AI found four hair compounds. The organoid graded them.

What the work claims

This is an industry method-and-discovery paper: a computational drug-discovery framework, applied to hair thinning, with in vitro validation. The authors argue that follicle decline is driven by several separable programs at once, and that a single compound acting on one pathway is therefore the wrong shape for the problem. Their framework combines a graph neural network trained on their own robotic screening data, which learns to predict whether a molecule boosts dermal papilla cell viability, with structure-based virtual screening against two defined targets: prolyl hydroxylase domain protein 2 (PHD2), whose inhibition stabilises the hypoxia regulator HIF-1 alpha, and the 5-alpha-reductases that convert testosterone to the follicle-shrinking androgen DHT.¹

The claimed output is four optimised compounds: two that increase dermal papilla proliferation, one PHD2 inhibitor, and one non-steroidal 5-alpha-reductase inhibitor with potency comparable to the drug dutasteride but without steroidal activity. The load-bearing efficiency claim is that computational prioritisation lifted experimental hit rates to between 7 and 21 percent depending on target, which the authors frame as more than a hundredfold above the roughly 0.1 to 1 percent typical of conventional high-throughput screening. The compounds were then combined into a single water-based formulation and tested in a follicle organoid.

How it works

The screening cascade is concrete. An ensemble of ten models was trained on robotic viability screens of primary dermal papilla cells, then used to score more than 11 million commercially available molecules; 22,506 passed the prediction threshold, structural and drug-likeness filters cut that to roughly 2,000, and 157 were tested, yielding 33 hits. The two target-based arms used physics-based binding calculations to rank candidates against PHD2 and the reductases. Each initial hit was then put through medicinal chemistry, mostly to fix poor water solubility, producing the four leads.

Target engagement was checked with orthogonal assays. The PHD2 inhibitor drove a hypoxia-response reporter and raised the downstream outputs VEGF-A in keratinocytes by about 19 percent and lactate in dermal papilla cells by about 10 percent, with RNA sequencing showing induction of hypoxia-adaptive genes. The 5-alpha-reductase inhibitor achieved near-complete enzyme inhibition at high dose, acted reversibly rather than by the irreversible mechanism of dutasteride, and, importantly, did not stimulate the androgen-sensitive prostate cancer line LNCaP, supporting the absence of androgenic agonist activity. Then came the tissue step. A follicle organoid, made by co-culturing primary dermal papilla cells with keratinocytes to recreate the epithelial-mesenchymal cross-talk of a real follicle, was used to measure sprouting length. The two proliferation compounds increased sprouting more than minoxidil; the reductase inhibitor reversed the growth suppression caused by added testosterone; and the four-compound formulation, stable for 150 days, outperformed a panel of commercial hair products on every in vitro axis the authors measured.

Where a skeptic should push

Start with the endpoints, because everything rests on them. The organoid readout is sprouting length, a morphological growth proxy, and the paper offers no evidence that sprouting length in a dermal-papilla-plus-keratinocyte aggregate predicts hair count or density in a person. That is the single most load-bearing assumption in the translational story, and it is unaddressed. The effect sizes on which the compounds are ranked are also modest: proliferation gains in the 11 to 23 percent range, a 19 percent VEGF bump, a 10 percent lactate rise, and organoid comparisons against minoxidil at p-values of 0.0243 and 0.0218 that are real but not commanding.

The reductase result carries a caveat the authors state plainly: the compound's benefit in the organoid is visible only when exogenous testosterone is added. That is not by itself a cheat, because a 5-alpha-reductase inhibitor blocks the conversion of testosterone to DHT and so cannot show benefit without androgen present, and androgen-driven miniaturisation is the genuine driver of the disease. The fair criticisms are narrower: the organoid does not develop miniaturisation on its own, the testosterone challenge is acute and may be supraphysiological, and unlike the established drugs dutasteride and finasteride this compound carries no clinical validation. So the experiment is a legitimate mechanism demonstration, not evidence that the compound reverses clinical hair loss. Separate the demonstrated from the asserted. Demonstrated: target engagement in reporter and enzyme assays, and morphological responses in a co-culture. Asserted: that any of this predicts regrowth in humans.

Two framing issues compound the endpoint problem. First, the hundredfold hit-rate improvement is benchmarked against generic literature figures for conventional screening, not against a matched control screen the authors ran themselves, and the hit thresholds, greater than 5 percent proliferation or 30 percent enzyme inhibition, are lenient. Second, this is a company validating its own product against retail competitors it selected, using assays it designed, with no in vivo data and no clinical data, and with RNA sequencing performed in flat culture where, the authors concede, off-pathway transcriptional responses linked to generic xenobiotic handling were prominent. None of that makes the chemistry wrong. It means the efficacy narrative is internally consistent and externally unproven.

The organoid as an AI screen's reality check

For organoid-based discovery, the constructive contribution is the shape of the workflow. Computational pipelines generate ranked lists cheaply and can overfit to whatever proxy they were trained on, here a single-cell-type viability score. The follicle organoid is used as the step that a two-dimensional viability assay cannot perform: it puts candidates into an epithelial-mesenchymal context where a compound that merely raises cell counts can still fail to produce tissue-level growth. Just as valuable, the organoid is where a four-compound polypharmacology cocktail was checked for compatibility and combined activity, something no single-target biochemical assay can do. That is a genuine and generalisable role for tissue models in the AI-discovery loop: not to originate hits, but to filter them against emergent, multicellular behaviour before anyone commits to in vivo work.

The non-obvious implication is a warning the paper embodies rather than states. An organoid lends a result the look of physiology, and that look is persuasive precisely when the underlying endpoint is weak. Here a growth surrogate with no validated tie to the clinical outcome sits at the end of an otherwise rigorous cascade, and its tissue-level appearance risks being read as the validation the cascade lacks. The discipline this argues for is specific: before an organoid endpoint is allowed to rank drug candidates, someone has to establish that the endpoint tracks the outcome that matters. Sprouting length, spheroid size, a morphology score, these are measurable, but measurable is not the same as predictive. The credibility an organoid confers should be earned by an anchoring study, not granted by the format.

The genuine threat is that AI plus organoids becomes an engine for manufacturing complete-looking evidence packages, mechanism annotation, target engagement, tissue morphology, formulation stability, that are entirely surrogate and self-benchmarked. For a consumer product this is a marketing risk; for a therapeutic program it is a capital-destroying one, because the polish of the pipeline can outrun the validation of the model and push weak candidates toward expensive clinical failure. The opportunity and the threat are the same instrument viewed from two sides: a follicle organoid is a good counter-screen for emergent tissue behaviour, and a bad substitute for the clinical anchor that would tell you whether its numbers mean anything.

The bottom line

Established here: a graph-neural-network and docking pipeline nominated four hair-follicle compounds across three pathways at experimental hit rates well above conventional screening; the compounds show target engagement in reporter and enzyme assays; and a dermal-papilla-plus-keratinocyte follicle organoid registered morphological growth responses and let the four be combined into a stable formulation. Still hypothesis, and central to the whole story: that any of these readouts predicts hair regrowth in people. The organoid endpoint is a growth surrogate of unproven clinical relevance, the reductase benefit is visible only against an artificially added androgen, the benchmarks are self-selected, and there is no in vivo or clinical evidence. What would confirm the approach is a study linking the organoid endpoint to a clinical outcome, followed by controlled human data. What would break it is evidence that sprouting length does not track regrowth. As a template for using tissue models to filter AI-generated hits, the workflow is sound; as proof that these particular molecules work, it is not yet evidence.