Valo Health

Opal's models do the target and chemistry work: AI represents millions of patient data points to find disease subtypes, causal inference and statistical genetics rank targets, and predictive chemistry models choose which molecules the lab makes next, learning from each round of results. Partnered programs with Novo Nordisk and Merck KGaA start from that engine.

The licensed patient data, the knowledge graphs and the laboratory chemistry carry value of their own alongside the models.

Models propose targets and molecules and the laboratory tests them in rounds, with the models refined after each round, so experimental results decide what advances. Biology, chemistry and engineering teams work together on each program, and partners take the decisions on clinical development.

No published rule says what result a target or molecule needs before it moves forward.

The approach has four named parts: AI that maps millions of patient data points into a multidimensional space to find subtypes, causal inference and statistical genetics for target selection, proprietary knowledge graphs, and closed loop predictive chemistry. Training data is characterized as more than 17 million de identified patient records with linked biobank samples, plus nference's multimodal data, and a 2018 paper describes geometric deep learning for molecular features.

No model, version or update practice is named.

Opal is Valo's own platform, and the outside parties are named: nference supplies its agentic AI platform and de identified multimodal data under a multiyear partnership, and Charles River's Logica platform carried the lupus work. Data access agreements supply the patient records.

The full list of data providers and service providers is not published.

Valo reported its one clinical readout openly: the Phase 2 SPECTRA study of OPL-0401, randomized, double blind and placebo controlled in 114 patients with diabetic retinopathy over 24 weeks, missed its primary and secondary endpoints, and the company noted the molecule did not use its platform. Platform evidence so far is partner reported, such as new cardiometabolic targets with Novo Nordisk and a lupus candidate found with Charles River, alongside method papers including a 2018 study of geometric deep learning for chemical features.

No platform derived candidate has yet entered a clinical study.

The patient data behind Opal is described as de identified: more than 17 million records through Valo's data agreements, and de identified multimodal data drawn from academic medical centers in the nference federated network. Valo's own privacy policy covers its website and says personal data is kept as long as needed for its purposes.

How the patient datasets are governed, kept and protected inside Valo is not described.

Valo works with de identified patient records licensed through data partners, and the only privacy document on its site is a website privacy notice updated in April 2025.

No statement of business associate status or of why the rule does not apply is published.

The website privacy notice says Valo uses safeguards to protect personal data, and the Opal platform holds licensed patient records and partner research data.

Valo names no security certification or outside assessment.

Opal is an internal discovery platform, not a product sold to clinicians, so no device question arises. Its outputs are targets and small molecule candidates that follow the drug pathway, and partners such as Novo Nordisk run clinical development of the programs they take forward.

OPL-0401 reached Phase 2 under Valo's own sponsorship before the company decided to seek a partner for it.

Causal inference is the stated method for separating cause from correlation in patient data, and patient subtyping is meant to reflect the real world variety of disease across the records Valo can access.

Nothing is published on how the models are evaluated across patient groups or how their behavior is governed.

Laboratory rounds test what the models propose and the models are refined on the results, and partners run their own development before anything reaches patients. Valo published the negative SPECTRA result in full.

No accuracy figure for the platform's predictions, and no remedy for a wrong one, is published.

Patient data arrives as de identified, longitudinal records through data partners such as nference rather than through live connections to hospital systems, and the platform works on that data in bulk for research.

No record system or interface standard is named.

Opal runs inside Valo, which is based in Lexington, Massachusetts, and partners receive programs and data rather than software. Patient data reaches it through data access agreements and nference's federated platform.

Where the platform and its patient datasets are hosted is not stated.

Partnership economics are published for the largest deal: Novo Nordisk's expanded agreement carries an upfront payment, equity investment and near term milestone totalling up to $190 million, about $4.6 billion in milestones across up to 20 programs, research funding and royalties. Valo describes its model as success based, keeping royalties while partners lead clinical development.

Terms for the other partnerships are not disclosed.

Small molecules are the modality, and the programs name the disease areas: obesity, type 2 diabetes and cardiovascular disease with Novo Nordisk, Parkinson's disease with Merck KGaA and The Michael J. Fox Foundation, lupus and other autoimmune disease with Charles River, Pioneering Medicines and Pfizer, and neurological disease with Lundbeck. Novo Nordisk says the collaboration has already identified several new targets.

No program discovered on the platform has yet reported clinical results.