Chemify

Machine learning powers a core module here rather than the whole business, and the reason is an open question about what Chemputation is. The company describes it as a universal chemical programming language combined with advanced robotics and machine learning, and elsewhere in its own press material describes Chemputation itself as its purpose built artificial intelligence. A chemistry society describing the same platform calls it a programming language, robotics and a database of validated reactions. A programming language is not a model and a reaction database is not a model, and nothing published reconciles the two accounts.

Remove the machine learning and a robotic facility that executes coded synthesis routes drawn from a validated library is still a business, and a substantial one: the chemical programming language, the automation and the Maryhill facility all stand without it. What the models do is the design and selection work, and the company names those components specifically, a quantum mechanics and inverse design process for proposing molecules and a route optimization system for choosing how to make them, with real time reaction data feeding predictive analytics that tune pathways and anticipate errors. Those capabilities do not exist without the learned layer. The center of gravity is automation and manufacturing.

Oversight is described with real mechanisms at three points. A design is constrained before it is proposed: every output is stated to be synthesis aware, drawn from a database of reactions the company has validated and automated on its own robotics rather than from literature it considers unreliable. A route is checked before it is run, with safety and viability verified and guardrails and error checks encoded into the chemical programming language. And the result is checked after: identity, purity, yield and properties are verified and linked to the procedure and telemetry that produced them. The partner sets the question, the constraints and the acceptance criteria, and the company states that the partner remains the author.

One part is missing, and one claim reads narrower than it sounds. Nothing published states at what point a person must decide rather than may, what a failed run does beyond marking a boundary on the map, or what happens when the models and the robot disagree. And the technology page states that the closed loop approach eliminates the synthesis bottleneck, AI hallucinations and the delay from idea to compound. In context that reads as a claim about what the loop removes from the workflow, since an unmakeable proposal fails at the bench, rather than a claim that every model output is correct.

Component by component, the approach is named. Quantum Chemputation is the design process, described as uniting quantum mechanics with inverse design so that what is physically allowable, what is wanted and what can be made are resolved together, with every design output stated to be synthesis aware. ASSEMBLER is the route optimization software, drawing on a proprietary database of reaction classes the company says have been validated, optimized and automated on its own robotics rather than taken from literature. Chi-DL is the chemical programming language, described as hardware agnostic, with guardrails and error checks encoded into every validated reaction and multi step route. The data that trains the models is characterized by type: reagents, catalysts, temperature, pressure and yield captured in real time from every reaction, feeding predictive analytics that optimize pathways and anticipate errors.

What is not identified is the model layer itself. The artificial intelligence carries the label ML Chemical Intelligence and nothing behind that label is named: no architecture, no model class, no parameter count, no training corpus size and no benchmark. Nothing states which parts of the system are learned and which are rule based, and the company's own descriptions vary, with press material calling Chemputation itself its purpose built artificial intelligence while an outside chemistry society describes the same system as a programming language, robotics and a reaction database. No version identifier, release history or update practice is published for any component.

The chain is structurally short and the company describes it that way throughout. Every named component is its own: chi-DL is stated to be invented by Chemify, ASSEMBLER is its route optimization software, the reaction database is proprietary and built from reactions validated on its own robotics, the robotics are its own, and the synthesis facility is one it built and paid for. No external model provider sits between a question and a compound, and none is named or implied anywhere. The company's argument for reliability rests on that closure: it rejects literature derived routes in favor of ones it has run itself.

What is not published is anything outside the model layer. No compute or hosting provider is named, no open source components or third party licenses are disclosed, and there is no subprocessor list, which matters now that Genesis promises encryption in transit and at rest and a protected domain a partner's data does not leave. On the physical side the same silence applies: no reagent, instrument or contract manufacturing supplier is named behind a facility whose whole proposition is that it makes things.

Published research short of independent validation, with a physical operation behind it.

Three peer reviewed papers appeared within two weeks of each other in April 2026, in PNAS, Nature Communications Chemistry and Nature Communications Biology, covering molecular discovery, automated synthesis translated from published literature, and application to cancer drug development. The Maryhill Chemifarm has been operating since June 2025 at a stated build cost of $16 million. Two funded programs exist with an external partner, Lgenia, running enzyme and whole cell assays against compounds Chemify designs and synthesizes, supported by Gates Foundation grants reaching $3.2 million cumulative, and a malaria program was described in December 2025 as moving toward investigational new drug enablement within twelve months and aiming at a single dose cure.

The qualifications are specific. The papers were authored by the founder with co-authors from the university where he holds his chair and from which the company spun out, so they are peer reviewed but not independent of the company. The superlative claims, the world's most advanced automated synthesis facility and the world's largest curated reaction library, carry no third party measurement and no published size for the library. Nothing has reached a clinical study, the malaria timeline is forward looking with no filing or registration published, and no pharmaceutical partner or commercial program outcome is disclosed anywhere.

Categorical commitments are published, and they are the ones a partner in this business would ask for. The Genesis product page states that a partner's molecules, routes and results are encrypted in transit and at rest and do not leave their protected domain, that a partner's world model is theirs alone and is never shared, pooled or reused across partners, and that every route is checked for safety and viability before anything is made and is then run under stated operating controls. The never pooled across partners clause is the equivalent here of a no training on customer data commitment, and it matters because the platform's whole economic argument is that every run compounds into a model.

What is missing is the schedule behind those commitments. No retention window is stated for a partner's routes, telemetry or analytics, nothing says what happens to a partner's data when an engagement ends, and no de-identification or aggregation practice is described for the reaction knowledge the company says accrues to its own validated database alongside the partner's private model. The only privacy document the company publishes is a website notice last updated 15 June 2022, which reaches visitors rather than partner chemistry and still lists the company's old university address.

No protected health information is at stake; the inputs are molecules, routes and analytical results. Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is partially addressed, since the company describes guardrails and error checks encoded in its chemical programming language and a safety and viability check on every route, and it publishes no account of how a safety event is handled when one occurs.

No status is stated and the only privacy document reaches the website. It defines the Service it governs as the Website, was last updated on 15 June 2022, and still gives the company's registered address as the School of Chemistry at the University of Glasgow rather than either of the two sites it now operates from. Protected health information is not mentioned, no business associate position is taken, and no statement says whether anything the platform handles touches patient data.

Chemify Ltd is a United Kingdom company, so an equivalent home regime documented to a comparable depth could cover the same ground. None is documented for the service rather than the website, and this notice does not reach the service. No published statement says the platform processes no patient data either, though the inputs are molecules, routes and analytical results, and the disclosed programs are preclinical tuberculosis and malaria work.

Controls are asserted with nothing independent behind them. The Genesis product page states that a partner's molecules, routes and results are encrypted in transit and at rest, that nothing leaves the partner's protected domain, that a partner's world model is never shared, pooled or reused across partners, and that routes are run under strict operating controls. Those are the right claims to make and none of them is backed by an attestation.

No security page, trust center, certification or attestation appears anywhere across the site's eight sections or its footer. The three legal documents published address information security only through a standard disclaimer that transmission over the internet cannot be guaranteed secure. No SOC 2 report, ISO/IEC 27001 certificate or independent penetration testing on a stated cadence is named.

The absence is more conspicuous here than at a design only company, because this one operates a physical chemical facility where certification against recognized safety and quality standards is ordinary industry practice, and because Genesis is sold on a promise of confidentiality between competing partners sharing one plant. Nothing published either claims or documents a certification of either kind. That is a statement about disclosure rather than about the underlying compliance, which United Kingdom law governs regardless.

The platform makes no device claim and is scoped accordingly. The design software, the chemical programming language and the robotics are not offered as a medical device and carry no clearance, which is the appropriate position for what they are.

Nothing is cleared and nothing adverse appears. No candidate has an approval or an investigational application on the public record, and as a Scottish company Chemify would engage United Kingdom and European regulators before the Food and Drug Administration in any case. A malaria program was described in December 2025 as moving toward investigational new drug enablement within twelve months and aiming at a single dose cure, and no filing, clearance or trial registration is published for it.

The company also operates a chemical manufacturing facility, which carries its own regulatory regime for process and workplace safety entirely separate from medicines regulation. Nothing published states the facility's standing under that regime either.

No governance framework, model card, evaluation charter or responsible use policy appears anywhere across the site's eight sections. What partially substitutes is the published science: three peer reviewed papers appeared in April 2026, in PNAS, Nature Communications Chemistry and Nature Communications Biology, covering molecular discovery, automated synthesis translated from published literature, and application to cancer drug work, and the founder's group has published on the underlying approach over two decades. That exposes the methods and their validation to external scrutiny.

What none of it supplies is a statement of when the company judges its own system unreliable. No results are published by reaction class, molecular complexity or target family, so a reader cannot see where route prediction degrades, and the stated size of the validated reaction library, which is what bounds everything the platform can make, is never given. That matters more here than at a design only company, because the failure mode is not a wrong answer on a screen but a physical synthesis that consumed reagents and produced something unintended.

One provenance point belongs with the papers. They were authored by the founder with co-authors from the University of Glasgow, where he holds his chair and from which the company spun out. The work is peer reviewed. It is not independent of the company's founder.

Mechanisms exist that let an output be challenged, and there are several of them because the output is a physical compound rather than a prediction. The Genesis product page states that successes, failures and partial outcomes stay linked to the exact procedure, telemetry and analytics, that routes, precedent, conditions, deviations and analytics remain inspectable, and that every delivered result is verified for identity, purity, yield and properties. Every suggestion carries an executable route or the experiment needed to establish one, so a proposal can be traced to the chemistry that would produce it, and the chemical programming language encodes guardrails and error checks.

Nothing stands behind the output. No success rate, failure rate or accuracy figure is published for the design or route prediction models with a method and a denominator, so there is no measurement to hold the company to. No guarantee, indemnity or remediation obligation is published for a partner whose program follows a route that does not work, and the commercial framing points at economics rather than recourse: the promise is a lower cost per validated decision rather than a remedy when a decision is wrong.

No clinical record integration exists. There is no electronic health record touchpoint, no clinical workflow surface and no provider connection, because the company designs and physically makes molecules rather than serving health systems.

Among the research systems of record a partner would run, nothing is named either. No electronic laboratory notebook, laboratory information management system, research data platform or partner cloud environment is named as connected, and no connector or application programming interface is offered. The company does publish an interoperability artifact, but it points the other way: chi-DL is described as a hardware agnostic universal chemical programming language, so a synthesis procedure written in it is portable across automation systems rather than tied to one vendor's robotics. That is a standard the company invented and uses, with no third party adoption, marketplace listing or partner deployment named to verify it against. A partner reaches this platform by reserving capacity and receiving compounds, not by connecting a system to it.

Deployment here is physical, and that half is disclosed in concrete detail. Two addresses are published on every page: the Advanced Research Centre at 11 Chapel Lane in Glasgow and the Maryhill Chemifarm on the Oakbank Industrial Estate, with the Chemifarm operating since June 2025. A network of facilities is stated as the intention and a Silicon Valley site is planned. The Genesis product page adds that a partner's access is reserved physical capacity at that facility, so where the work happens is not in question.

The data half is thinner. Genesis states that a partner's molecules, routes and results are encrypted in transit and at rest and do not leave their protected domain, which is a security commitment rather than a residency one: nothing names a cloud or compute provider, states which jurisdiction holds a partner's design data, describes retention, or says what happens to the synthesis telemetry a run generates. A United Kingdom company planning a United States site would face cross border questions that nothing published anticipates.

The selling model is easy to find, even though no price is. The site carries a partnering section, a Ways We Work section describing engagement models, and a Genesis product page that states the unit the business is sold in: reserved physical capacity at the synthesis facility, taken by a founding cohort of partners, with the stated economic argument being a lower cost per validated decision rather than a cheaper synthesis run. How this is sold and what drives the cost are both public before any contact is made.

What is published in figures is grant income rather than price, and it is money received rather than a ceiling, which is unusual: $1.6 million in follow on funding from the Gates Foundation announced in December 2025, against $3.2 million cumulative across both phases of the tuberculosis and malaria work. A trade report renders the same cumulative total in sterling where the company's own release states dollars, so two figures circulate for one grant.

Beyond the grants there is nothing. No pharmaceutical contract, license, synthesis rate or facility access price appears anywhere, despite a facility that has been operating since June 2025 and a unit of charge the Genesis page describes but does not price. For roughly $107 million raised, not a single commercial agreement has been disclosed.

Coverage is named across two dimensions and part of it has work behind it.

Across the workflow this company spans design, route selection, physical synthesis, purification and analytical verification, closing the make step itself rather than handing it to somebody else, with a facility operating since June 2025 and reserved capacity sold as the unit of access. Across application the reach is stated beyond medicine by design, with catalysts, electronic materials, agriculture and green energy named alongside drugs, and disease work covering tuberculosis, malaria and cancer.

Validation reaches part of that. The Gates funded tuberculosis and malaria programs are named with an external partner running the assays and a stated development timeline for one of them, and three peer reviewed papers in April 2026 cover discovery, literature to synthesis translation and cancer drug work.

The reach is asserted more widely than it is evidenced. No materials, agriculture or energy customer or program is named, nothing has reached a clinical study, and the stated intention to build a global network of facilities currently rests on one operating site with a second planned.