Verge Labs

The model is the product. The company now sells a foundation model for neurology, vBx-1.0, which reconstructs a patient's brain molecular state from measurements that can be taken from a living person, and the platform around it is the data that trained it and the applications it enables: target discovery, target characterization, biomarker identification and patient stratification.

The CONVERGE analysis that finds disease modules in post-mortem tissue feeds the model rather than standing on its own. Remove the model and what is left is a tissue archive and a database; the "virtual biopsy" that is the company's offer exists only as a model output.

The oversight structure is described and the threshold is missing. What the model produces are predictions that inform decisions people make: which targets to pursue, which patients to enroll, which patients may be at risk of an adverse event. The company describes those as inputs to trial design and program decisions taken by drug developers, and the disease modules the model reads out are defined on post-mortem tissue by its own analysis and described as interpretable rather than opaque, so a scientist can see what is being predicted. Predictions are validated against held out donors with true brain values before being offered.

What is not published is any threshold or guidance: no confidence measure a partner sees on a per patient prediction, no statement of the correlation below which a readout should not be relied on, and nothing on what happens when the model's stratification is used to exclude a patient from a trial, which is the consequential use.

The model is identified and the training data characterized, without the version and update discipline behind it. The technical report of 16 June 2026 describes vBx-1.0 concretely: a multimodal transformer of roughly 90 million parameters that fuses a patient's blood RNA sequencing, brain RNA sequencing and genetics into a single 512 dimensional embedding, trained with a contrastive objective that masks modalities and forces the masked view toward the full modality view, so the full picture can be reconstructed from the modalities a living patient can supply. The training set is stated: 3,272 human brains from 1,586 patients across five diseases, each contributing multiple modalities, drawn from an atlas the company puts at 12,000 brain samples across 6,500 patients. The company also says the gains so far have come from scaling paired data rather than compute, and that the model embeds 1,000 patients in 16 seconds on a laptop.

What is not published is version practice: vBx-1.0 is the first release, with the roadmap stated as intent rather than a release cadence, and nothing says how a partner would learn that the model changed under them.

A structurally short chain: the model is the company's own and small enough to run anywhere. vBx-1.0 was built and trained by Verge on its own patient data atlas, at roughly 90 million parameters, and the company states it runs on a single GPU and embeds 1,000 patients in 16 seconds on a laptop, so no outside model service sits between a patient's data and an answer. The comparator methods named in the technical report are baselines it tested against rather than components it uses.

What is not published is the party list around it: no hosting or cloud provider, no subprocessor list, and no statement of who else touches patient material inside the platform or what governs a partner's data once it arrives.

Published research short of independent validation, and no candidate of its own in the clinic now. The technical report of 16 June 2026 is substantive: held out donors, named comparator methods, correlations by disease and brain region, and two translational results stated as trial consequences, a 43 percent reduction in required Parkinson's trial enrollment at fixed power and a 34 percent reduction in dropouts from an on-target adverse event when screening on predicted baseline biology, the latter benchmarked against the company's own ALS trial. The company reports 280 or more novel targets surfaced, an 83 percent preclinical validation rate and two AI discovered drugs reaching clinical candidate stage, and states $1.6 billion in partnerships with top 20 pharmaceutical companies.

Those headline figures carry no method or denominator. The technical report is the company's own rather than peer reviewed or independently replicated, and its own clinical program in ALS missed on efficacy in Phase 1, after which it refocused on the platform.

Nothing published reaches the patient data the platform is built on. This company's asset is human material and the data derived from it: a collection it describes as 12,000 brain samples across 6,500 patients with paired proteomic, genomic and clinical data, more than 900 frozen brain tissue samples, and 15 million single cell profiles, with a model trained on 3,272 brains from 1,586 patients.

The only privacy document published is a website privacy statement, last updated 30 August 2023 and still in the former corporate name, which by its own terms applies to personal information collected through the company's sites. Nothing published states the consent basis under which the tissue and clinical data were obtained, the de-identification standard applied, how long identifiable material is kept, or whether a partner's own patient data entering the platform would train models used for other partners. Safety engineering is not addressed.

No statement of status and no privacy document that reaches the product. The published privacy statement, last updated 30 August 2023 and issued in the company's former name, applies to personal information collected through its websites; it does not reach the research platform or the patient data atlas behind it.

No business associate status is stated, and no scope position explains how the company treats patient derived material, which matters for an organization holding brain tissue, blood profiles and clinical trajectories from thousands of patients and offering to stratify a partner's trial population. Silence is not a stated position.

Security controls are asserted with nothing independent behind them. There is no SOC 2 report, ISO 27001 certificate, penetration testing cadence, security page or trust center on the site. The website privacy statement says administrative, physical and technical safeguards have been implemented to protect personal information and that no security measures can be entirely secure, which is an assertion about website data with no outside check behind it, and it advises the reader to take their own precautions.

That is the whole published position for a company holding human tissue derived genomic and clinical data on thousands of patients and inviting pharmaceutical partners to send trial data into its models.

The platform is scoped as research technology, and no device claim is made. Verge sells target discovery and patient stratification to drug developers; nothing presents the model as a regulated product, and the company holds no clearance and claims none. Its own drug program carried its regulatory position as sponsor: the ALS candidate reached a Phase 1 study that missed on efficacy, after which the company refocused on the platform.

The risk lies in how the model is marketed rather than in any claim it makes. Predicting which patients will respond to a therapy and which are at risk of an adverse event is a clinical use if it ever leaves the research setting, and nothing published states where the company draws that line.

Results are published within the groups where the model is most at risk of failing, with numbers and named baselines. The technical report holds out donors with matched blood and brain tissue entirely from training, feeds the model blood alone, and scores predicted brain module activity against those donors' true brain values, benchmarked against a regularized linear model, a principal component analysis standing in for a published projection method, and scVI, a deep learning expression model. Performance is broken out by disease and brain region rather than reported as one number: an average Pearson correlation of 0.40 for Parkinson's modules in the substantia nigra against baselines topping out at 0.27, and 0.41 in cortex and 0.49 in dorsolateral prefrontal cortex for Alzheimer's, on 26 to 57 patients per task. The report shows both where a competitor method collapses and where its own correlations are weakest.

What it lacks is a comparison across the groups the product serves in the demographic sense. There is no published result by ancestry, sex or age, and for a model trained on a brain bank collection, representativeness is the open question. No governance framework or model card is published either.

A measured error rate is published with its method, its denominators and its baselines. The technical report scores vBx-1.0 on held out donors whose blood and brain samples were kept entirely out of training, reporting average Pearson correlations for predicted brain module activity by disease and region (0.40 in substantia nigra for Parkinson's, 0.41 in cortex and 0.49 in dorsolateral prefrontal cortex for Alzheimer's) against three named baselines on 26 to 57 patients per task, and it states the downstream consequences as figures a partner can check: enriching a Parkinson's trial for likely responders cuts required enrollment by 43 percent at fixed power, and screening on predicted baseline biology would have removed 34 percent of the patients who dropped out of the company's own ALS trial because of an on-target adverse event. Benchmarking against its own failed trial is the unusual part.

The figures are the company's own, in a technical report rather than a peer reviewed paper, and nothing attaches to them commercially: no warranty, remediation or fee consequence is published for a partner whose stratification proves wrong, and the terms of a collaboration are private.

No integration evidence. The platform does hold clinical data: the model is trained on patient clinical trajectories alongside brain and blood molecular profiles, and the company offers to stratify a partner's trial population. How that data arrives is not described as a connection to any system: the atlas was assembled over a decade from tissue collections and studies, and a partner engagement is a collaboration in which data is transferred under an agreement.

No electronic health record, clinical data warehouse, trial data system or interoperability standard is named, no integration is claimed, and no production connection is cited. For a research platform that is an absence rather than a defect, though the route for moving a partner's patient level trial data in and stratification results back is left unstated.

An internal platform, with location implied rather than committed. The model is operated by Verge rather than delivered to partners: a collaboration brings a partner's questions and data to the company, whose operation is in San Francisco. The technical report notes that the model is small by design, runs on a single GPU and could embed a thousand patients on a laptop, and mentions fine tuning embeddings into custom predictors for a partner's program, but nothing states that a partner can run it in its own environment.

Nothing published says where the platform runs, whether outside cloud providers hold the patient data atlas, where a partner's transferred data rests, or how one partner's material is separated from another's.

No price is published, and the commercial structure is discoverable. The platform is not sold as software: the company works through multi year discovery collaborations, and the shape of the terms is public, combining upfront fees, equity, near term payments, milestones on target validation and development, and downstream royalties, with the partner selecting targets and holding an option to license and advance them. One announced four year collaboration carried up to $42 million in upfront fee, equity and near term payments against a potential deal value of $840 million plus royalties, and another triggered near term milestones when the partner elected to pursue two validated targets. The company states $1.6 billion in partnerships with top 20 pharmaceutical companies.

Those headline values are full milestone stacks contingent on clinical success and describe what particular partners negotiated, not a price a new partner can apply; no rate card, unit of charge or access tier is published for the model itself.

Coverage is named narrowly and validation sits behind part of it. The company states its scope plainly: frontier AI for human biology starting with the brain, with an all in human platform for neuroscience drug development. The data and the model are built for neurodegenerative disease, with the model trained across five diseases and results reported for Parkinson's and Alzheimer's by brain region, and the company's own clinical experience is in ALS.

What the platform does within that scope is also named: target discovery, target characterization, biomarker identification and patient stratification. Validation is strongest on the prediction task the technical report measures and thinner on the downstream uses, which rest on modeled trial simulations rather than a completed trial run on the model's stratification. Nothing outside neurology is claimed.