One donor, three platforms: how far does an Alzheimer's organoid proteome really travel?

What the work claims

The paper is a benchmarking study, not a discovery of new biology, and it should be weighed as one. Its argument is that patient induced pluripotent stem cell (iPSC) derived cerebral organoids, three-dimensional balls of human brain tissue grown from a patient's reprogrammed cells, are a translationally relevant model of Alzheimer's disease because their proteome matches what large human cohorts see.¹ To test this the authors grew organoids containing neurons, astrocytes and microglia from one Alzheimer's donor and one control donor, then measured proteins in both the organoid lysate and the conditioned media (the fluid the organoids secrete into) on the three platforms that now dominate clinical Alzheimer's proteomics: label-free mass spectrometry, the SomaScan aptamer assay, and the Olink antibody-based proximity extension assay.

They compared what the organoids contained against Alzheimer's biomarkers and drug targets nominated by 121 published proteomic studies, and against differentially abundant proteins from clinical cohorts totalling more than 17,000 samples: 13,623 plasma, 2,110 cerebrospinal fluid (CSF), and 1,275 dorsolateral prefrontal cortex donors. The headline results are two. First, the organoids detect almost every nominated candidate: every protein in the most heavily replicated sets (46 plasma proteins reported in more than three studies, 13 such CSF proteins, 17 such cortex proteins) was measurable in the matching organoid fraction. Second, and stated far more quietly, only roughly one in four of those most reproducible candidates were themselves differentially abundant in the organoid at all: about 22 percent of plasma candidates (10 of 46), 23 percent of CSF candidates (3 of 13) and 29 percent of cortex candidates (5 of 17). That is the lenient test, whether the protein changes at all in the model, not whether it changes in the same direction as patients. The stricter same-direction requirement was imposed only when the authors narrowed to 13 high-priority drug targets from the lysate (including CHCHD6, DLG4, NLGN3, NPTXR and NEK7), by demanding a concordant direction between organoid and cortex plus preferential brain expression.¹

How it works

The design is clever precisely because the three platforms disagree. The three platforms share only around 9 percent of CSF proteins and about 4 percent of plasma proteins, and they often diverge even where they overlap.¹ A candidate nominated by a mass spectrometry cortex study therefore cannot be checked in a model profiled only by an antibody panel. By running all three platforms on both the secretome and the lysate, the authors let each cohort be benchmarked on its own measurement technology. Coverage bears out why this matters: in the conditioned media only 252 proteins were shared across all three platforms, while in the lysate 2,090 were, because deep mass spectrometry of lysate recovered 9,971 proteins and approached the affinity panels.

To call a protein Alzheimer's-associated in the organoid the authors used differentially abundant proteins (DAPs) between the Alzheimer's and control organoids, tiered by how many platforms agreed. In the lysate, 59 DAPs were supported by all three platforms, 391 by two, and 2,419 by a single platform. Pairing the organoid secretome with patient biofluids and the lysate with cortical tissue, they then asked, protein by protein, whether the model detected the candidate and whether it moved the same way. Seven proteins recur across plasma, CSF and cortex nominations at once: APOE, BDNF, GFAP, HGF, HP, SMOC1 and SPP1. The regulatory backdrop the authors invoke is the FDA Modernization Act 2.0, which removed the mandatory-animal-testing requirement and named organoids among human models that can support evidence of drug efficacy and safety.¹ That is a governance fact about where the field is being pushed, not evidence that any given organoid is fit for the purpose.

Where a skeptic should push

The single most load-bearing assumption is that a change measured in one Alzheimer's organoid versus one control organoid is a valid estimate of the disease-associated change seen across thousands of patients. The authors are commendably explicit that they profiled organoids from one Alzheimer's and one control donor, both female, aged 84 (the Alzheimer's donor) and 72 (the control), and they argue, correctly, that the customary demand for three to five iPSC lines has no firm empirical basis and that truly capturing donor heterogeneity can need dozens of lines. But that candour does not rescue the direction claim. With a single genome on each side, the DAP statistics and their false-discovery control are computed across organoid replicates from the same two lines. They therefore measure technical and organoid-to-organoid variance, not the donor-to-donor and disease heterogeneity that the cohorts of 13,623 plasma donors are built to average over. A protein that moves up in this one Alzheimer's line could move up because of Alzheimer's, or because of that donor's genetic background, or the twelve-year age gap baked into the pair, or a differentiation quirk. With one genome on each side, disease and donor are perfectly confounded, and nothing in a one-versus-one design can separate them.

The second push is on what "most reproducible" means. The replicated sets were defined by how many studies reported a protein, which is a publication-frequency filter, not an effect-reliability filter. Proteins get reported often for reasons that include assay availability and prior attention, so the denominator against which one-in-four is computed is itself biased. And the concordance is measured against reference directions that the paper's own opening concedes are unstable: when SomaScan and Olink disagree on the sign of a change in real patients, "the organoid reproduces the disease change" has no single ground truth to reproduce. Finally, detection and reproduction are being narrated as if they were one achievement. Detecting almost every candidate is a statement about expression, that the protein is present and measurable. Showing a disease-associated change is the disease-relevant test, and even its lenient form, whether the protein changes at all, passes for about a quarter of the best candidates. The strongest case for the paper is the detection result plus the honest reporting of the change result; the weakest move a reader could make is to hear "detects almost every candidate" and infer near-perfect fidelity.

Why detection is not fidelity for drug targets

For organoid models of human organs and the drug discovery built on them, the useful reframing is that this study measured two very different properties and reported them under one banner. The genuine opportunity is real and specific: an organoid profiled on the same platform that nominated a candidate is a legitimate triage bench. If a plasma SomaScan cohort flags a protein, you can now ask whether a human brain model even expresses it and secretes it, on the identical assay, before committing a medicinal chemistry program. The completeness of detection is the valuable part, because a target the model cannot express is a target the model cannot mechanistically interrogate at all. The 13 direction-concordant, brain-preferential targets are a defensibly prioritized shortlist for follow-up, not a validated one.

The non-obvious implication cuts the other way. "Almost every candidate detected" is exactly the kind of number that will be quoted, under the FDA Modernization Act 2.0 banner, as evidence that cerebral organoids are ready to substitute for animals in nominating Alzheimer's targets. But detection is the cheap property. Any sufficiently deep proteomic assay of human brain tissue will detect most brain-expressed proteins; that is a statement about the assay's sensitivity, not the model's disease validity. The property a drug hunter actually needs, that the model shows the same disease-associated change the patients do, held for only about a quarter of even the best candidates on the lenient changes-at-all test, from a single donor, and full same-direction concordance was verified only for a filtered subset. The genuine threat is that a program in-licenses the concordant shortlist as though it were population-validated, builds an assay around a target whose organoid behaviour happens to reflect one patient's genome, and discovers the mismatch only in a failed trial. There is also a subtler risk from researcher degrees of freedom: because the reference platforms disagree with each other, comparator choice becomes a free parameter, and apparent concordance can be an artifact of which platform you benchmark against rather than a property of the biology. The discipline this paper should impose on the field is to report detection and change-reproduction as separate metrics, always, and to treat any single-donor result as a hypothesis awaiting a multi-line replication.

The bottom line

What is established here is narrow and worth having: patient cerebral organoids, profiled across mass spectrometry, SomaScan and Olink, detect essentially all of the most-replicated Alzheimer's proteins nominated by human cohorts, and the authors have built the field's first multi-platform reference proteome for prioritizing those candidates. What is only hypothesis is the translational headline: that these organoids reproduce human Alzheimer's biology well enough to nominate drug targets. A disease-associated change reproduced in only about one in four of the best candidates on the lenient changes-at-all test, and even that drew on one donor pair, with full same-direction concordance checked only for the 13-target subset, which is not enough to know whether one-in-four is a property of cerebral organoids or of these particular cells. The claim would be confirmed by a benchmark across many independent Alzheimer's and control lines showing that direction reproduction is stable and donor-independent, ideally with the concordant target shortlist holding up. It would be broken if adding lines collapses or flips the concordance. Until then, the right posture is to use these organoids as an expression and triage system, where they are strong, and to withhold the word "validated" from their disease-direction claims, where they are not yet earned.