Sage — Forced Differentiation Is Not a Moved Set Point
What the KAIST colon-cancer reversion experiment does and does not show
DISCLAIMER
This essay is an interpretive commentary. It draws on published experimental work and on a previously stated mitochondrial-metabolic hypothesis of cancer. It is not medical advice, not a diagnosis, and not a treatment recommendation. It does not establish that any diet, drug, supplement, or training protocol reverts cancer, and it should not be used to delay or replace care directed by a qualified physician. Individual decisions require personalized evaluation. Mechanisms discussed are hypotheses except where a specific experimental result is cited.
A team at KAIST reported that colorectal cancer cells can be pushed back toward a normal-looking intestinal identity without being killed. Professor Kwang-Hyun Cho and colleagues, writing in Advanced Science (“Control of Cellular Differentiation Trajectories for Cancer Reversion,” DOI 10.1002/advs.202402132), built a Boolean model of the gene network governing human intestinal differentiation and asked which nodes, flipped together, would send a cell back down the road it had traveled in reverse. The model named three: MYB, HDAC2, and FOXA2. Simultaneous knockdown in HT-29, HCT-116, and Caco-2 cells raised enterocyte markers and collapsed proliferation. The same pre-silenced cells, implanted as subcutaneous xenografts in nude mice, formed smaller tumors over roughly three to four weeks, and in part of the HCT-116 arm formed none. Single knockdowns were weaker. Nothing was poisoned, irradiated, or cut.
The result is real. The meaning being attached to it is not. Public accounts have treated the experiment as proof that cancer cells can change their minds, and that oncology has acquired a second verb: convert, rather than destroy. Conversion is an attractive word. It is not what was tested. What was tested is whether a malignant phenotype can be forced, for a few weeks, while three regulators are held down, in cells that were silenced before they were ever asked to form a tumor. That is a different claim, and a smaller one. Read from the mitochondrial-metabolic hypothesis, it is also the claim one would have expected.
What the experiment actually was
The network was inferred from 4,252 normal human intestinal cells, not from tumors. BENEIN, the group’s control framework, treated differentiation as a trajectory through a gene-regulatory state space and searched for a small set of nodes whose joint inhibition would move a cell toward the enterocyte attractor. MYB is a transcription factor that maintains intestinal stem-cell self-renewal and cooperates with Wnt. HDAC2 is a class I histone deacetylase that compacts chromatin and helps hold cell-cycle and anti-differentiation programs in place. FOXA2 is a pioneer factor that bookmarks endoderm and epithelial identity; in this circuit it matters as part of the triple, not as a classical oncogene. The in vitro finding is that the combination is synergistic. The in vivo finding is reduced tumorigenicity of cells knocked down before implantation.
Two design facts set the ceiling. The knockdown preceded engraftment. There was no treatment of an established tumor, no orthotopic colon, no intact immune system, and no arm in which the three genes were restored and the phenotype watched. The assay measures compliance under a clamp. It does not measure memory.
The genetic reading, and where it stops
The dominant paradigm treats cancer as a genetic disease: mutations in proto-oncogenes and tumor suppressors confer the hallmarks, and therapy wins by inhibiting the driver or killing the clone. On that reading, the KAIST result is a genuine embarrassment for the idea that the genome is a sentence. The cells still carried their mutations. The phenotype moved anyway. Proliferation fell because the cells differentiated, not because they died. A network state was shown to be sufficient to hold a malignant appearance, and insufficiently fixed by mutation alone to prevent a short-term reversal.
That is worth having. It is the same class of observation already implied by differentiation therapy in acute promyelocytic leukemia, by lineage plasticity under targeted pressure, and by the everyday fact that mutational burden and malignant behavior come apart. Clonal hematopoiesis and field cancerization put driver mutations in histologically normal tissue. Some cancers carry little mutational burden and behave badly. The KAIST experiment adds a clean experimental instance: phenotype is a state, and a state can be forced.
It does not identify the origin of the state. A control lever is not a cause. Finding that three nodes hold a cell in an immature attractor tells you where the lock is. It does not tell you who installed the lock, why that configuration was retained, or whether the cell will stay there when the lever is released. The genetic model’s error was to treat downstream molecular events as the disease. Repeating the error one level up — treating a differentiation-network controller as the disease — is not a correction. It is a relocation.
Where the metabolic hypothesis places the result
The mitochondrial-metabolic hypothesis begins one layer upstream of both stories. The root proposal is a cellular energy crisis: impaired oxidative phosphorylation and ATP supply in a precancerous or early transformed cell. That insufficiency initiates a primitive glycolysis rescue, an ancestral fermentation program suited to hypoxia and unstable fuel. In the resulting pleiomorphic population, selection then retains subclones optimized for carbon-unit acquisition and glycolytic flux, because that configuration has a fitness advantage where fuel is contested. Mutations arise secondarily from mitochondrial reactive oxygen species, replication stress, and failed quality control, and they are kept when they stabilize the carbon-efficient, apoptosis-resistant state. The Warburg phenotype is not a byproduct of growth. It is a rescue program that selection locks in.
On that account, MYB, HDAC2, and FOXA2 are not the origin. They are the information-layer face of the locked rescue.
MYB is the stemness node selection is expected to favor. It keeps the cell on the LGR5 and Wnt self-renewal program: immature, proliferative, oriented toward biomass rather than toward the ATP cost of a finished enterocyte. A differentiated enterocyte is post-mitotic and oxidative. A MYB-high crypt-like cell is the opposite. Forcing MYB down is pushing the cell off the program that carbon-capture selection had reason to keep.
HDAC2 is where the metabolic claim and the network claim actually touch. It is not a driver mutation. It is a chromatin eraser whose activity sits inside the cell’s acetyl-CoA and NAD+ economy. Mitochondrial acetyl-CoA writes acetyl marks. Class I deacetylases strip them. Sirtuins compete for NAD+. A cell under chronic ATP shortfall and high glycolytic flux is a cell whose histone code is being funded to drift. HDAC2 is a hinge between engine failure and the repressed, undifferentiated attractor. Knocking it down forces the attractor open. It does not repair the engine that was paying for the closed state.
FOXA2 is the awkward node, and it is instructive. Alone it is often a differentiation factor. In the triple circuit it helps stabilize the wrong configuration. That is what an attractor looks like. Identity is a joint state, not a list of oncogenes. A pioneer factor can bookmark a regressive trajectory when the other two nodes are holding the cell immature.
The experiment is therefore compatible with the upstream hypothesis, and easily misread as a replacement for it. Compatibility: the malignant appearance is a selectable state, reversible at the network surface without first editing the genome. Misreading: the network surface is the disease. The hypothesis predicts a controllable downstream lock. It also predicts that the lock was installed and retained for a bioenergetic reason. The paper tests the first prediction and is silent on the second.
Forced phenotype, defended set point
The durability problem is the same problem as drug-forced weight loss. A GLP-1 agonist occupies a receptor, lowers intake, and produces a lower weight while occupancy continues. Withdraw the drug and most of the weight returns, because the defended set point and the food environment were not rewritten. Occupancy forced a phenotype. It did not move the quantity the organism defends.
The KAIST design is occupancy. Knockdown is the drug. Enterocyte markers and slow growth are the forced phenotype. The untested question is whether the attractor moved or was only held. Nothing in the study releases the clamp. If the suppression was transient, the three genes return on a timescale of days, and no one looked. If it was stable, the clamp never came off, and the twenty-five-day xenograft is continued occupancy in a new location. Either way, durability was not measured. Reduced tumor volume under ongoing suppression is compliance. It is not memory.
The metabolic hypothesis makes a directional prediction here, and it is falsifiable. If the immature state is retained because glycolytic carbon-capture wins under ATP shortfall, then restoring MYB, HDAC2, and FOXA2 under glucose-replete, insulin-high conditions should recapture the cell. The mutations are still present. The mitochondria were not repaired. The selective landscape that favored the rescue is still there. The same restoration under conditions that lower glucose and insulin availability, and that support host and cellular oxidative capacity, should recapture more slowly or not at all, if engine function and selective pressure are doing the work the hypothesis assigns them. Recapture in the first condition and persistence in the second would favor a moved set point only when the bioenergetic boundary conditions change. Recapture in both would mean the network lock is self-sufficient and metabolism is irrelevant to its maintenance. Persistence in both would mean the knockdown rewrote identity regardless of fuel. The paper contains none of these arms.
This is also why the nude-mouse flank matters. Subcutaneous implantation in an immunodeficient animal removes immune editing, the colonic nutrient and microbial niche, and any competition with an intact epithelium. It is a permissive site for measuring whether silenced cells can assemble a mass. It is a poor site for asking what a metabolically stressed, immune-competent colon defends.
What the study cannot license
It cannot license the sentence that cancer cells changed their minds. A mind, in any sense that matters here, would be a stable preference that survives the removal of the force. The force was not removed.
It cannot license conversion as a clinical verb. There was no established tumor, no human, no drug that hits all three nodes with acceptable selectivity, and no evidence that normal intestine would tolerate the same suppression. HDAC2 is ubiquitous. MYB maintains normal stem-cell compartments. FOXA2 bookmarks endoderm far beyond the colon. A triple inhibition that is elegant in a Boolean model is a toxicity problem in a person. Broad HDAC inhibitors already exist. They are not HDAC2-specific, and they have not turned colorectal cancer into normal mucosa.
It cannot license the claim that computational control of differentiation trajectories has found the origin of cancer. BENEIN is a map of a lock inferred from normal differentiation. A good map is valuable. The Cancer Genome Atlas was also a good map. Large-scale cataloguing of drivers improved subsets of care and left the common solid tumors, once metastatic, largely uncured. Applying a better controller to the same layer — state description without a cause of state retention — risks the same ceiling. AI inside a genetic paradigm produced sharper targeting of downstream events. AI inside a differentiation-network paradigm can produce sharper targeting of a different set of downstream events. Neither asks why the configuration was selectable.
It also cannot be used against the metabolic hypothesis. Failure to measure mitochondria is not evidence that mitochondria are irrelevant. The study was not designed to test fuel, oxidative capacity, acetyl-CoA supply, NAD+ state, insulin, or intratumoral selection. Absence of those measurements is a boundary on interpretation, not a result.
What would make the result larger
The result becomes larger if the converted state survives restoration of the three genes inside a tumor that was established first, in an orthotopic and immunocompetent setting, and if that survival depends on the metabolic conditions the hypothesis specifies. Serial sampling across withdrawal would tell phenotype memory from continued occupancy. A comparison of glucose-replete and ketogenic or otherwise oxidative-supported conditions would tell a self-sufficient network lock from a lock maintained by carbon-capture selection. Cybrid or mitochondrial-transfer arms, of the sort that have already suppressed tumorigenicity in other models, would ask whether engine replacement moves the same attractor the triple knockdown forces. Those are the experiments that distinguish a lever from a cause.
Until then, the honest summary is narrow. Three regulators can hold a colorectal cancer cell line in an immature, tumorigenic state. Forcing all three down before implantation reduces that tumorigenicity for a few weeks in a nude mouse, by differentiation rather than by killing. The genome was not repaired. The mitochondria were not restored. The clamp was not released. The selective pressure that, on the mitochondrial-metabolic account, installs and retains the rescue program was not changed.
Oncology does not yet have a second verb. It has a clearer description of a lock, and a reminder that phenotype can be forced from the surface. Forcing is not the same as restoring the engine, and it is not the same as removing the reason the primitive configuration keeps winning. Therapy that only occupies the lock should be expected to behave like therapy that only occupies the receptor: effective while present, and silent on the set point the tissue defends when the occupancy ends.
A working hypothesis, open to correction and revised as new evidence arrives. Last revised October 17, 2026.
Related: Sage — Cancer (Oncology chapter) · Sage — GI · Sage — Warburg and Crabtree
See the notice at the top of this page. Working hypothesis, not medical advice.