Variational AI

Enki creates molecular structures from a preclinical target product profile. Researchers specify biological targets, unwanted activity and chemical properties, and the learned model searches for compounds that balance those requirements. This generation step is the service sold to biotechnology and pharmaceutical partners, including projects with Merck and ImmVue. A medicinal chemistry team then selects structures for synthesis and experimental testing.

Lead optimization can also preserve a chosen scaffold while changing other properties. The product therefore depends on learned molecular representations and generative optimization, with human chemistry work downstream. Its commercial output is candidate chemistry rather than a database of existing compounds or a research staffing service.

Chemists define the target profile, inspect generated structures and choose which compounds advance to synthesis and testing. The Life Chemicals example makes that selection concrete: 21 of 40 generated structures entered route development. Experimental findings can then guide another optimization cycle. The product's synthetic feasibility claim includes a 90% success figure and describes design intent, so it does not promise that every generated structure will work in the laboratory.

This is a meaningful human decision point before physical experimentation. Public workflow descriptions do not specify uncertainty thresholds, automatic stop conditions or a formal process for investigating a misleading prediction after it has influenced a program.

The company identifies Enki as its proprietary generative foundation model and describes a variational autoencoder that learns a continuous chemical representation. Training combines experimental and computational observations, curated public information and internal data; partner data can support a tailored model.

The April 2026 Enki 4 release names an architecture update and expanded target coverage, while the current product page lists a larger target set than the original release announcement. This provides a model family, technical method and visible release history. It does not provide a complete model card, dataset inventory, benchmark split or customer change notification policy that would establish how every production version is evaluated and replaced.

Enki is identified as Variational AI's own generative model, built from curated public chemistry information and internal data, with partner specific adaptation in some agreements. That makes the central learned component identifiable without relying on an unnamed external language model. AWS supplies the cloud environment, and the privacy policy names Google Analytics and HubSpot for broader service analytics and communications.

Those providers are not described as the source of molecular inference. The public account does not enumerate every library, specialist model or processing subcontractor involved in a project, nor their individual retention and training restrictions. It supports an identifiable core model and host, with remaining dependencies incompletely mapped.

The Life Chemicals dual kinase project provides assay counts and a reproducible experimental outline. Enki generated 40 structures, 21 advanced to route development and 19 were synthesized. Biochemical testing found activity against at least one of EGFR or FGFR1 in 15 of those 19, with five active against both. The report identifies potency thresholds, comparator kinases and the assay method.

Separately, Rakovina selected 35 of 138 generated ATR candidates for laboratory work and reported subsequent mouse exposure studies. These are collaborative preclinical results, with selection between generation and testing. They do not establish clinical benefit, broad success across all targets or independently replicated platform performance.

Partner research data can be used to adapt the model for a particular program. Merck's collaboration expressly involves training a tailored Enki model on Merck's proprietary data, while the cloud architecture separates each partner's environment from internal computation. The general privacy policy limits personal information retention by purpose and legal need rather than a fixed deletion period.

It does not establish a complete lifecycle for compound uploads, model checkpoints or derived predictions. A public promise excluding all customer material from shared model training, detailed deidentification procedures and separate retention windows for research inputs and outputs are not specified. Dedicated isolation does not by itself answer those reuse questions. Public materials also do not define model specific safety testing or the handling of an AI safety incident.

The December 2025 privacy policy explicitly includes the Enki application and describes Variational AI as a Canadian company acting as a data controller under GDPR. It provides access, correction, deletion and objection rights, explains processing purposes and requires confidentiality and purpose restrictions when personal information is disclosed for business operations. Those are substantive service privacy commitments for users of the application.

No public US business associate agreement or plan specific PHI terms accompany them. The service concerns molecular research, but that focus alone does not establish a contractual prohibition on patient data or permission to submit protected health information.

The security description centers on separate AWS environments for pharmaceutical partners and the segregation of proprietary research information. The application privacy policy also describes protection of personal information and contractual limits on service providers. These are architecture and handling statements. They do not amount to a public independent security assessment of Variational AI's operating controls.

No company specific SOC 2 report, ISO 27001 certification claim or accessible audit artifact accompanies the reviewed product and collaboration materials. AWS's own assurance programs would concern the infrastructure provider rather than establish the vendor's access administration, development practices or incident response performance. The public record remains limited to the stated controls.

Enki is positioned as software for preclinical molecular design. Its outputs are chemical structures and predicted properties that researchers evaluate through synthesis, biochemical assays and later studies. The Rakovina material distinguishes initial target inhibition from mouse pharmacokinetic work and further characterization still required. Merck retains responsibility for developing and commercializing compounds from its collaboration.

These activities establish a discovery scope rather than a patient diagnosis or treatment recommendation. No FDA clearance or approval is presented for Enki itself. The success of a design run, or a compound's activity in an assay, does not mean the software or resulting investigational medicine has received regulatory authorization.

The scientific workflow uses explicit target objectives and experimental checks rather than an unreviewed list of generated molecules. Project reports identify potency, selectivity and synthesis constraints, with follow up assays informing further chemistry. These practices make particular design results inspectable, but they do not describe a companywide AI governance framework.

Public materials do not define a model risk committee, a release approval standard or a procedure for addressing systematic performance differences across underrepresented chemical classes. The low data dual kinase example is a research demonstration, not a population fairness evaluation. For drug discovery, the relevant unresolved issue is reliability across target classes and training coverage.

The dual kinase report makes the model's practical limitations measurable. Of 40 generated structures, 21 were selected for synthesis development, 19 were made and five showed the requested dual activity in biochemical testing. That denominator and assay description allow a reader to distinguish predicted suitability, synthetic feasibility and measured activity. The result is not a promise that all future designs will succeed.

Researchers retain the selection and laboratory validation steps, and no public remedy guarantees replacement chemistry, reimbursement for failed experiments or compensation for downstream development decisions. Standard website terms also do not establish the negotiated recourse available under a scientific collaboration contract.

Input to Enki is a target product profile, with optional proprietary scientific data for a tailored collaboration. Output consists of candidate structures and properties that chemists use in subsequent synthesis and testing. The Merck agreement confirms use of partner data, and the product describes fitting into medicinal chemistry workflows. Neither establishes a named production connector, a documented API or a transfer contract with a laboratory information system.

Public materials do not identify EHR connections, electronic notebook integrations or bidirectional laboratory updates. The evidenced interoperability is therefore a scientific collaboration workflow, with the technical transport and automation between the customer's systems and Enki left unspecified.

A July 2026 architecture account identifies AWS and a dedicated virtual private cloud for each partner. Proprietary partner information is separated from the company's internal compute environment, supporting projects that need isolated data and model work. This is more specific than a generic reference to cloud hosting. The account does not name the AWS region used for a customer's project, available residency choices or the locations of backups and support access.

The company's Vancouver address is an office location, not a hosting commitment. Public materials also do not establish an on premises deployment option or a complete cross border processing arrangement for the application.

Commercial work is organized around a partner's target profile and the rights to resulting compounds. The ImmVue agreement describes an initial fee and a payment tied to success, with exclusive ownership rights for the partner. Merck's two target collaboration also includes an upfront payment and milestones, potentially totaling $349 million. That ceiling describes a particular development agreement, not the cost of accessing Enki or ordering a design run.

Public pages do not give a standard project minimum, per target rate, subscription price or implementation charge. The main economic details available are the payment structure, scientific scope and allocation of compound rights.

Enki serves early drug discovery teams working with small molecules and related therapeutic modalities. Current target coverage includes kinases, GPCRs, proteases, ion channels, oxidoreductases and hydrolases. Enki 4 also extends the design approach to degraders, molecular glues and antibody conjugate payloads. Named work includes dual kinase inhibitors with Life Chemicals, CNS directed ATR inhibitors with Rakovina and immuno oncology chemistry with ImmVue.

Those projects demonstrate use in particular discovery settings. They do not validate every advertised target or modality. Clinical care, patient selection and prescribing are outside the described workflow, while the broader target list represents available design coverage rather than demonstrated drug efficacy.