Latent Labs

The models are the whole company. Latent-X1, Latent-X2 and the Latent-Y agent take a target and a binding site and generate an antibody or a protein binder, sequence and bound structure together, and that generated design is what a partner pays for. There is no immunization campaign, display library or physical screening deck underneath that would survive if the model were removed.

The published validation is zero shot, against targets the company states had nothing like them in the training data, which only sharpens the point: the design comes out of the model first and the wet lab confirms it afterward.

The Latent-Y agent is built to run an entire design campaign on its own, from reading the literature and picking an epitope through generating candidates and selecting lab ready sequences, and the company reports that its validated results were produced this way, without a human filtering step before the lab. It also offers a collaborative mode in which a researcher reviews progress and directs each subsequent step, so the oversight structure and its two settings are described.

The part that is missing is what governs the autonomous run. Nothing published states the confidence level at which the agent stops and asks for a person, what it does when a lab result contradicts its own prediction, or which decisions are reserved for a human. The control is visible in how the system is built and undocumented as a rule a partner could hold the company to.

The models are named, versioned and tied to dated technical reports. Latent-X1, Latent-X2 and Latent-Y form a lineage with public release dates. Each is an all atom generative system that produces the sequence and the bound structure of a protein complex at once, from a target structure, an epitope specification and, for antibodies, an optional framework to build onto. Latent-X2 adds antibody formats, nanobodies and scFvs, to the macrocycles and mini binders of Latent-X1, and Latent-Y wraps the generator in an agent that plans and runs a campaign.

Under that, the detail thins out: no model class, parameter count or training procedure, and the training data characterized only as large protein sequence and structure data, with no size or cutoff. The dated reports mark the versions, but none ties a change in the production model to a date.

The chain is short by design, and the company describes it that way. The models are built in house, the Latent-Y agent generates its candidates through Latent Labs' own Latent-X2 rather than a third party model, and no outside model provider sits between a target and a design. There is nothing that could be swapped out without the change being visible.

The part a partner's review would still chase is the infrastructure around that: the compute and hosting behind the platform, any third party components, and the terms that govern a partner's target and design data once it lands, since several partners' programs run on the same system. None of that is named.

There is a lot of published validation, none of it yet independent. Each of the three models ships with a technical report carrying lab results: per target hit rates, scFv affinities into the low picomolar range, a ten donor immunogenicity panel, and a bench experiment against prior design methods. Named academic groups at UC Davis, LMU University Hospital and the Translational Genomics Research Institute are reported as early users, and the validated sequences for each target are posted on the platform for a reader to inspect.

Two things cap it. None of the reports has been through peer review; they are company authored documents on arXiv. And no molecule designed on the platform has entered a clinical study, so every result is biochemical or from an early lab assay rather than from use in a patient.

The commitment about platform data is categorical: the company states it does not train on user data or outputs, and that researchers keep the sequences they generate. For a platform whose users upload the targets and epitopes they care about, that is the stewardship question that matters, answered plainly rather than left to inference.

What sits behind it is thinner. There is no retention schedule for how long a user's inputs and designs are held, and the company reserves the right to learn from aggregate anonymized usage. No protected health information is involved, since the inputs are protein targets and sequences. And safety engineering, meaning guardrails and how a safety event is handled, is not described.

The data protection frame here is European rather than United States health. The published commitments, that the company does not train on user data or outputs and that researchers keep their sequences, reach the platform service a customer actually uses, and the service is described as GDPR compliant. No business associate status is stated anywhere, and none would be expected, since the inputs are protein targets and sequences rather than patient records. There is simply no HIPAA position taken, in either direction, because the United States health privacy rule is not the framework this product is sold under.

A pharmaceutical partner's security team, before putting target data on a shared platform, looks for an independent attestation it can open: a SOC 2 report, an ISO 27001 certificate, a trust portal. Latent Labs offers none of these. What it offers is a statement that platform data is kept confidential and meets GDPR, which speaks to privacy obligations rather than to how the environment is secured and audited.

No device claim is made, and the product is scoped to match. The models design candidate molecules at the discovery stage; nothing is presented as a diagnostic or a medical device, and nothing carries a clearance, which is the right position for what this is. No regulatory milestone appears on either side of the work: the company has no pipeline of its own, every published result is preclinical, and any regulatory path that follows a design belongs to the partner that develops it.

Three technical reports set out how the models perform, with the counts attached. The Latent-X1 report gives hit rates target by target, 10 to 64 percent for mini binders across five targets and 91 to 100 percent for macrocycles across three, and runs a bench experiment in which its best binders outbind the strongest produced by RFdiffusion and AlphaProteo. The Latent-X2 report validates across 18 targets with binders confirmed on about half of them from 4 to 24 designs each, and puts nanobodies against TNFL9 through a ten donor human immunogenicity panel using ex vivo T cell and cytokine assays. The Latent-Y report records binders on six of nine campaign targets. Developability is profiled against approved therapeutics.

What the reports do not carry is the step above this. They are posted on arXiv and have not been through peer review, their authors are current or former employees, and the misses are reported in aggregate, as the share of targets that yielded no binder, rather than named target by target the way a subgroup breakdown would name them.

The failure rates come with their method and their denominator. Across the three technical reports a reader can see how often a design works: 10 to 64 percent of mini binders tested per target, binders on about half of 18 antibody targets from 4 to 24 designs each, six of nine Latent-Y campaigns, with the affinities and the design counts stated alongside. A buyer can work out the odds before committing a program.

What the reports do not add is recourse. No guarantee, indemnity or remediation obligation stands behind a design that does not work, and commercial use runs through partnership agreements whose terms are private, so the risk of a failed design sits with the partner. The figures are the company's own, in documents it posted itself rather than ones that went through review.

There is no clinical record surface here, which is expected for a protein design platform, and there is also nothing connecting to the systems a research buyer runs. No electronic lab notebook, laboratory information management system or research data platform is named as integrated, and no standard or connector is offered for moving designs into a partner's own environment. The platform has an API, but it reaches Latent Labs' own models rather than plugging the service into a customer's systems of record, so for a buyer there is nothing to integrate with.

The platform runs one way: hosted by Latent Labs, reached through a browser or an API, with no setup on the user's side. The company describes the service as GDPR compliant and restricts access in line with EU and UK sanctions lists, which places it in a European regulatory frame.

Where the data actually rests is implied rather than committed. No statement names the cloud region a user's targets and designs sit in, and none describes tenant isolation between one customer and another on the shared platform. GDPR compliance describes a legal posture, not a residency guarantee a buyer could hold the company to.

A buyer can size up the model before talking to anyone. The platform has a free allowance, 250 designs a day for an approved researcher, and a credit model for larger campaigns, and a researcher applies for access rather than booking a sales call. The unit of charge, a design, is clear.

The figure a buyer cannot get yet is the price: the credits carry no number, and pharmaceutical use runs through private partnership deals, so the cost of anything past the free tier still takes a conversation.

Coverage is named at the level of target class and individual target, with results behind a good part of it. The models design antibodies, including nanobodies and scFv fragments, alongside macrocyclic peptides and small protein binders, and the hard case the company puts forward is targets embedded in the cell membrane. Named targets run through the published work: macrocyclic peptides against K-Ras that the company reports reaching the affinity of a trillion scale mRNA display screen while testing far fewer sequences, scFvs against benchmark antigens binding into the low picomolar range, nanobodies against TNFL9, and the human transferrin receptor as a route across the blood brain barrier.

What is not established is therapeutic reach. No disease area, indication or clinical program is named, and the company runs no pipeline of its own, so the coverage is described in targets and molecular formats rather than in the conditions a drug would treat.