dysfunction, persistent microglial activation, endothelial injury as a unifying lesion, and dysbiosis as a primary driver. Some will mature. Many won't. I'd treat any single-lab finding in this tier as a lead, not a result. Be especially wary of models built around one assay that hasn't been run blinded across sites.

One honest observation: almost every mechanism above can be connected to almost every other. Persistent antigen drives immune activation, which drives autoantibodies, which drive endothelial injury, which drives clotting, and so on. That interconnectedness makes for satisfying review figures. It also makes the field vulnerable to post hoc storytelling. The models that deserve the most attention are the ones that make a falsifiable, therapeutically testable prediction.

Treatment trials so far

Here the summary is short and sobering. No pharmacological treatment has robust randomized evidence of benefit for established long COVID.

The null antiviral trials deserve a careful reading rather than a dismissive one. They don't disprove the persistence hypothesis. They show that a specific drug, given for a specific duration, to a broadly defined population, didn't move symptom scores enough to detect. If persistence matters in only a fraction of patients, an unselected trial could easily miss it. Ritonavir's interaction profile also limits how long and how aggressively you can dose. I'd much rather see the next antiviral trials enrich for participants with detectable antigen.

The metformin finding is promising but narrow. It concerns prevention during acute infection, in a specific trial population, and it came from a prespecified secondary outcome. It needs confirmation before it changes practice.

On rehabilitation, I'll say plainly what I think. Exercise programs designed around progressive increases in load, applied to people with post-exertional malaise, risk causing harm. The muscle biopsy work makes that concern biological rather than merely anecdotal. Pacing and symptom-titrated approaches are the defensible default until better trials exist. You can track active and completed studies on ClinicalTrials.gov, and the NIH's RECOVER initiative remains the largest coordinated US effort.

Limitations that run through the whole literature

• Case definitions differ between studies, so prevalence and effect sizes rarely compare cleanly.

• Controls are often wrong. "Recovered" controls are not the same as uninfected ones, and many people have now had undocumented infections.

• Selection bias is common. Clinic-based cohorts overrepresent severe, persistent and well-resourced patients.

• Outcomes are mostly self-reported, which matters a great deal in unblinded or small trials.

• Mechanistic studies are cross-sectional, so it's hard to tell cause from consequence.

The open debates follow from these limitations. One is whether long COVID is one disease or several. Another is how much it overlaps with ME/CFS and other post-infectious syndromes. A third is whether the mechanisms converge on a shared final pathway or stay separate. I lean toward several overlapping subtypes. I also think the ME/CFS research community was right about post-exertional malaise long before long COVID put it in high-impact journals, and the field should acknowledge that more often than it does.

What I'd watch next

A few developments would genuinely change my thinking:

1. Biomarker-stratified trials. An antiviral or monoclonal trial restricted to antigen-positive participants would be the most informative experiment in the field. Positive or negative, it would tell us something.

2. Replication of immune signatures in population cohorts. Cortisol, herpesvirus reactivation markers and complement profiles all need testing outside specialist clinics.

3. Passive transfer at scale. If IgG transfer reliably reproduces a phenotype, B cell-directed therapies become a rational subgroup strategy.

4. Objective endpoints. Actigraphy, cardiopulmonary exercise testing with repeat-day protocols, and validated cognitive batteries would make small trials far more interpretable.

5. Cross-condition work with ME/CFS. Shared mechanisms would shift funding and trial design for both conditions.

If you're trying to keep up, the volume is the hard part. Hundreds of papers appear each month, and the ones that matter are often not the most cited. I find it helps to keep a living evidence map organized by tier, as above, rather than by mechanism. I keep mine in PACR, an AI-native research workspace that keeps literature search, reading notes and drafting in one place. That way, when a new trial reports, I can see immediately which hypothesis it bears on and what else sits at that level of evidence. You could start the same way with a focused query through PACR's research search, such as "long COVID randomized trial", and sort what comes back by design before you sort it by topic.

The image I opened with, a few antigen-positive cells in an otherwise normal biopsy, is a good reminder of where we are. We can see things we couldn't see five years ago. We don't yet know which of them matter. The next round of trials should be built to tell us. If you're designing one, or reviewing the literature for a grant, set up a workspace in PACR and build your evidence map before you write a word of the protocol.