There is a moment, now common in memory clinics, that the science has not quite caught up with. A person eighteen months into an anti-amyloid infusion schedule is handed a blood result. A tau marker is down. Down by how much, and down by enough to mean what? The honest answer from the clinic has been a shrug dressed up as caution. A paper published on 30 April 2026 in Frontiers in Pharmacology tries to put arithmetic behind the shrug, and it does so without enrolling a single person.
What Hugo Geerts, Shaina M. Short, Athena Grant and Piet H. van der Graaf built is a quantitative systems pharmacology model, which is to say a computer simulation of biology rather than an experiment on it. The authors work at Certara Predictive Technologies, a biosimulation company, with academic affiliations in Leiden and Cincinnati. Their amyloid half tracks the aggregation cascade explicitly: free Abeta42 monomers, oligomers (dimers up to 16-mers), protofibrils (17 to 23-mers) and plaques (24-mers and above), each with forward and backward reactions, dosed on the real titration schedules of seven antibodies including lecanemab, donanemab, aducanumab and gantenerumab. That was bolted onto a previously calibrated neuronal network model of a cortical microcolumn, with soluble amyloid and tau oligomers acting on specific voltage- and ligand-gated ion channels, an effect taken from preclinical work. The model's output was then checked against published numbers: 35 clinical data points on plasma p-tau181 change, and 34 reported changes in CDR-Sum of Boxes, across 32 drug-dose-time combinations. These are group averages from trial reports, not individual patient records.
The fit tells you where the action is. Simulated change in protofibrils tracked observed p-tau181 change best (r squared 0.62), followed by oligomers (0.56), with a multivariate fit reaching 0.69; overall the model accounted for about 70 percent of the variance in observed p-tau181 change and about 50 percent of the variance in CDR-SOB change. In other words, within this model it is the soluble, hard-to-measure intermediate species, not the plaque bulk that lights up a PET scan, that best explain why blood tau moves.
That sets up the finding that matters at the bedside. The team asked what plasma p-tau181 would read at the precise moment central amyloid load crossed below 25 Centiloids, the usual amyloid negativity threshold, in 1,000 simulated patients. The predicted drop was not one number. It was about 15 percent for donanemab, about 45 percent for aducanumab and about 75 percent for lecanemab. The logic is almost mundane once stated: an antibody that strips amyloid quickly reaches the threshold sooner, and tau simply has had less time to respond. If that prediction holds, a single universal blood-tau cut-off for declaring the amyloid job done would be wrong in opposite directions depending on which drug a person was given.
The second prediction is about why clearing amyloid helps some people and barely touches others. In the model, lowering soluble oligomers eases their effect on ion channels and so lowers neuronal firing, which the authors frame as counteracting the hyperactivity seen in mild cognitive impairment and early Alzheimer's, and which simultaneously pulls down p-tau181. Push the simulated patient further along the disease, with more neuronal pathology and a higher baseline tau load, and the cognitive benefit shrinks regardless of how much amyloid is removed. The authors' proposed reasons are loss of the compensatory upregulation of the enzyme ChAT that raises cholinergic tone, and loss of synapses and neurons as the illness advances.
Hold all of this at arm's length, because the authors largely do. The 1,000 virtual patients had an average age of 74.4 years (range 61 to 90), an average baseline amyloid load of 120 Centiloids and about 75 percent were APOE4 carriers, and none of them exist. The cognitive engine was calibrated on ADAS-Cog while the antibody trials used CDR-SOB, a gap the authors bridge by citing close to 300 parallel readings of the two scales. The calibration used p-tau181, while much of current practice has moved toward p-tau217, and the authors concede baseline levels differ by epitope even if they expect relative change to behave similarly. And the whole structure rests on an assumption they name plainly: that human cognition can be simulated by the spike train activity of a cortical microcolumn.