Two enzymes sit on opposite sides of the same switch. DNMT3A adds methyl groups to DNA, silencing genes. TET2 initiates their removal, reactivating them. Together they regulate the methylation landscape that determines which genes a cell expresses. They are opponent arms of a single system, pushing the epigenome in opposite directions.
When either enzyme is lost to mutation in blood stem cells, the result looks similar from a distance. The mutant clone expands, outcompeting normal blood cells in a slow takeover called clonal hematopoiesis. It happens in most people over 70. Both TET2 and DNMT3A mutations drive this expansion. Both trigger inflammatory signaling. Both send mutant immune cells infiltrating into the central nervous system.
Same system. Same superficial phenotype. Opposite effects on Alzheimer's disease.
Katie Matatall and colleagues at Baylor College of Medicine published the protective side in Cell Stem Cell in August 2025. In the UK Biobank, TET2-mutant clonal hematopoiesis is associated with a 47% reduced risk of late-onset Alzheimer's. No other clonal hematopoiesis driver gene shows anything like that protection.
The mechanism, worked out in 5xFAD mice, runs against expectation. Mice transplanted with TET2-mutant bone marrow showed more myeloid cells infiltrating the brain, more activated microglia, more inflammation. Everything that should worsen a neurodegenerative disease. Instead, amyloid plaques shrank and cognitive performance improved on both novel object recognition and conditioned fear memory tests.
The reason is that TET2-mutant immune cells don't just inflame. They clear. iPSC-derived TET2-mutant microglia showed elevated phagocytosis of amyloid-beta compared to wild-type or DNMT3A-mutant microglia. The mutant cells produce hyperinflammatory chemokines (CCL2, CCL7, CCL5, CX3CL1) that recruit more phagocytes to sites of damage. The inflammation is productive. It summons cleanup crews that actually eat the debris.
Break the other arm and the picture inverts. At the American Society of Hematology meeting in December 2024, the same research group reported that 5xFAD mice transplanted with DNMT3A-mutant bone marrow showed the opposite: worsened cognitive impairment, increased anxiety, and decreased microglia activation. DNMT3A-mutant immune cells expanded, just as TET2-mutant cells did. They entered the brain. They caused inflammation. But they did not clear plaques. The inflammation was decoupled from productive phagocytosis.
The title of their ASH abstract named the finding directly: "Dnmt3a and Tet2-Driven Clonal Hematopoiesis Have Opposing Effects on the Pathogenesis of Alzheimer's Disease."
Opposing effects from the same system. Not different magnitudes of the same effect. Opposite signs.
Christopher Walsh's team at Boston Children's Hospital ran into this phenomenon from the other direction. In a 2026 Cell paper, they sequenced 311 brains, 190 from people with Alzheimer's and 121 controls. They found TET2, DNMT3A, and ASXL1 mutations enriched in Alzheimer's microglia-like cells, with inflammatory and proliferative transcriptional signatures. Walsh described the finding bluntly: Alzheimer's disease is driven by the same mutations that drive blood cancers.
But the paper was descriptive. It identified the mutations. It characterized the inflammatory signatures. It did not distinguish between the per-gene effects. The mutations appeared as a single phenomenon, clonal hematopoiesis in the Alzheimer's brain, when they were two phenomena with opposite valences. The very feature that makes them scientifically important, the sign difference, was invisible at the resolution the paper used.
This is not a criticism of Walsh's work. It is the natural result of studying clonal hematopoiesis as a category rather than as a family of mutations with directional effects. When you lump TET2 and DNMT3A together under the heading "clonal hematopoiesis," you lose the sign. The aggregate tells you mutations are present. It cannot tell you what the mutations mean.
The broader point extends beyond DNA methylation. Opponent-process architecture is everywhere in biology. Kinases add phosphate groups to proteins; phosphatases remove them. Excitatory neurons drive activity; inhibitory neurons suppress it. Sympathetic nervous system accelerates; parasympathetic decelerates. In each case, two arms push a shared variable in opposite directions, and the functional state of the system depends on the balance between them.
When you disrupt one arm of an opponent process, you don't get generic dysregulation. You get dysregulation with a specific direction. Remove the demethylator and the epigenetic landscape shifts toward hypermethylation. Remove the methylator and it shifts toward hypomethylation. Both produce clonal expansion, the shared consequence of imbalance in either direction. Both produce inflammation, the shared immune response to abnormal cells. But the downstream functional state, whether inflamed microglia clear amyloid or fail to, depends on which direction the landscape was pushed.
The surface phenotype carries no information about the sign. You have to know which arm was broken.
This is what makes the Matatall finding difficult to absorb. A 2023 Nature Medicine paper by Bouzid and colleagues had already reported that clonal hematopoiesis was associated with reduced Alzheimer's risk. But the association was aggregate, across all CHIP mutations. The field's instinct was to treat clonal hematopoiesis as a single variable that was either protective or harmful. Matatall's contribution was to show that it is neither. TET2-mutant clonal hematopoiesis is protective. DNMT3A-mutant clonal hematopoiesis is not. The aggregate statistic reflected a weighted average of opposite effects, dominated by whichever mutation was more prevalent in the cohort.
The sign was hiding inside the average.
There is a design principle buried here that goes beyond medicine. Any system built on opponent processes contains a latent asymmetry: disruption of one arm produces the opposite functional outcome from disruption of the other, even when both disruptions share a surface phenotype. In metamaterials, composing two elements with opposite rate-sign responses creates a passive switch. In epigenetics, losing one arm of a methylation/demethylation pair creates inflammation that clears debris, while losing the other creates inflammation that doesn't.
The phenotype-level description is always the same: something broke, and the system is inflamed. The sign-level description is where the information lives. It tells you whether the inflammation is productive or empty, whether the disruption is protective or pathological, whether the switch flipped toward clearance or toward failure.
The hardest part of the opponent-sign principle is that it requires abandoning phenotype-level reasoning. Inflammation is not a single thing. Clonal expansion is not a single thing. Every surface behavior that can be produced by disrupting either arm of an opponent process is two things masquerading as one. The sign is invisible until you ask which arm was lost.