Deep Genomics

There is nothing else here. The company presents no pipeline, no clinical asset and no service other than what its models produce, and it now describes itself around the biological foundation model platform rather than around a therapeutic ambition.

Five models are named and each does a different job: BigRNA for tissue specific regulation and variant effect prediction, REPRESS for microRNA binding and messenger RNA degradation, DeepADAR for guide RNA design in RNA editing, and FlashRNA and Enigma for regulatory genomics. The laboratories exist to generate training and validation data for those models rather than to run a discovery operation the models assist, which the company states directly: datasets are designed to fit the purpose of training from scratch or fine tuning for a specific application.

Remove the models and nothing remains to sell, and that holds in a strong sense, because everything that was not the platform has fallen away. A stated plan in September 2023 to become a forward integrated biopharmaceutical company advancing multiple therapeutics is no longer reflected anywhere in what the company publishes today.

The validation architecture is described concretely and backed by physical capacity. The company operates lab in the loop workflows across two facilities totaling more than 10,000 square feet in Toronto and Cambridge, Massachusetts, generates datasets fit to the purpose of training or fine tuning rather than reusing whatever exists, and validates predictions using orthogonal assays, including functional assays, before feeding results back into training. Orthogonal validation is the substantive commitment there: checking a prediction by a method that does not share the original's assumptions is the honest form of the test, and the model papers carry it through to published task suites.

The software side is described too, with model evaluation frameworks and experiment tracking integrated into the flagship model's workflow, so a prediction and the run that produced it are linked.

What is absent is the human layer. Nothing published states who decides which predictions are taken forward, what a contradicting assay result does beyond retraining, at what confidence a prediction is not acted on, or which judgments require a scientist rather than allowing one. The company also publishes no round by round hit rates from inside the loop, so the loop's own yield is not visible.

The models, their training data and how they change over time are all published.

The models are identified individually rather than gestured at as a platform. BigRNA is a biological foundation model with close to two billion tunable parameters, trained on over a trillion signals derived from high throughput sequencing, used for variant effect prediction, on and off target effect prediction for genetic medicines, and sequence embeddings feeding other models. REPRESS predicts cell type specific microRNA binding and messenger RNA degradation directly from sequence. DeepADAR designs guide RNAs for ADAR mediated editing. FlashRNA and Enigma are named as efficient models for regulatory genomics, each with a preprint and a digital object identifier. Training data is characterized per model: the trillion signals for BigRNA, sixteen million endogenous editing sites plus a custom synthetic guide RNA screening dataset for DeepADAR.

Versioning and update practice is published with specifics. The company states that BigRNA has evolved since its initial development in 2021 and names the changes: prediction resolution moved from 128 base pairs to one, new training datasets relevant to disease contexts were integrated, and architectural improvements were made for predictive power and efficiency. The workflow tooling is named too, including experiment tracking integration and model evaluation frameworks.

The papers describing the commercial models, though, sit on a preprint server rather than in journals: the BigRNA manuscript has been public since September 2023 with no peer reviewed version, which is three years for the company's flagship claim, and FlashRNA, Enigma and REPRESS are all preprints from 2025. The peer reviewed record, which includes Science, Nature Biotechnology, Nature Medicine and Molecular Systems Biology, is largely the founder's academic lineage from before the product era.

Disclosure here is partial but substantial, and it rests on specifics rather than on policy. A third party dependency is named outright: an experiment tracking platform integrated into the flagship model's workflows, which puts an actual vendor inside the machine learning stack on view. Training inputs are characterized by scale and type rather than waved at: over a trillion signals from high throughput sequencing for BigRNA, sixteen million observed editing sites plus a custom synthetic screening dataset for DeepADAR. The models are described as trained from scratch or fine tuned on the company's own data rather than adapted from an unnamed base, so no external model provider sits between the data and the output.

What remains absent is the provenance that matters most. No dataset, cohort or consortium is named as the source of the paired human genotype and RNA expression data the flagship model learned from, and no data use terms are stated for material that originated with human participants. No compute or hosting provider is disclosed, there is no open source component inventory, and with no partner named anywhere there is nothing stating what governs partner data on arrival.

The evidence splits into two halves, and almost all of it sits in one.

The scientific half is deep: seventeen listed publications including Science, Nature Biotechnology twice, Nature Medicine, Molecular Systems Biology and Proceedings of the IEEE, with model specific work reporting checkable results, including a study across fourteen genes in which the flagship model designed effective tissue specific steric blocking oligonucleotides for targets including Wilson disease and spinal muscular atrophy. Two peer reviewed papers address off target liability directly and one of them is published as updated industry recommendations.

The operational half is close to empty. No pharmaceutical partner is named anywhere, no collaboration or program outcome is disclosed, no pipeline exists and no molecule is in a clinical study. There is nothing to count deployments against.

The science comes with three qualifications. The landmark peer reviewed papers are largely the founder's academic lineage from 2015 to 2018, from around or before the product era. Every paper describing the current commercial models sits on a preprint server, the flagship BigRNA manuscript public since September 2023 and still not published in a peer reviewed journal. And the most quoted external endorsement of that work came from a member of the company's own scientific advisory board, which makes it a related party statement rather than independent commentary.

The only privacy document is a website notice, and it answers none of the questions about how the platform handles data. Titled the Deep Genomics Website Privacy Policy and last updated 24 August 2026, it explains the collection, use and disclosure of personal information through the website and other interactions with a visitor. Its safeguards section offers reasonable administrative, technical and physical measures and need to know access, and its retention clause says information is kept no longer than necessary. No retention schedule, no statement of what reaches a model and no de-identification practice appears anywhere.

No protected health information is obviously at stake, though that is less settled than it looks. The platform trains on paired genotype and RNA expression data across many individuals, which is human genomic data whatever its provenance, and the company says nothing about where that data came from, under what consent, or how it is governed. Its own notice describes the business as an AI powered solution that identifies RNA therapeutic candidates, without connecting that to the data question at all. The company also states that it works in support of major pharmaceutical partners, and nothing published says whether partner data is retained, whether it trains models serving other work, or how programs are separated.

Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is not published.

No status is stated and the only privacy document reaches the website. It is titled the Deep Genomics Website Privacy Policy, was last updated 24 August 2026, and scopes itself to personal information collected through the website and other interactions with a visitor. Protected health information is not mentioned, the United States health privacy rule is not mentioned, and no statement says whether the platform processes patient data.

Deep Genomics Incorporated is a Canadian company with a Cambridge, Massachusetts facility, so an equivalent Canadian regime could in principle stand in for the United States rule. Nothing published applies one to the platform: the notice reaches the website rather than the platform, and no Canadian federal or provincial privacy regime is invoked for the research data either. The company also takes no published position on whether its models touch patient data, which matters because the training data is paired human genotype and RNA expression.

Nothing is published. There is no security page, no trust center, no certification and no attestation anywhere in the site's menus or its footer, which offers a privacy policy and a cookie preferences control and nothing else. No SOC 2 report, ISO/IEC 27001 certificate or independent penetration testing on a stated cadence is named, and the website privacy notice's safeguards section offers reasonable administrative, technical and physical measures with need to know access, which is an assertion with nothing independent behind it.

The absence lands hard here. This company's training corpus includes human genomic and transcriptomic data, its models are its asset rather than a means to one, and it publishes nothing about how either the data or the weights are protected. It is also a Canadian company running a United States facility, so the material crosses a border by construction with no control statement attached.

Deep Genomics makes no device claim for the platform, which is scoped accordingly. The models are not sold as a product, are not presented as a medical device and carry no clearance, which is the appropriate position for what they are.

There is nothing to clear. No candidate is in the clinic, no pipeline is presented anywhere in the company's material, and no investigational application, designation or regulatory milestone is disclosed for anything. That is a change from the September 2023 statement of intent to advance multiple RNA targeted therapeutics, and no announcement accounts for it.

One item counts in its favor. The company contributed to updated industry recommendations on assessing hybridization dependent off target risk for therapeutic oligonucleotides, published in a peer reviewed journal. Helping write the safety guidance for its own modality is a different and stronger signal than complying with it, and it is the closest thing to a regulatory posture the company has.

Model behavior is evaluated in published work rather than asserted, and some of that work is specifically about where predictions fail. Two peer reviewed papers address off target liability directly: Transcriptome-wide off-target effects of steric-blocking oligonucleotides, and Assessing Hybridization-Dependent Off-Target Risk for Therapeutic Oligonucleotides, published as updated industry recommendations, which puts the company's method in front of its own field as guidance rather than as a claim. A third, in Frontiers in Molecular Biosciences, is a curated census of pathogenic and likely pathogenic untranslated region variants together with an evaluation of deep learning models for variant effect prediction, which is an evaluation of the model class this platform belongs to.

The REPRESS preprint reports the platform's own comparative results, stating that the model outperforms other advanced methods and neural architectures across a suite of orthogonal tasks that are named individually: identifying genetic variants affecting microRNA binding, predicting out of distribution data from massively parallel reporter assays, and predicting canonical and non canonical repression. Out of distribution performance is the hard boundary question, and it is addressed rather than avoided.

What is absent is the breakdown and the outside check. No results are published by tissue, cell type, gene family or disease context, so nothing shows where the models are weaker across the groups the platform serves. There is no model card, no independent audit of model behavior and no governance framework document, and the comparative claims for the commercial models sit in preprints the company wrote rather than in reviewed venues.

Mechanisms exist that let a prediction be challenged, and they are described in engineering detail. The platform runs lab in the loop workflows in which predictions are validated using orthogonal assays, including functional assays, with results fed back into training, across two experimental facilities totaling more than 10,000 square feet. The company states a commitment to reliability, reproducibility and correctness as a system property, names experiment tracking integration and model evaluation frameworks inside the BigRNA workflow, and has published off target risk assessment methodology as industry guidance, so the way it checks its own predictions is inspectable.

Nothing stands behind the output. No accuracy, success or error rate with a method and a denominator is published for any model on the company's own site; the comparative statements about REPRESS name the tasks but not the numbers, and the figures sit in the preprints. No guarantee, indemnity or remediation obligation is published, and with no partner, customer or agreement disclosed anywhere, there is no counterparty relationship on the public record for a commitment to run toward in the first place.

Because the platform predicts RNA biology and designs genetic medicines rather than serving health systems, no clinical record integration exists: there is no electronic health record touchpoint, no clinical workflow surface and no provider connection.

On the research side, among the 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, no interoperability standard is claimed, and no connector or application programming interface is offered. The one external research tool named, an experiment tracking integration inside the BigRNA workflow, is part of the company's own machine learning stack rather than a connection a partner could use. With no partner named anywhere, there is also no deployment against which an integration claim could be checked.

The platform runs for Deep Genomics rather than being deployed to anyone, so there is no tenancy model, hosting option or isolation commitment to describe. Location is implied rather than committed: corporate headquarters on University Avenue in Toronto, a second office at Canal Park in Cambridge, Massachusetts, and two experimental facilities totaling more than 10,000 square feet across those two cities.

Two facts sit alongside that. The company names a third party experiment tracking platform integrated into its model workflows, which is a processor a partner would want listed. And this is a Canadian headquartered company operating a United States facility and training on human derived genomic data, which raises a cross border data question with real regulatory content behind it. Nothing published touches it: no cloud or compute provider is named, no residency or transfer position is stated, and no retention or segregation statement covers partner data, of which none is acknowledged in any case.

Nothing commercial is public ahead of a sales conversation. The company states that its platform advances drug discovery problems both internally and in support of major pharmaceutical partners, and it names none of them. There is no collaboration announcement, no deal value, no upfront, no milestone structure, no licensing arrangement and no customer of any kind on the site or in the news archive. Neither is there a funding figure, an investor list, a pricing page nor a pipeline.

The company is not new and not small: it was founded in 2015, employs roughly 100 people, runs two laboratory facilities and is led by a chief executive with three decades of pharmaceutical experience. So the silence is a choice rather than a stage.

One change compounds it. In September 2023 the company announced a strategic plan to become a forward integrated biopharmaceutical company advancing multiple RNA targeted therapeutics, which implies a pipeline and a development model. Nothing in the company's current material reflects that plan, and the news archive carries no announcement explaining the change. A reader cannot establish what this company sells, to whom, or on what terms.

Coverage is claimed broadly without specifics. The company describes itself as a fully integrated AI driven drug discovery company advancing a range of drug discovery problems internally and in support of major pharmaceutical partners, and states the domain as genetic medicines and RNA therapeutics. It names no therapeutic area, no indication, no target and no program anywhere in its published material: there is no pipeline page, no therapeutics section and no development stage disclosed for anything.

What the material shows is capability scope rather than operating scope. The models address variant effect prediction, tissue specific regulation, microRNA binding and messenger RNA degradation, on and off target effects of genetic medicines, and ADAR mediated RNA editing. Diseases appear only inside the publication list, in studies of Wilson disease, autism and splicing defects, which are scientific results rather than statements of where the platform is meant to operate.

This is a change rather than a constant. In September 2023 the company announced a strategic plan to become a forward integrated biopharmaceutical company advancing multiple RNA targeted therapeutics, which would have implied a named pipeline. Nothing the company currently publishes reflects that plan, and no announcement explaining the change appears in its news archive.