Using brain organoids to reverse-engineer clozapine

What the work claims

This is a primary mechanism-discovery study built on a substantial resource: three-dimensional human cortical organoids (hCOs) derived from 28 induced pluripotent stem cell lines, 14 from schizophrenia (SCZ) patients and 14 from matched controls, of whom 25 of 28 were of European ancestry and only one had confirmed treatment-resistant disease.¹ Clozapine (CLZ) is the only drug approved for treatment-resistant schizophrenia (TRS), yet decades on, no one knows precisely how it works downstream of its messy receptor binding. The authors set out to find out by exposing the organoids to clozapine for four to 24 weeks and reading the molecular consequences across several data layers at once.

Their central claim is that schizophrenia organoids carry a distinctive alternative-splicing signature, that this signature is enriched in glutamatergic (excitatory) neurons, that clozapine reverses it, and that the reversal is independent of how fast the organoids metabolize the drug. Alternative splicing, the process by which one gene's exons are assembled into different mature messenger RNAs, is thereby nominated as a druggable node for next-generation antipsychotics. The bold part is not the schizophrenia biology, which remains associative, but the proposal that an organoid panel can serve as a mechanism decoder for a drug whose target was never understood.

How the splicing program was found

The organizing move is layering. At day 180, after four weeks of clozapine, the authors ran deep bulk RNA sequencing and mass-spectrometry proteomics on the same organoids, quantifying 19,416 protein-coding transcripts and 6,849 proteins, with 6,472 gene products overlapping.¹ A first, sobering result set the tone: messenger RNA and protein levels correlated only weakly, a median Pearson coefficient of 0.19 across the overlap, a reminder that transcript abundance is a lossy proxy for the protein that does the work. Comparing patients with controls, the strongest disease signal was an upregulation of genes involved in RNA splicing, while synaptic genes were downregulated; clozapine treatment pushed in the opposite direction, most clearly by lowering splicing-related proteins. That inverse pattern, disease raises the splicing program and the drug lowers it, was only visible when proteomics was added to transcriptomics.

To resolve the splicing itself, the authors quantified alternative-splicing events using percent-spliced-in, the fraction of transcripts that include a given segment. Of 94,499 confident events, 71 percent were intron retention and 14.2 percent were exon skipping, and it was exon skipping that was over-represented among the disease-changed events far beyond chance.¹ The change-in-inclusion values for shared exon-skipping events at baseline and after treatment were negatively correlated, meaning clozapine tended to undo the schizophrenia-associated skip, with the histone-demethylase gene KDM2B highlighted as the top reversed event. A connectivity-mapping analysis, borrowed from drug repurposing, formalized this as a statistically significant reversal of the exon-skipping, intron-retention and alternative-splice-site signatures.

Single-cell sequencing of 16,274 cells from older, 350-day organoids then localized the program. Glutamatergic neurons, the largest population at 48.7 percent, carried the bulk of the disease-associated intron retention, and clozapine inverted splicing signatures across astrocytes, inhibitory neurons, glutamatergic neurons and progenitors, with the excitatory neurons additionally showing reversed exon skipping.¹ Two orthogonal anchors gave the signature external traction: projected onto a population-scale atlas of adult postmortem prefrontal cortex, the organoids' disease signature correlated most strongly with real schizophrenia, and the treatment signature anti-correlated with it; and a splicing-focused association analysis tied 1,044 exons in 369 genes to schizophrenia genetic risk, overlapping the organoid protein-level signatures. The drug metabolism control matters to the story: the organoids formed only trace clozapine metabolites, 0.03 to 0.52 percent of the parent compound, with no case-control difference, which the authors take to mean the splicing effect is not an artifact of differential drug breakdown.

Where a skeptic should push

The most load-bearing assumption is that a signature-level reversal in these organoids reflects clozapine's therapeutic mechanism in adult treatment-resistant schizophrenia. Several gaps sit under that. The organoids are developmentally young: the authors' own atlas comparison matches them to second-trimester cortex, a prenatal-like state, while the disease presents in adulthood and the drug is given to adults. The cohort is schizophrenia in general, not treatment-resistant schizophrenia; only one donor had confirmed TRS, yet the therapeutic framing is entirely about TRS, and whether TRS is simply severe schizophrenia or a distinct entity is unsettled. And the exposure used 10 micromolar clozapine, which the methods themselves call supraphysiological and which sits several-fold above typical clinical plasma levels, so the magnitude of the effect may not scale to the clinic.

The evidence is also associative at the point where it most wants to be causal. Nothing in the study manipulates the splicing program itself: there is no experiment knocking down a splicing factor to see the schizophrenia phenotype improve, no functional or electrophysiological readout of rescue, and no behavioral endpoint, because an organoid has none. The connectivity-mapping reversal is a statistical signature match, not a demonstration that fixing the splice fixes anything. Two of the load-bearing signals lean on the weakest data layers: the disease-versus-drug inverse pattern was clearest in proteomics, which here detected mostly highly expressed proteins and correlated with mRNA at a median of only 0.19, and the single-cell splicing analysis could not be done on real single cells at all. Because splicing needs junction reads that sparse single-cell data lack, the authors pooled cells into 96 synthetic pseudobulk samples, which recovers coverage but forfeits genuine single-cell resolution and can inflate apparent significance. Finally, the metabolism-independence claim is softer than it sounds: the organoids barely metabolized clozapine in the first place, so concluding the effect is metabolism-independent is close to observing that little metabolism occurred.

To the authors' credit, they flag most of this, including the prenatal state, the single TRS donor and the open question of whether other antipsychotics share the effect. The honest separation is between what is shown, that schizophrenia organoids carry a splicing signature that correlates with real disease and genetic risk and that clozapine statistically reverses, and what is proposed, that this splicing program is the drug's mechanism and a target for new drugs.

Decoding a drug's mechanism, and the blind spot

For organoid models of human organs and the drug discovery built on them, the interesting contribution is not a schizophrenia target but a workflow. Most organoid drug work is framed as efficacy prediction: put a compound on a patient's cells and see if it kills or corrects. This study runs the arrow backwards, taking a drug whose target was unknown and using a genetically stratified organoid panel plus layered omics to localize a candidate site of action to a cell type, glutamatergic neurons, and a molecular process, exon skipping. That is mechanism-of-action hypothesis generation rather than a decoded mechanism, because nothing here perturbs the splicing program to prove it is causal, but it unlocks a capability the field undersells: an organoid biobank as an instrument for nominating where an existing drug may act and generating a fresh target hypothesis from it, rather than only ranking candidates. The proteomics lesson is part of the blueprint. Because the decisive disease-versus-drug inversion was invisible in transcriptomics and only appeared at the protein level, any organoid screen that reads out messenger RNA as a stand-in for function inherits that median 0.19 decoupling and can miss the real signal.

The non-obvious threat is a dual-use safety trap, and it is grounded in the paper's own words. The authors report, as a strength, that prolonged clozapine did not harm cell viability or induce stress, and describe the exposure as non-toxic. In a brain organoid that is true and beside the point. Clozapine's clinically limiting toxicities, the reasons it is a last-line drug requiring blood monitoring, are agranulocytosis and myocarditis: they are hematologic and cardiac, they live in organs a cortical organoid does not contain, and they are therefore structurally invisible here. An organoid safety readout is organ-scoped by construction, so a compound can be pronounced non-toxic in the exact model that cannot host the toxicity that matters. The authors' own statement is scoped and fair: clozapine did not harm their cortical organoid at the doses used. The hazard is downstream of the paper, when an organ-scoped result is read or cited as a general safety signal it was never meant to be, because the model has no marrow or myocardium in which the drug's defining toxicity could appear. Read as mechanism discovery, this study is a strong template; the caution attaches to how such a finding travels, not to the claim the authors actually made.

The genuine opportunity and the genuine threat thus share a mechanism. The organoid's power is that it isolates one organ's molecular response with enough depth to decode a drug; its danger is that the same isolation hides every liability that lives elsewhere in the body. A drug-discovery program that uses organoids to nominate targets and decode mechanism is on firm ground here. One that reads an organoid's silence on toxicity as evidence of safety is being misled by the model's boundaries, not informed by its biology.

The bottom line

Established result: patient-derived cortical organoids carry an alternative-splicing signature, concentrated in glutamatergic neurons, that correlates with real schizophrenia transcriptomes and with splicing-linked genetic risk, and clozapine statistically reverses it without detectable cytotoxicity in the model. That is a real, multiomics-anchored observation. Still hypothesis: that this splicing program is clozapine's therapeutic mechanism, or a viable drug target for treatment-resistant schizophrenia. The organoids are prenatal-like, the cohort is not treatment-resistant, the drug dose is supraphysiological, the causal link is untested, and the key single-cell splicing result rests on synthetic pooling. What would confirm the claim is a causal experiment, correcting the splicing program by an independent means and recovering a disease-relevant phenotype, ideally in more mature organoids and in confirmed treatment-resistant donors. What would break it is evidence that other antipsychotics reverse the same signature without clozapine's clinical superiority, which would sever splicing reversal from efficacy. Either way, the transferable lesson for the field is the method and its limit: organoids can decode what a drug does to one organ, and cannot tell you what it does to the rest of the body.