Relay Therapeutics

The model is the engine of core modules inside a platform whose organizing idea is physical. Relay's Dynamo platform puts protein motion at the center: full length proteins are synthesized in house, imaged by cryogenic electron microscopy and ambient temperature crystallography, and simulated by long timescale molecular dynamics to find binding pockets that a static structure would not show.

Machine learning enters at two stated points: the data from activity based, ligand centric and computational screens feeds the machine learning components that identify chemical starting points, including a proprietary machine learning powered DNA encoded library capability the company calls REL-DEL, and in optimization the platform combines machine learning models with molecular dynamics to predict and design compounds with the properties a candidate needs. Remove the learned components and the simulation, structural biology and medicinal chemistry still work; remove the physics and the approach loses the insight it is named for.

The oversight structure is described and the threshold is missing. Computation proposes and experiment disposes throughout: simulations generate a target modulation hypothesis that structural and functional work then tests, machine learning identifies chemical starting points that are screened, and designed compounds are synthesized and characterized before a development candidate is selected. Medicines then pass through preclinical evaluation and clinical trials, where the decisions that reach a patient are made under protocol and regulatory oversight rather than by the platform.

What is not published is any internal standard: no confidence measure with a value at which a predicted pocket or compound is set aside, no hit rate across programs, and no statement of where the models are unreliable.

The scientific approach is described in general terms and nothing about the models is identified. The Dynamo pages set out the three phases of motion based drug design in detail, naming the experimental methods (full length protein synthesis, cryogenic electron microscopy, ambient temperature crystallography) and the computational one (long timescale molecular dynamics of the full length protein), and they place machine learning at two points: components fed by screening data that identify chemical starting points, and models used with simulation during optimization. REL-DEL is named as a machine learning powered DNA encoded library capability.

What is not published is any architecture, training data description, benchmark, prediction accuracy figure or version practice for those machine learning components; the peer reviewed literature the company lists describes the molecules and their clinical results rather than the platform's models.

The work is Relay's own, and while the architecture is described, no model provider is named. The platform integrates the company's protein engineering, structural biology, simulation and machine learning, and the one named outside relationship brings a partner's DNA encoded library screening inward rather than sending data to a model service.

No outside model provider, hosting arrangement or compute supplier is named, no subprocessor list is published, and nothing states what happens to a collaborator's material inside the platform or whether it informs models used elsewhere. Relay is not selling the platform, so these gaps concern the company itself; for a collaborator, the supplier list, the compute arrangement and the position on data contributed to a joint program are all unpublished.

Candidates designed on the platform are in named clinical studies with disclosed timelines, and the methods behind them are published. Zovegalisib (RLY-2608), an allosteric, pan mutant and isoform selective PI3K alpha inhibitor, is in the first in human ReDiscover trial with fulvestrant and a CDK inhibitor in PI3K alpha mutated HR positive, HER2 negative advanced breast cancer, and the company initiated the Phase 3 ReDiscover-2 trial in mid 2025 against capivasertib plus fulvestrant; the same molecule is in the Phase 2 ReInspire study in adults and children with PIK3CA related overgrowth spectrum. The design route is published: the company solved the full length cryogenic electron microscopy structure of PI3K alpha, ran long timescale molecular dynamics to find the conformational differences between wild type and mutant, and used that to design the molecule. Method papers on the platform's molecules are peer reviewed, including the discovery of lirafugratinib (RLY-4008), a selective irreversible FGFR2 inhibitor, in the Proceedings of the National Academy of Sciences in 2024, alongside published discovery and clinical proof of concept work on zovegalisib.

The publications page lists posters and presentations beside journal articles without printing venues, so the page alone does not show where each item appeared. Trial results to date are interim and no product is approved.

How protected information moves through the company's systems is not explained anywhere on the site. Relay publishes a privacy policy at relaytx.com/privacy-policy, which may cover the question. No other page states a retention period, a de-identification standard, or what happens to data used in the discovery platform.

What the platform itself handles is protein structures, simulation output and screening data rather than patient records. The company does run its own clinical trials, so participant health data exists on the sponsor side, governed by trial protocol and site ethics oversight rather than by anything the company publishes. Safety engineering is not addressed.

Relay states no business associate position anywhere on the site. The privacy policy at relaytx.com/privacy-policy may settle whether the company addresses health information or scopes the policy to the website. No other page states business associate status or explains a scope.

The company's own position is structurally different from a vendor's: Relay is a trial sponsor developing its own medicines rather than a business associate processing a covered entity's data, and participant data is governed by trial protocol and site agreements. Nothing published says what agreement governs any health data the company receives.

Controls are described with a check behind them, and no attestation against a recognized framework. No SOC 2 report, ISO 27001 certificate or trust center is published. What exists is the cybersecurity disclosure the company files as a listed registrant, and it is specific rather than boilerplate: a risk management program covering identification, assessment and mitigation, with periodic security assessments, audits and testing informed by industry standards and supported by automated monitoring, internal information security policies, and an incident response plan that is reviewed and periodically updated. Third party risk is handled continuously rather than once, with the cybersecurity practices of vendors and service providers assessed before onboarding and reviewed periodically through the engagement using questionnaires and contractual requirements. Governance is named down to the roles: a director of information technology operations and information security runs the program, reporting to a vice president who meets regularly with the audit committee.

A filing describes a program; it does not evidence an examined one. The standards the assessments are informed by are not named, and neither is the party performing them.

Dynamo is discovery technology used inside the company, not a product offered to anyone, and it is scoped that way: no device claim is made and nothing presents it as a regulated device. The regulatory activity belongs to the medicines instead. Zovegalisib is in a Phase 3 trial in advanced breast cancer and a Phase 2 study in PIK3CA related overgrowth spectrum, and lirafugratinib has been studied in FGFR2 driven solid tumors. Those are the drugs' positions, held by Relay as sponsor under trial protocols, and they say nothing about the software that helped design them. No product is approved.

Nothing published examines how the models behave. No governance framework, model card, evaluation method or performance breakdown appears on the site, and the peer reviewed literature the company lists describes its molecules and their clinical results rather than the machine learning behind them.

The question is narrower here than for a purely data driven platform, because a simulation of protein motion is anchored to the target being studied rather than to a training corpus: where the approach would vary is across target classes, and the company has prosecuted kinases most visibly. The learned components, the screening models that identify chemical starting points and the DNA encoded library capability, are the parts where training data composition would matter, and no coverage or performance analysis is published for either.

Predictions are tested before anything advances, with nothing published standing behind them. The platform's own description puts experiment on both sides of the computation: structural and dynamic data go in, and the compounds the models help design are made and characterized, with candidates then going through preclinical work and clinical trials under regulatory oversight before reaching a patient. That is review before commit, and it is the mechanism a reader can point to.

Nothing is published that measures the computational contribution: no prediction accuracy, no hit rate against a conventional baseline, and no error characteristic for the machine learning components. No warranty, indemnity or remediation commitment applies, and none would: Relay develops its own medicines rather than selling the platform, so a counterparty is a partner or an investor rather than a customer buying an output.

No clinical record integration exists: Relay is a drug developer running its platform internally, and no clinical record system is involved. For a pharma or biotech partner the record systems that matter are its research systems: electronic lab notebooks, laboratory information systems, research data platforms and cloud data environments.

None is named and no integration is claimed; the platform's own instruments and screening data feed it, and the clinical programs reach patients through trial sites rather than through any system Relay connects to. For a company that develops its own medicines, that is an absence rather than a defect.

An internal platform, with location implied rather than committed. Dynamo runs inside Relay: the protein engineering, structural biology, simulation, screening and medicinal chemistry all sit in the company's own operation at 60 Hampshire Street in Cambridge, Massachusetts, and no software is delivered to anyone.

Nothing states where the computation runs, whether outside cloud or high performance computing providers are used, or how a partner's material would be separated from the company's own programs where a collaboration exists.

No price is published, and the commercial posture is discoverable. Dynamo is not sold: Relay develops its own medicines, and the partnerships it enters bring capability inward rather than licensing the platform out, such as the December 2023 agreement applying a partner's DNA encoded library screening alongside Dynamo against a GPCR target, announced without economics.

A reader can therefore establish quickly that there is nothing to buy here and that the relationship on offer is investment or collaboration. As a listed company Relay files financial statements, program status and risk factors on a regular cadence, which describes the company's own economics rather than a price.

Coverage is named and validation sits behind part of it. Relay states an initial focus on precision oncology and genetic disease, and the pipeline names what that means: zovegalisib against PI3K alpha in HR positive, HER2 negative advanced breast cancer, now in Phase 3, and the same molecule in PIK3CA driven vascular anomalies and PIK3CA related overgrowth spectrum in Phase 2, with earlier work on a selective FGFR2 inhibitor in FGFR2 driven solid tumors. Those are settings with clinical data behind them rather than claimed breadth.

Modality is small molecules, which is the real boundary, and the platform's applicability to targets outside the ones it has prosecuted is argued from the approach rather than demonstrated.