Massive Bio

The matching models are the engine of the core service, and the company sells more around them. The solutions page describes language understanding that extracts more than 500 clinical parameters from records, trial and drug matching models, a de-identification model and two large language model assistants (Dr. Arturo AI for physicians, Ask Fiona AI for patients), coordinated since January 2026 by Reticulum Nexus, which the company describes as more than 100 agents.

Around that engine it sells work done by people and data products: patient relations coordinators and oncology nurses who review matches and guide enrollment, concierge logistics, recruitment reporting for sponsors (Trial Vision), and commercial analytics for drug makers (the Drug Utilization Optimizer and NexusPulse). The concierge and navigation service carries value without the models; matching at the stated scale does not exist without them.

Review is built into the matching workflow, and the newer autonomy claims come with no limits. In the company's published study, extraction and matching agents produced candidates that clinicians verified against linked record passages, ambiguous timing was routed to clinician judgment, an undocumented mandatory criterion was treated as not met, and matches reached an oncologist-validated list; the service pages say oncology nurses review matching reports and that final eligibility rests with the investigator.

Two things fall outside that structure. Patients who upload records are told they will see preliminary matches instantly, ahead of the tailored report within 24 hours, and the patient assistant, Ask Fiona AI, answers patients directly. And Reticulum Nexus was announced in January 2026 as a fully autonomous operating system whose agents coordinate next-best actions across patients, sites and sponsors. What the preliminary matches, the assistants and the autonomous agents may not do, and what stops them, is not published.

The architecture is named in detail; a buyer cannot tell when it changes. The company's 2026 study describes OncoAgents (domain-tuned language model agents for extraction, normalization and matching), OncoGraph (an oncology knowledge graph that applies inclusion, exclusion and timing rules deterministically, with version-controlled updates), OncoRecommend (prioritization) and OncoSet (more than 2,000 annotated records, 14,000 trials and 50 tumor subtypes used to train and calibrate the agents).

The solutions page adds language understanding for records, a de-identification model and large language model assistants. The base models under the agents are not named in the study; company research compares GPT-4.1 with GPT-4.0 for report extraction and describes a GPT-4o platform for cell therapy matching, and the company has announced OpenAI work on protocol parameterization and pre-screening. No release history or update notice is published.

The model provider is named and most other parties are listed by type. The company announced in August 2026 that it is an OpenAI Select Partner, building on earlier OpenAI work for protocol parameterization and real-time patient pre-screening, and a 2023 partnership brings Datavant's record retrieval into its operations.

The patient authorization lists recipients by type: laboratories and genomic testing facilities, contract research organizations, pharmaceutical and diagnostic companies, real world data partners and trial sites. What reaches OpenAI and under what retention and training terms, the cloud host, and a named subprocessor list are not published.

A prospective, peer reviewed evaluation of the product with its method open to read, run by the company. Loaiza-Bonilla and colleagues (ESMO Real World Data and Digital Oncology, 2026, open access; five of seven authors company founders or staff) screened 3,804 consecutive patients with metastatic or progressive solid tumors at US academic and community sites from January to December 2024: 157,367 record pages and 23,912 patient and trial pairs, against a gold standard from two oncologists.

The system reached F1 0.82 against 0.47 for GPT-4 zero-shot and 0.67 for GPT-4 chain of thought; screening time fell from a median 120 minutes to about 30 (15 automated, 15 clinical review), and median time to an oncologist-validated list was under 7 days. Ablations credit the knowledge graph and the split into agents.

The authors state the limits: identification only, with no consent, enrollment or completion outcomes; the sites are not named; and they call for external validation across record vendors, sites and populations before broad deployment. Hematologic cancers and the non-US markets the company serves were not in the cohort. Scale figures on the site (200,000 onboarded patients, more than 19,000 trials) describe use, not benefit.

Retention and training use are stated; the de-identification method is not. The patient authorization (last updated 11/10/2025) covers complete medical records including genomic results, mental health treatment, substance use disorder treatment, HIV status and sexually transmitted diseases, with psychotherapy notes only on separate authorization, and states that anonymized and aggregated data created by the company may be used for analytics, AI training and real world evidence research.

The privacy policy (last updated 04/06/2026) keeps patient profile data for ten years after collection or last activity and de-identified or aggregated data without a fixed term; the authorization runs ten years unless revoked, and revocation stops new processing of identifiable data within 30 days without reaching what was already used. No de-identification method is named, which matters for genomic data. Two uses need a direct answer.

The Drug Utilization Optimizer is described as integrating consented patient data with physician profiles to give drug makers' commercial teams real-time notifications of cases, and NexusPulse turns consented clinical and biomarker data into signals for commercial and market access teams; the authorization's stated purposes are eligibility, diagnostic and therapeutic analysis, and the privacy policy says customer information is not sold. Nothing is published on safety engineering for the language model assistants: guardrails, hallucination handling or how a safety event is handled.

Business associate status is stated and a substantive patient document stands behind it; no agreement is published. The patient authorization (last updated 11/10/2025) says the company and its subcontractors operate as business associates bound by business associate agreements, and sets out the records covered, purposes, recipients by type, revocation and access rights, with a named data protection officer.

The privacy policy says the company is not a covered entity, and the company's study ran at health networks under a business associate agreement. Records gathered on a patient's own authorization, as the patient's agent, travel a different path from records a provider shares under an agreement, and the documents do not say which protections follow each.

A recognized attestation is named without the report. The terms say the company maintains SOC 2 certification, the privacy policy says it maintains SOC 2 controls, and the home page carries an AICPA SOC badge. The type, period and scope are not stated and there is no trust center. The privacy policy describes encryption in transit and at rest, access controls, pseudonymization, VPN and firewalls. The data held is sensitive: full records including genomic and mental health information for a reported 200,000 onboarded patients.

No device claim is made, and trial matching is scoped as information. The company says it does not practice medicine, its reports are for educational and informational purposes, treatment decisions rest with the treating physician, and trial participation is decided by the sponsor or investigator.

No regulatory position is stated for the clinical functions offered beside matching: the solutions page lists drug matching among its AI tools, the patient authorization covers analysis of diagnostic options and therapeutic optimization, and Dr. Arturo AI summarizes a patient's disease for physicians.

A subgroup evaluation is published with its groups and one gap, and the per group figures sit in a supplement. The company's prospective study (Loaiza-Bonilla and colleagues, ESMO Real World Data and Digital Oncology, 2026; first and corresponding author the company's co-founder and chief medical and AI officer, five of seven authors company founders or staff; announced by the company in April 2026) compared sensitivity, specificity, precision and F1 across age, sex, race and ethnicity against a pre-specified alert at a 10 point F1 gap.

The text reports that no group crossed it and that the largest difference was about 7 points, with Black patients below white patients, which the authors attribute mainly to fragmented care histories and missing staging. Group sizes and intervals are in Supplementary Table S5, outside the open article text. The figure of F1 by race is captioned illustrative, and its bars (white near 0.87, Black near 0.75) sit further apart than the 7 points in the text. Two questions stay open.

Patients with thin records are less matchable regardless of eligibility, and the study measures accuracy among patients who reached the service, not who never did. And sponsors fund part of the service: the patient authorization discloses that trial sponsors may pay the company for concierge and navigation once a trial is chosen, and nothing states whether sponsor relationships affect which trials are shown or in what order.

A published error rate with its method and denominator, adopted by the company, and nothing behind it beyond that. The prospective study the company announced in April 2026 (Loaiza-Bonilla and colleagues, ESMO Real World Data and Digital Oncology, 2026; 3,804 consecutive patients with metastatic or progressive solid tumors at US academic and community sites during 2024) reports sensitivity and specificity of about 0.84, precision 0.81 and F1 0.82 (95 percent interval 0.81 to 0.83) over 23,912 patient and trial pairs, against eligibility labels from two oncologists who agreed at kappa 0.92.

Pairs with ambiguous timing were routed to clinician judgment and left out of the F1, so the rate describes the pairs the system could decide. At a sensitivity near 0.84, about one eligible pair in six was missed. This is a study result on the version tested, not performance in use. The patient authorization says reports are informational, the company is not liable for clinical outcomes, and enrollment is decided by sponsors and investigators; the terms exclude damages. Patients can request access, amendment or restriction of their records by email.

Records reach the company three ways, one of them through a named network. Patients and physicians upload records in any format; the patient authorization invokes the right of access, under which providers must release records within 30 days to the patient's authorized agent; and a partnership announced in June 2023 brings Datavant's medical record retrieval (Switchboard) into the company's operations.

The company's study describes health network deployments running on the network's own record data, with an oncologist-validated list returned to the record, and names no record system. NexusPulse structures data to the mCODE model. No named record system connection, marketplace listing or FHIR interface is published.

A hosted service with location implied rather than committed. The privacy policy says personal information may be stored and processed in the patient's region or another country where the company and its service providers keep servers, and uses standard contractual clauses for transfers from the EU and EEA to the United States.

The company's study describes a different option for health networks: the system ran in the networks' own environment, on premises or in a role-controlled virtual private cloud, with agent messages kept there. The hosting regions, the cloud provider, tenant isolation and where patient-uploaded records rest are not stated.

No price is published, and how the business is paid is discoverable. The service is free to patients and caregivers; revenue comes from pharmaceutical, contract research and provider contracts, and the patient authorization discloses that a trial sponsor may pay the company for concierge, navigation and last-mile coordination once a patient pursues a specific trial. Products for drug makers (Trial Vision, the Drug Utilization Optimizer, NexusPulse) carry no published terms.

Coverage is named and validated for part of it. The service covers oncology and hematology trials, including pediatric trials, for patients and physicians in 18 countries and 11 languages, searching more than 19,000 active trials.

The company's prospective study validates matching for patients with metastatic or progressive solid tumors at US academic and community sites; hematologic cancers and non-US records were not in it, pediatric patients are not reported, and the authors give extraction fidelity above 95 percent on internal records and about 89 percent on external scanned referrals.