Cellarity

The model is not an accelerant here, it is the method. Designing against a cell behaviour rather than a molecular target is only possible if something can represent that behaviour quantitatively, and the computational model of cell state is that representation. Take the modelling away and there is no way to define what the compound is supposed to do, because the objective itself is a learned description of a network state rather than a protein anyone could name in advance. Few companies in this lane depend on their models this completely.

The laboratory is the check, as it is throughout this category: a model proposes and an experiment overrules it. What is unusual here is that one ordinary human check has been removed by design. A medicinal chemist working from a target can ask whether a molecule does what it was built to do, and in a behaviour led programme there is no such question available at selection time, because the target is characterised afterwards.

That places more weight on the empirical result and less on human interpretation than elsewhere in the lane. Nothing published describes how model output is gated on the way to a development candidate decision, what a contradictory experimental result does to the model, or at what point mechanism must be understood before a programme advances.

The concept is explained clearly and the implementation is not. A reader can understand the approach: single cell technologies capture the molecular network of a cell, machine learning builds a quantitative representation of the behaviour that network produces, and compounds are sought that shift it. The company also publishes a visual artefact it calls a Cellarity Map showing modelled changes in cell behaviour under a drug candidate.

Below that level the description becomes proprietary framing rather than method. No model architecture, data volume, benchmark, validation result or peer reviewed methods publication was located, and network biology and cell state correction function as brand terms rather than as specified techniques.

The modelling appears to be built in house on the company's own single cell and multi omics data, and the platform is consistently described as proprietary, which is itself a partial disclosure. What is missing is any account of what sits underneath it.

No architecture, framework, external model provider, compute partner or third party dataset is named, and nothing distinguishes data generated internally from public reference atlases incorporated into training, which is the distinction that matters most for a platform of this kind.

Clinical stage with a first in human study running since June 2025, in healthy volunteers and people with sickle cell disease, and 28 day data on CLY-124 guided for the end of 2026. Alongside that sits genuine third party validation of a kind money cannot buy quickly: a partnered programme in metabolic dysfunction associated steatohepatitis with Novo Nordisk, meaning a large pharmaceutical company examined the platform and committed to it.

Guidance is also specific rather than vague, naming quarters for a preclinical readout and a development candidate nomination. It stays below the top grade because nothing has read out. The platform thesis, that behaviour led design translates to humans better than target led design, is precisely what the coming data tests, and until then it remains an argument.

Not applicable in the provider sense, since no clinical records flow into the platform. One distinction is worth drawing that does not arise for the chemistry led companies in this lane. Single cell and multi omics training data is derived from human tissue, so the corpus behind these models originates with donors and patients even though it never passes through an electronic health record. Nothing published describes consent, provenance or governance for that material. The representativeness question that follows from it is examined on the governance axis.

Not applicable in the provider sense. There are no covered entity customers and no protected health information arriving from health systems, because the company develops its own therapeutics. The same qualification recorded elsewhere in this lane applies: the not applicable convention was written for platform companies selling software, and a sponsor running a first in human study is handling identifiable subject data under the clinical research regime. Nothing published describes how that data is governed, and the grade follows the lane convention rather than penalising the company for a convention that has moved.

No certification, attestation or security page was located. As with others in this lane the exposure differs from a software vendor's, since the assets at risk are the company's own multi omics corpus and pipeline chemistry rather than customer data. Two things keep the absence from being weightless.

The company is running a clinical study and therefore holds human subject data, and it exchanges information with a large pharmaceutical partner, both of which normally attract contractual security requirements that nothing public describes.

Regulatory standing is evidenced by a decision rather than a description. CLY-124 is in a first in human study begun in June 2025 enrolling healthy volunteers and people with sickle cell disease, which means the investigational new drug pathway was cleared and the platform's output has been through regulatory review. A second programme has a development candidate nomination targeted for the fourth quarter of 2026.

It sits below the top grade because this is a single asset at the earliest clinical stage, and because nothing published addresses the regulatory question specific to this approach, which is how mechanism of action is established for a compound selected on phenotype before its target was known.

No governance position, model documentation or data composition statement was located, and for this company the representativeness question is unusually pointed. Models of cell state are trained on single cell and multi omics reference data, and the ancestral composition of publicly available reference datasets in this field is well known to be uneven. The lead clinical programme is in sickle cell disease, which overwhelmingly affects people of African ancestry.

Whether the reference data underlying a cell state model adequately represents the population the lead drug is intended for is therefore a direct scientific question rather than a general fairness concern, and nothing published addresses it. Recorded as the question a careful reader should ask rather than as a finding about the answer, which is not available.

Accountability for anyone actually exposed to the output runs through the clinical trial regime, which is a stronger apparatus than most of this index can point to: regulatory clearance before first dose, institutional review board oversight, informed consent and adverse event reporting.

That structure matters more than usual for this company, because a compound selected on phenotype carries an unresolved mechanism into the clinic and the trial system is what stands between that uncertainty and a participant. The grade stops below the top because this is the law rather than the company's own commitment, and nothing published describes internal accountability when a model driven selection proves wrong.

Not applicable. No electronic health record touchpoint, clinical workflow surface or provider integration exists, because the platform serves the company's own discovery programmes rather than external customers. The research analogue is the coupling between single cell assay data and the modelling layer, which is internal by design and involves no third party system.

Not applicable. The platform is operated internally and is not deployed, licensed or hosted for customers, so there is no tenancy, region or residency commitment for a buyer to assess. The partnered programme with Novo Nordisk implies some data exchange with a partner, and nothing published describes how that is structured, which is ordinary for a collaboration of this type.

Not applicable in the vendor sense, as the company sells nothing and has no price or customers. What stands in for it is disclosed with more discipline than most in this lane: a published pipeline page identifying the lead asset, its indication and its trial status, dated guidance tied to specific quarters, named therapeutic areas, a named pharmaceutical partner, and an identified investor syndicate. Publishing a timeline you can be held to is a form of commercial candour. Deal terms for the Novo Nordisk partnership are not disclosed, which is ordinary for this category.

Scope is defined and deliberately narrowed rather than left expansive. Hematology and immunology are named as the first therapeutic areas following what the company describes as proof of concept across diverse cell types, and the individual programmes are specified: sickle cell disease, myelofibrosis, metabolic dysfunction associated steatohepatitis and inflammatory bowel disease.

A platform company that publicly restricts itself to two therapeutic areas is making a harder claim than one that says its approach applies everywhere. It remains below the top grade because only one indication has reached the clinic.