Cellarity

The models are the method. Cellarity designs drugs against the behavior of a whole cell rather than against a named protein, and the only way to work that way is computational: high dimensional single cell transcriptomics maps the network state that defines how a cell behaves, and learned models then link chemistry to that state to predict which compound moves a diseased cell back toward a healthy one.

There is no version of this company without the modeling. The lead asset makes the point concretely. CLY-124 was reached by transcriptomic analysis of fetal and adult erythropoiesis that pointed to a non cytotoxic route to fetal hemoglobin induction, and only afterward was the target deconvoluted to DCN1 and the mechanism to partial suppression of CUL3 neddylation, a target the company notes had no inhibitors in clinical development. Remove the models and the program does not exist, because nobody was looking at that target.

Oversight here is a described mechanism, not an asserted phrase. The platform runs what the company calls a lab in the loop method: the model predicts, an experiment tests the prediction, and the outcome refines the next round of predictions, published as an active learning framework in Science 2025. The experiment overrules the model by construction, and the published account says the loop is what improves both the efficacy and the safety profile of selected compounds.

The safety model carries published thresholds. Nature Communications 2025 states a probability cutoff of 0.7 and a margin of safety cutoff of 80, says both were optimized on held out training data to reach performance plateaus while minimizing false positives, and explains why the margin sits above the usual literature range of 10 to 100. That gives a stated point at which a compound is classified high risk rather than low.

The missing part is the human one. Nothing published describes which decisions require a person rather than allowing one, how a model selected compound is gated on the way to a development candidate nomination, or what the company does when the loop produces a candidate whose mechanism cannot be established. That last question is specific to this platform, because a compound here can be chosen from its effect on a cell state before anyone knows what it binds.

The marketing pages describe the approach in general terms, naming single cell foundation models and dynamic AI modeling without identifying anything. The publications are where the substance is, and they name a great deal.

ToxPredictor is identified in Nature Communications 2025 as a random forest classifier chosen from 193 configurations across eight model classes and deployed as an ensemble of 30 models trained on different cross validation splits, with the reasons for choosing it over boosting and deep learning approaches stated. Its training data is characterized precisely: DILImap, full transcriptome RNA sequencing of 300 compounds at multiple concentrations in sandwich cultured primary human hepatocytes, with compound categorization drawn from the named public resources DILIrank and LiverTox. The discovery framework is published in Science 2025 as an active, lab in the loop deep learning system powered by high throughput transcriptomics, released with three open single cell datasets including a perturbational set of more than 1,700 samples and 1.26 million single cells.

The gap is version and update discipline. No version identifier, release history or update practice is published for any model, so nothing tells a reader whether the system running today is the system the 2025 papers describe. The single cell foundation models named on the platform page are not identified anywhere. The Science paper's method is summarized from the company's own account of it.

The chain is structurally short. The models are built in house and published as such: a random forest ensemble for the safety model, a deep learning active learning framework for discovery, both described in the company's own papers. No third party model provider sits between the data and the output, and there is no supplier model here that could be swapped without the change being visible in the literature.

The outside parties that do appear are named in the method sections rather than in a disclosure document. Compound categorization comes from DILIrank, maintained by the Food and Drug Administration, and from LiverTox; the earlier proof of concept used the public TG-GATES microarray database; compounds were purchased from MedChemExpress; sample handling used a named liquid handling instrument. Those establish where the inputs come from.

What is not published is a party list a buyer or partner could rely on. There is no subprocessor list, no named hosting or compute provider, and no terms governing data once it reaches any of them. For a company running partnered discovery programs with Novo Nordisk and Pioneering Medicines, that is the disclosure gap that matters most here.

Both halves a drug discovery platform needs are present: a candidate in a clinical study with a disclosed timeline, and peer reviewed method papers behind the platform that produced it.

CLY-124, a first in class oral globin switching candidate for sickle cell disease, entered a global Phase 1 proof of concept study in June 2025 in healthy volunteers and people with the disease. The timeline is published rather than implied: 28 day data in the fourth quarter of 2026, multiple ascending dose cohorts in the United States, Kenya and other countries, those cohorts run as monotherapy and in combination with hydroxyurea, and a myelofibrosis development candidate nomination in the fourth quarter of 2026 with head to head in vivo data against a JAK2 inhibitor in the second quarter.

The method is published in two peer reviewed journals. Science 2025 describes the active, lab in the loop deep learning framework behind the platform and reports a 13 to 17 fold improvement in recovery of phenotypically active compounds over industry standard approaches, released alongside three open single cell datasets for outside benchmarking. Nature Communications 2025 publishes ToxPredictor and the DILImap library of 300 compounds profiled at multiple concentrations in primary human hepatocytes, with 88 percent sensitivity at 100 percent specificity in blind validation against more than 20 preclinical safety models, and names three Phase 3 failures it flagged that animal studies missed. Preclinical mechanism for the lead asset was presented in an oral session at the 67th American Society of Hematology Annual Meeting in December 2025, identifying the target as DCN1 and the mechanism as partial suppression of CUL3 neddylation, with in vitro dose response to 25 percent fetal hemoglobin at an EC50 of 17.5 nM, in vivo humanized mouse data against hydroxyurea, and GLP toxicology showing no cytopenias, organ dysfunction or genotoxicity.

No clinical readout has been published yet, the 28 day data is still guided rather than reported, and no trial registration identifier appears anywhere on the company's estate.

The only privacy document published is a website notice last revised on 10 December 2019, scoped in its own opening paragraph to access and use of the Cellarity website. It does not reach the clinical trial route the company operates, where a person with sickle cell disease is invited to contact the company and a member of the clinical team responds, and it takes no position on health information at all.

No retention window is stated, only that information is kept for as long as needed; nothing says what data reaches a model or what happens to it there; and no de-identification practice is described for anything beyond a clause permitting the company to create anonymized and aggregated datasets from website information for research and other functions. The document's security section offers general assurances and a warning that information sent electronically may not be secure.

What is clear is that the platform's inputs are transcriptomic and chemical rather than clinical records: DILImap profiles compounds in primary human hepatocytes, and the released single cell datasets are laboratory data. Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is not published.

Nowhere does Cellarity state a business associate status, and the only privacy document reaches the website alone. The notice was last revised on 10 December 2019, describes itself as covering access to and use of the Cellarity website, and mentions neither protected health information nor the company's position under the United States privacy rule.

There is no plain statement that the product does not process protected health information, no explanation of scope, and no service level privacy document behind one. Whether it handles such data is a live question: the company sponsors a Phase 1 study with sites in the United States and Kenya, publishes a page inviting people with sickle cell disease to contact its clinical team, and is therefore handling identifiable subject data under the clinical research regime. Nothing published describes how that data is governed, and the notice covering the website is more than six years old.

No certification, attestation, trust center or security page is published. There is no SOC 2 report, ISO/IEC 27001 certificate or HITRUST certification, and no independent penetration testing cadence. The only security material is a section of the website privacy notice last revised on 10 December 2019, which tells a reader that information sent electronically may not be secure, that the company reviews its security procedures from time to time, and that no security measures are perfect. Those are controls asserted with nothing independent behind them.

The exposure that leaves unaddressed is not hypothetical. Cellarity holds partner data under discovery collaborations with Novo Nordisk and Pioneering Medicines, and as sponsor of a Phase 1 study with sites in the United States and Kenya it holds identifiable human subject data. Nothing published describes the controls around either.

The platform makes no device claim and is scoped accordingly. Cellarity is a clinical stage therapeutics developer: the discovery platform is not sold as a product, is not presented as a medical device, and carries no clearance. That is the appropriate position for what it is.

The regulatory positions the company does hold belong to the medicine. CLY-124 entered a global Phase 1 proof of concept study in June 2025 in healthy volunteers and people with sickle cell disease, and GLP toxicology results were presented at the 67th American Society of Hematology Annual Meeting in December 2025. No trial registration identifier is published, even on the page written for prospective participants, so the protocol has to be found elsewhere. A second regulatory thread runs through the safety work rather than the pipeline: the DILImap resource is framed against regulators' stated aim of reducing reliance on animal testing, which is a positioning claim rather than a regulatory status.

Where the models underperform is out in the open, because the company published the evaluation, the breakdown and the data.

The safety model, ToxPredictor, is described in Nature Communications 2025 (doi 10.1038/s41467-025-65690-3) with its method stated: a random forest classifier selected from 193 tested configurations across eight model classes, run as an ensemble of 30 models, with all doses and replicates of a compound held out together in each fold so the test mimics generalization to unseen compounds. In blind validation it reached 88 percent sensitivity, 29 of 33 DILI positives, at 100 percent specificity, 14 of 14 negatives, with a ROC AUC of 0.96 against 0.65 for a cell viability readout alone. Performance is reported across the risk strata the model serves rather than only in aggregate: no DILI, likely DILI, known DILI and withdrawn due to DILI, each with its count, in blind validation and again across the full library (52, 73, 25 and 13). The four misses are visible, as is the weaker earlier proof of concept on a public microarray database at 62 percent sensitivity and 92 percent specificity. Decision thresholds are published and justified, a probability cutoff of 0.7 and a margin of safety cutoff of 80, with the reason the latter sits above the usual literature range.

The model and its validation data are released publicly at dilimap.org, and the discovery framework paper in Science 2025 released three single cell datasets for outside benchmarking, including a perturbational set of more than 1,700 samples and 1.26 million single cells used to benchmark the platform in that paper. Both papers went through peer review. What is not published is a fairness evaluation in the demographic sense. The Science paper's figures are as the company reports them.

Cellarity publishes an error rate with its method and its denominator. ToxPredictor's performance is stated in Nature Communications 2025 as 88 percent sensitivity, 29 of 33 DILI positive compounds detected, at 100 percent specificity, 14 of 14 negatives, in blind validation, with a ROC AUC of 0.96 against 0.65 for cell viability alone, and 0.82 against 0.66 in cross validation. The comparator set is named as more than 20 preclinical models, the held out design is described, and the model and its validation data are published at dilimap.org so the figure can be checked rather than taken on trust. Four missed positives are stated rather than rounded away.

What is absent is any commitment running toward anyone. Nothing published states what Cellarity is answerable for if a model selected compound proves unsafe, no guarantee or remediation obligation attaches to the platform's output, and the recourse available to a trial participant comes from the clinical trial regime that binds any sponsor rather than from anything this company has undertaken. The interpretability gap the platform's own approach creates, that a compound can be selected before its mechanism is known, is now partly answered for the lead asset: the target and pathway were deconvoluted and presented at the 67th American Society of Hematology Annual Meeting in December 2025. Nothing published describes how that is done systematically, or what happens when it cannot be.

No clinical record integration exists. There is no electronic health record touchpoint, no clinical workflow surface and no provider connection, because the platform serves Cellarity's own programs and its partners' programs rather than health systems.

The research side is just as bare. No electronic laboratory notebook, laboratory information management system, research data platform or partner cloud environment is named as connected, and no interoperability standard or integration is claimed on the website or in the published methods, including for the partnered programs with Novo Nordisk and Pioneering Medicines where data moves between two organizations. The named external resources that do appear, DILIrank, LiverTox and TG-GATES, are public reference databases the company drew on for a published study, not connections a buyer could use.

The platform is operated internally rather than deployed to customers, so there is no tenancy model, hosting option or residency commitment for a buyer to evaluate. Location is implied rather than stated: the company's laboratories and single address are in Somerville, Massachusetts, and the published methods describe the work being done in its own facility.

Nothing published names a cloud or compute provider, states where data rests, or describes how processing crosses a border. Two places where that would matter are published without it. The discovery collaborations with Novo Nordisk on metabolic dysfunction associated steatohepatitis and with Pioneering Medicines on inflammatory bowel disease move partner data and results between organizations, and the Phase 1 study runs dose cohorts in the United States, Kenya and other countries. No statement covers either.

Nothing is sold, so there is no price to publish and no cost for a buyer to size. What is visible is the shape of the business: a Flagship Pioneering company monetizing through its own pipeline and through partnered discovery, with a $123 million Series B raised in 2021 and an investor base named as Flagship Pioneering, Baupost Group, BlackRock, Pictet, Hanwha Impact and Kyowa Kirin.

The partnering side is named but not quantified. The metabolic dysfunction associated steatohepatitis program runs with Novo Nordisk and the inflammatory bowel disease program with Pioneering Medicines, and no upfront payment, milestone structure, royalty arrangement or term is disclosed for either, so the economic weight of the partnerships cannot be judged from published material. No cumulative funding total more recent than the 2021 Series B is published, not even in a corporate outlook published in January 2026 that otherwise sets out milestones in detail.

Coverage is named narrowly and deliberately, and part of it has validation behind it. Hematology and immunology are stated as the first two therapeutic areas, chosen after what the company describes as proving the platform across diverse cell types and cell behaviors. Four programs are named with their stage: sickle cell disease in Phase 1, myelofibrosis in preclinical work with a development candidate nomination targeted for the fourth quarter of 2026, metabolic dysfunction associated steatohepatitis partnered with Novo Nordisk, and inflammatory bowel disease in collaboration with Pioneering Medicines. Naming four areas and refusing the everything platform posture is a disciplined choice.

Validation reaches one of them. CLY-124 is in human study with published preclinical mechanism, in vivo efficacy in humanized mice and GLP toxicology presented at a hematology meeting; the myelofibrosis program has a published preclinical approach and no candidate yet; the two partnered programs have no published data at all. The platform's stated reach across diverse cell types is a company assertion rather than something the published work demonstrates by cell type.