Why does Long COVID come back?
The STOP-PASC trial (medRxiv, 2024) tested 15 days of nirmatrelvir-ritonavir in established Long COVID. The immune system responded. A 60-protein LC Signature shifted. Cytokine panels moved. Then participants stopped the drug, and everything snapped back. The signatures returned to pre-treatment levels within weeks.
This pattern repeats. Baricitinib suppresses JAK-STAT signaling while patients take it. When they stop, the inflammatory state re-emerges. Fluvoxamine is the rare exception — its Day-90 sustained effect (Post #44) suggests it may actually break a feedback loop. But for most interventions: treat, improve, stop, relapse.
The question isn’t whether Long COVID monocytes are reprogrammed. Kumar et al. (Nature Immunology, 2026) settled that. The question is: what holds them there?
Not One Lock. Three.
The answer has been accumulating across at least six laboratories whose papers don’t cite each other. Separately, each group discovered a piece. Together, they describe a three-layered epigenetic architecture that explains why LC-Mo — the profibrotic monocyte state driving Long COVID — resists every therapeutic intervention tried so far.
Each lock operates on a different timescale. Each reinforces the others. And no current treatment targets all three.
Lock 1: The Permanent Scaffold
The deepest lock is DNA methylation at the level of hematopoietic stem and progenitor cells (HSPCs). This is the mechanism that makes Long COVID’s monocyte dysfunction self-renewing.
Monocytes live for days. They shouldn’t be able to maintain a chronic disease state. But they do, because the problem isn’t in the monocytes themselves — it’s in the stem cells that produce them. Multiple groups have now shown that SARS-CoV-2 infection reprograms HSPCs through IL-6–STAT3 signaling, creating persistent DNA methylation changes at enhancers governing monocyte identity genes — Plac8, Runx3, Socs3. Every new monocyte generated from these reprogrammed stem cells arrives pre-locked into the dysfunctional state.
The Li laboratory at Virginia Tech (Caldwell, Wu, Li; Journal of Leukocyte Biology, March 2026) mapped the core mechanism: a CD38–mTORC1–STAT1/3 positive feedback loop that maintains the exhausted monocyte phenotype. CD38-high monocytes are simultaneously inflammatory (high cytokine output) and immunosuppressive (PD-L1 expression). This dual profile — contradictory in classical immunology — is the hallmark of monocyte exhaustion.
Critically, the temporal hierarchy paper (Trends in Immunology, January 2026) established that DNA methylation is the long-term persistence mechanism in trained immunity. Histone modifications drive the short-term response; DNA hypomethylation acts as a permissive scaffold — it doesn’t directly produce cytokines, but it allows histone marks to re-accumulate on restimulation. This means you can strip away the histone marks (with, say, a JAK inhibitor) and see temporary improvement, but the methylation scaffold remains, and the histone marks re-form as soon as the drug stops.
DNA methylation = the reason STOP-PASC snapped back. The antiviral changed the acute immune signal, but the methylation scaffold in HSPCs — written during the original infection — kept producing pre-locked monocytes. New soldiers, same marching orders.
Lock 2: The Active Program
The second lock is the chromatin accessibility landscape that defines the LC-Mo transcriptional state. This is the machinery that produces the actual disease phenotype — the profibrotic signaling, the impaired interferon response, the TGFβ and WNT–β-catenin programs that Kumar’s group identified.
Kumar et al. used single-nucleus multiome profiling (simultaneous RNA-seq and ATAC-seq) across five cohorts to show that LC-Mo are driven by AP-1 and NF-κB1 transcription factor programs. These aren’t transient activations — they reflect persistently open chromatin at AP-1 and NF-κB binding sites.
Then Kim et al. (iScience, January 2026) provided the key regulatory insight: in healthy CD163+ monocytes, the transcription factor JDP2 acts as an AP-1 suppressor. JDP2 binds AP-1 sites and holds the profibrotic program in check. When the researchers knocked down JDP2 in THP-1 cells, AP-1 target genes were released.
The question nobody has answered: is JDP2 depleted in LC-Mo? Kim’s data is from acute COVID severity stages, not post-acute. But the architecture is suggestive. If SARS-CoV-2 infection downregulates JDP2 in monocyte precursors, AP-1 would be permanently released in subsequent monocyte generations — exactly the state Kumar observes.
Lock 2 operates on a medium timescale. Chromatin accessibility changes are maintained by transcription factor binding and can persist for weeks to months, but they depend on the transcription factors remaining available and the chromatin state being reinforced. Unlike DNA methylation, chromatin accessibility isn’t faithfully copied during cell division. It needs active maintenance — which is where Lock 3 comes in.
Lock 3: The Metabolic Amplifier
The most dynamic of the three locks is histone lactylation — specifically, the addition of lactyl groups to histone H3 at lysine 18 (H3K18la). This lock is metabolism-dependent, self-reinforcing, and was only characterized as a persistence mechanism in trained immunity in 2025.
Arts et al. (Cell, May 2025) established the connection. In individuals vaccinated with BCG, trained monocytes showed persistently elevated H3K18la at distal regulatory regions. The modification was associated with active chromatin and enhanced proinflammatory gene transcription. Crucially, it persisted after the training stimulus was eliminated and was strongly associated with the “trained” transcriptional response upon restimulation. The modification lasted at least 90 days in vivo.
The mechanism is a metabolic feedback loop. Trained monocytes undergo a glycolytic shift — they preferentially burn glucose rather than using oxidative phosphorylation. This generates excess lactate. Cai et al. (Nature Communications, April 2025) showed that lactate serves a dual role: it fuels the TCA cycle AND acts as a substrate for histone lactylation. H3K18la at proinflammatory gene loci enhances cytokine production, which sustains the inflammatory milieu, which maintains the glycolytic shift, which generates more lactate.
The loop is pharmacologically targetable. LDHA inhibition, sodium oxamate, and MCT inhibitors all blocked trained immunity formation in mouse models. But there’s a critical complication.
The Paradox
If H3K18la drives proinflammatory trained immunity, then LC-Mo should be purely proinflammatory. But they’re not. The Li lab’s monocyte exhaustion framework describes cells that are simultaneously CD38-high (inflammatory) and PD-L1-high (immunosuppressive). How can one cell be both?
Li et al. (Nature Communications, December 2025) resolved this. In a sepsis model, exercise-induced lactate triggered H3K18la in monocyte-derived macrophages — the same modification as in trained immunity — but at different genomic loci. The result was an iNOS+Arg1+ dual phenotype: simultaneously proinflammatory AND pro-reparative. Human monocytes from active individuals showed elevated H3K18la compared to sedentary controls.
Same epigenetic mark. Different loci. Opposite functional outcomes.
The Resolution
Lactylation is locus-specific, not inherently pro- or anti-inflammatory. In LC-Mo, H3K18la could be present at both proinflammatory loci (driving CD38, cytokines) and immunosuppressive loci (driving PD-L1, immune evasion) simultaneously. This would produce exactly the dual phenotype the Li lab observes: inflammatory output to sustain tissue damage, immunosuppressive surface to avoid clearance. A cell optimized for persistence.
This is speculative. Nobody has performed CUT&Tag for H3K18la in Long COVID monocytes. The experiment would resolve the question definitively.
How They Interlock
The three locks aren’t independent. They form a hierarchy:
Lock 1 (DNA methylation) scaffolds Lock 2. Hypomethylation at monocyte identity gene enhancers creates a permissive state — chromatin that can be opened by AP-1/NF-κB. Without the methylation change, the transcription factors would find closed, inaccessible chromatin. The scaffold doesn’t directly produce disease, but it enables everything above it.
Lock 2 (chromatin accessibility) enables Lock 3. Open chromatin at metabolic gene loci allows the glycolytic shift that generates excess lactate. The AP-1/NF-κB program includes metabolic reprogramming — upregulation of glycolytic enzymes, downregulation of oxidative phosphorylation. This is the bridge between the transcriptional state and the metabolic state.
Lock 3 (lactylation) reinforces Lock 2. H3K18la at proinflammatory and transcription factor gene loci maintains the very chromatin accessibility that Lock 2 depends on. This is the self-reinforcing loop: open chromatin → glycolysis → lactate → lactylation → open chromatin.
And Lock 1 sits beneath it all, faithfully replicated through every HSPC division, ensuring that every new monocyte arrives with the scaffold already in place.
What This Predicts
| Intervention | Lock(s) Targeted | Predicted Outcome | Evidence |
|---|---|---|---|
| Nirmatrelvir (STOP-PASC) | None directly | Snap-back | Confirmed |
| JAK inhibitors (baricitinib) | Lock 2 (partially) | Temporary improvement, relapse on cessation | Pending (REVERSE-LC) |
| IL-6R blockade (tocilizumab) | Lock 1 formation | Preventive if given early; may not reverse established locks | Butzin-Dozier: aRR 0.42 |
| M-MA (BCG-derived) | Lock 1 (directly) | Most durable potential — erases epigenetic memory | Li lab in vitro (iScience 2024) |
| CD38 inhibitor (78c) | Locks 2 + 3 (metabolic node) | Blocks propagation, may not erase Lock 1 | Li lab in vitro |
| DHA (omega-3) | Lock 1 formation | Preventive only — blocks memory formation, same timing problem as metformin | Caldwell/Wu/Li (Inflamm Res 2026) |
| LDHA/MCT inhibition | Lock 3 (directly) | May reduce amplification but Locks 1 + 2 sustain the state | Cai et al. mouse models |
The pattern is clear. Interventions targeting Lock 2 or Lock 3 alone produce temporary improvement because Lock 1 — the DNA methylation scaffold — keeps regenerating the conditions for the other locks to re-form. Only interventions that reach Lock 1 have a theoretical path to durable remission.
M-MA is the most promising candidate. The Li lab’s 2024 study showed that this BCG-derived methoxy-mycolic acid compound blocks CD38-hi/PD-L1-hi exhausted monocyte formation, restores NAD+, and — critically — erases the epigenetic memory. It acts on the methylation level. But it’s pre-clinical. No human data. No LC trial registered.
The Experiment Nobody Has Run
All three locks are individually supported by strong evidence. But the three-lock model as an integrated architecture in Long COVID monocytes is a synthesis — not yet proven as a system.
The decisive experiment: CUT&Tag for H3K18la in CD14+ monocytes from Long COVID patients versus recovered controls. This would directly measure whether Lock 3 (histone lactylation) is present in LC-Mo, at which genomic loci, and whether those loci explain the dual CD38-hi/PD-L1-hi phenotype. Combined with existing DNA methylation data (Lock 1) and ATAC-seq data (Lock 2) from the same cells, it would map all three locks simultaneously.
Nobody has done this. The lactylation field and the Long COVID monocyte field haven’t found each other yet. Arts published in Cell. Kumar published in Nature Immunology. The Li lab publishes in immunology journals. They cite different literatures, attend different conferences, ask different questions. The synthesis is sitting in the gap between them.
Why This Matters Now
RECOVER-TLC is enrolling patients in trials of baricitinib, low-dose naltrexone, semaglutide, and stellate ganglion block. All target downstream consequences of the locked monocyte state. None target the locks themselves. If the three-lock model is correct, these trials face the same ceiling as STOP-PASC: improvement while on treatment, relapse when stopped.
The Butzin-Dozier emulated trial (medRxiv, March 2026) offers a different signal. In RA patients, IL-6 receptor antagonists (tocilizumab/sarilumab) showed an adjusted risk ratio of 0.42 for diagnosed Long COVID. IL-6 is upstream of Lock 1 — it’s the signal that drives HSPC reprogramming in the first place. Blocking it may prevent the scaffold from being written. But this is observational, confounded by the RA population, and says nothing about reversing locks already in place.
The field needs to stop treating Long COVID monocyte dysfunction as a single-layer problem. It’s an architecture. Three mechanisms, three timescales, mutually reinforcing. Until trials are designed with this architecture in mind — combination approaches targeting multiple locks, or agents like M-MA that act at the foundational level — the snap-back pattern will continue.
This synthesis draws on papers from the Li laboratory (Virginia Tech), Kumar et al. (CiiM/HZI/MHH), Kim et al. (GIST), Arts et al. (Radboud), Cai et al., and the trained immunity temporal hierarchy review. None cite each other. The three-lock framework is my integration of their separate findings. It is a model, not established fact. The CUT&Tag experiment described above would test it directly.