Relay Therapeutics

A three way hybrid in which machine learning is integral but not the differentiating insight. What distinguishes Relay is a scientific thesis about protein motion, executed through molecular dynamics simulation and experimental structural biology, with machine learning applied to the resulting data.

The company's own description leads with conformational dynamics and with synthesising full length proteins through protein engineering, and positions machine learning as components of the Dynamo platform fed by screening data rather than as the platform itself. Machine learning is genuinely load bearing, most explicitly in the REL-DEL DNA encoded library capability which the company describes as machine learning powered and which the ZebiAI acquisition strengthened. Graded alongside Schrodinger and XtalPi as physics and experiment anchored rather than model led.

Oversight is structural and reciprocal by design: computational predictions generate hypotheses that are tested in activity based, ligand centric and computational screens, and the resulting experimental data feeds back into the machine learning components. Neither half runs unchecked, which is the intended shape of the platform rather than a control added afterwards. Held at B because no confidence thresholds, hit rate disclosures or documented decision gates between computational output and synthesis commitment were located.

The scientific thesis is explained clearly and the computational method is not. Company materials describe the three phases of motion based drug design, the role of full length protein synthesis in characterising conformational dynamics, and REL-DEL in outline. What was not located in this review is any model architecture description, benchmark, prediction accuracy figure or peer reviewed methods paper for the machine learning components.

A further caution for diligence: much of the circulating technical detail about this platform originates in secondary commentary and aggregator profiles rather than company or peer reviewed sources, so a reader can easily mistake third party description for company disclosure.

The clinical form of this axis does not reach the discovery workflow, which operates on protein structures, simulation output and screening data with no patient records involved, and that bounds the question as a matter of what the product does rather than what it promises.

Human subject protection in the ongoing trials runs through trial protocol and site ethics oversight rather than through anything this company publishes, which is the correct and ordinary arrangement and means a participant's protections come from the site. What is absent is enumeration on the computational side.

No model or model family is named, no hosting arrangement is described, and no sub processor list was located, and the same caution recorded on the other axis applies here: where technical detail about this platform circulates, much of it traces to secondary commentary rather than to the company, so a reader assembling a picture of the stack from what is available may be assembling something nobody has vouched for.

One question specific to a company that both runs its own trials and operates a discovery platform is worth asking directly: whether anything derived from its own trial participants, including tissue, imaging or response data, reaches the discovery side, and under what basis. Ask that, plus the hosting arrangement and a sub processor list.

Genuinely clinical stage, which places it ahead of most of this category, with multiple oncology programmes in trials and RLY-2608, a mutant selective PI3K alpha inhibitor for breast cancer, reported as the most advanced and reported to have produced tumour responses in patients who had progressed on prior PI3K therapy. Unlike several peers the clinical assets are platform originated rather than in licensed.

Held at B rather than A on evidence provenance: the specific clinical results located in this review came largely from secondary coverage rather than from company releases or peer reviewed publication, so a buyer should confirm current programme status and data directly from company filings before relying on it. No approval exists.

Not applicable in the provider sense and rated accordingly rather than penalized. The platform operates on protein structures, simulation output and screening data with no patient records in the discovery workflow. Human subject protection in the ongoing oncology trials runs through trial protocol and site institutional review board oversight rather than vendor policy.

Not applicable. The company is a clinical stage therapeutics developer rather than a business associate of covered entities handling protected health information.

No SOC 2, ISO 27001 or equivalent attestation was located and no trust centre was found. The substance sits in the company's annual report instead, where as a Nasdaq registrant it files a required cybersecurity disclosure, and that disclosure is specific rather than boilerplate.

It describes a cybersecurity risk management programme covering identification, assessment and mitigation, with periodic security assessments, audits and testing informed by industry standards and supported by automated monitoring tooling, and internal information security policies including an incident response plan that is reviewed and periodically updated.

The standout is third party risk handling, and it is better than most peers manage. The company states it assesses and reviews the cybersecurity practices of vendors and service providers before onboarding and then periodically throughout the engagement, using vendor questionnaires and contractual requirements. Reviewing a supplier once at selection is common; reviewing it continuously is not, and for a business whose computational work depends on external screening and compute relationships that is the right control to have.

Governance is named down to the individuals: a director of information technology operations and information security runs the programme day to day, reporting to a vice president of information technology who meets regularly with the audit committee to review information security risk and the threat landscape, and the filing states both roles are held by people with over twenty years of relevant experience.

Held at B. A filing describes a programme; it does not evidence an examined one, and unlike the strongest disclosure in this category this one says its practices are informed by industry standards without naming which. Ask which framework, and ask for the results of the periodic assessments it describes.

Correctly structured, with the Dynamo platform not presented as a regulated device and regulatory activity at the asset level. Multiple small molecule oncology programmes are in clinical trials under active INDs, with the most advanced reported in Phase 2. No approval has been obtained and no breakthrough or expedited designation was located in this review.

No governance framework, model card or performance characterisation was located, and a second search directed at the scientific literature rather than product pages did not change that. That matters because the same approach did overturn absence findings for several peers in this category, so this is a tested absence rather than an unretrieved one.

The domain relevant question is narrower here than for purely data driven peers, because a physics simulation of protein motion is anchored to the target rather than to historical chemistry and is therefore less exposed to inheriting past medicinal chemistry preferences. The exposure that remains sits in the machine learning components, and the company's own platform description confirms where: screening data feeds the machine learning parts of the platform, and it operates a proprietary encoded library capability whose data trains models used for hit finding. A model built on encoded library screening data carries the chemical coverage limits of those libraries, so performance would be expected to differ between chemical space the libraries sample well and space they do not. Nothing published addresses it.

One piece of language is worth flagging rather than crediting. The acquisition that brought this capability in was announced as delivering validated machine learning models alongside curated experimental datasets. Validated is a commercial characterisation in a transaction announcement, not evidence: no validation study, benchmark or performance breakdown accompanying those models was located then or since.

A sourcing caution belongs on this record too, matching one already held against a peer. Several third party profiling sites attribute specific capabilities to this company, including foundation models trained on billions of molecular interactions and petabyte scale proprietary datasets. Those figures do not trace to company filings, the platform description or peer reviewed work, and should not be used in diligence.

The scientific thesis is explained clearly and the computational method is not, and that split decides the grade. Company materials describe the phases of motion based drug design, the role of full length protein synthesis in characterising conformational dynamics, and the discovery engine in outline, which is a real account of why the approach might work.

What was not located in this review is any model architecture description, benchmark, prediction accuracy figure or peer reviewed methods paper for the machine learning components, and no warranty, indemnity or remediation commitment. A thesis about protein motion is a biological argument; the claim being sold is that a computational system exploits that biology better than alternatives, and nothing published tests the second.

A further caution belongs on the record for diligence purposes and it is the second instance of this pattern in the lane. Much of the circulating technical detail about this platform originates in secondary commentary and aggregator profiles rather than company or peer reviewed sources, so a reader can easily mistake third party description for company disclosure, and a diligence process that quotes it back is quoting something the company never said and cannot be held to. Check the provenance of any technical detail you have been shown. Ask for the model architecture, prospective hit rates against a conventional baseline, and a peer reviewed methods paper.

Not applicable. This is a discovery and clinical development organization with no provider workflow surface and no EHR touchpoint.

Not applicable in the software sense. Dynamo is operated internally and coupled to the company's own protein engineering and structural biology capability rather than licensed to customers, so tenancy and residency questions do not arise for a counterparty. Partnerships such as the December 2023 X-Chem agreement bring external screening capability inward rather than pushing the platform outward.

Public listing on Nasdaq under RLAY means financial statements, programme status and risk factors are filed on a regular cadence and are inspectable by anyone, which is the baseline this index credits. Held at B rather than A because no segment level economics, partnership deal terms or platform commercial disclosure of the kind Schrodinger, XtalPi and Generate:Biomedicines provide was located in this review, and because the company sells no platform access, so the only commercial surface is the equity and the pipeline behind it. The X-Chem partnership was announced without disclosed economics.

Concentrated by design. Modality is small molecules only, and the disclosed therapeutic focus is precision oncology with additional reported work in fibrosis and genetic disease. The platform thesis narrows scope further in a way worth understanding: motion based design is most valuable against targets whose conformational dynamics gate druggability, which is a specific subset of hard targets rather than a general capability, and the company positions it that way rather than claiming universal applicability.