The Shuffle Inside the Cell: A Computer Model Asks Whether Mitochondria Mix Themselves Into Dementia

Inside every neuron you own, something is quietly coming apart and putting itself back together. Mitochondria, the organelles that make the cell's energy, do not sit still: they fission into fragments and fuse back into larger structures, over and over, for as long as the cell lives. Textbook wisdom treats this as good housekeeping, a way of sharing out working parts so that no single compartment is left short. Alan G. Holt and Adrian M. Davies, both listed as independent researchers, wanted to know what that same habit looks like across a human lifetime when some of the parts being shared are broken. So they built the neuron in software and let it run. Their paper appeared in Clinical Bioenergetics on 13 July 2026, having been submitted in April and accepted in June; an earlier version was posted as a bioRxiv preprint in December 2024.

The study is a simulation, and nothing else: no patients, no animals, no cultured cells. What the authors wrote, in Python 3, is a single post-mitotic neuron containing a population of mitochondrial DNA molecules that can replicate, expire, mutate and compete, with no cell division to reset the deck. They then gave the organelle two possible geometries. In the closed cristae model, mtDNA is penned inside an individual crista, one of the folded internal compartments, and can only move elsewhere when fission and fusion carry it there. In the open mitochondrion model, mtDNA drifts freely throughout the organelle. Each configuration was run over a simulated human lifespan of 100 years, with 200 replications for every combination of factor and level, and the resulting neuron loss was read against thresholds the authors took from the published literature: below 20 percent cumulative loss counts as healthy ageing, 20 percent marks the onset of dementia, and 40 percent is severe. The source code sits in a public repository, which is more than most modelling papers offer.

The headline result is unglamorous and fairly stark. In the closed cristae model, the faster the mixing, the worse the outcome: neuron loss rose with the fission and fusion rate, the onset of decline moved earlier, and severity increased, although the losses appeared to reach a ceiling at higher mixing rates rather than climbing forever. Turn the mixing rate down towards nothing and, in the model, neuron loss becomes almost non-existent. The logic the simulation encodes is simple enough to follow: a deletion mutant confined to one crista is a local problem, but a deletion mutant redistributed around the organelle gets the chance to clonally expand, and clonal expansion is the prerequisite for the simulated cell's death. It is worth being precise about what this is. It is a hypothesis rendered in code, calibrated against human neuron-loss figures, not a measurement taken from a brain.

Then the authors did the thing that makes the paper worth reading. They added selective mitophagy, the quality-control step in which unhealthy fragments are not allowed to fuse back and are instead expunged from the organelle. In silico, it worked far too well. Neuron loss was essentially eliminated, even when mitophagy events were made rare relative to mixing, and even when the simulated host was given a very high predisposition to deletion mutations. A model in which nobody loses neurons is a model that cannot account for the amount of dementia that actually exists in human populations. In the single run under selective mitophagy that did produce neuron loss, the authors report that background heteroplasmy accounted for 60 percent of the mtDNA population and the clonal expansion for only 37 percent, with 3 percent wild-type, a pattern that does not match what has been observed in real tissue, where background heteroplasmy is low and a few clonal expansions dominate. Their conclusion is a piece of honest self-criticism: given how well selective mitophagy performs in their simulation, they are skeptical that it operates in living neurons the way their function models it.

For a family living with a diagnosis, the useful translation is narrow, and it is not a treatment. Nothing here identifies a drug, a supplement, a dose or a target; nothing here says that any particular person's dementia was caused by mitochondrial mixing. The authors do float one speculation, that compounds which stress mitochondria into excessive fragmentation could matter, and they mention pharmaceuticals and pesticides in that context, but that is an idea generated inside a simulation to explain a simulated pattern, not an observation in people, and it should be read as exactly that. What the paper contributes to the grammar of living systems is a grammatical point: it treats organelle architecture as a variable rather than a backdrop, and it shows that whether mtDNA is fenced into cristae or free to roam changes the shape of the neuron-loss curve across a lifetime. That is a claim experimentalists can go and test.

There is a quieter value here too. Published models usually arrive tidy, with the mechanism confirmed and the curve fitted. This one reports its own mismatch with reality and asks the field to explain it, which is how modelling is supposed to work: the failure is the finding, and somebody with a microscope now has a reason to go looking.