Thesis
Why an instrument, and why now
The prediction layer is free, and getting freer
Drug and disease scoring across every approved molecule is public and downloadable under a permissive licence. Zero-shot repurposing over thousands of diseases is on GitHub with a paper in a major journal behind it. Proteome-wide variant pathogenicity is integrated into the browsers every geneticist already opens. Interaction-level methods for how a mutation changes a binding energy are being published monthly by academic groups.
The consequence is not that the field is crowded. It is that a score is a commodity, and every improvement in this layer makes a prediction cheaper and a trustworthy threshold more valuable. So we do not intend to sell a score.
Never compete with free on a score.
Be the referee
There are now several interaction-level models and nobody has calibrated any of them to a clinical threshold, or to a specific population. That measurement is the thing a hospital actually needs and the thing no publication currently provides.
We intend to publish it, including the cases where our own numbers lose to a free tool. Stated plainly: our own figures are as unmeasured as everyone else's today. The difference we are proposing is that the measurement gets published.
They answer the previous question
The installed interpretation platforms are good, consolidated, and already inside the buildings we care about. They answer which variant caused the diagnosis. The question immediately after it is what that variant breaks and what might be tried, and that is a different question with a different kind of evidence behind it. The intent is to sit after the report rather than to replace it.
That the thesis works, and what it costs to run as a drug company
The best-funded company running this thesis has taken roughly 115 million dollars and about ten years to reach a single Phase 2 study. That is not a criticism. It is what being the sponsor of record costs, and it is why such a company will address perhaps twenty diseases in its lifetime out of the thousands its own mission statement cites.
The gap between twenty and thousands is the market for an instrument. There is a second lesson in the same place: an off-patent generic in a new indication carries no exclusivity, which is why companies that must own an asset are pushed toward reformulations, combinations and compounds that never launched. An instrument does not have that problem, because the customer keeps the finding.
The gap between twenty and thousands is the market for an instrument.
Four positions that are actually open
Where there is no literature to read
A knowledge graph has no edges for a variant nobody has published. Sequence and structure do not need a prior publication. In a consanguineous population much of the burden is private variants in poorly described conditions, and only the second approach reaches them at all.
Inside the perimeter
A company whose value sits in a proprietary database cannot install it inside a national genomics programme. A model-agnostic layer built on openly licensed weights can, and in that market sovereign delivery is the premium tier rather than a discount.
Joined to what is on the shelf
What is registered and stocked in one country is irrelevant to a global sponsor and decisive for a hospital pharmacist. Mechanism and availability live in separate systems today, and nothing joins them.
The same pipeline on animals
No commercial offering was found selling cross-species target engagement, and animal health carries a fraction of the regulatory drag with decades of labelled data behind it.
Every external number on this site, and where it comes from
Nothing here is our measurement. These are third-party figures, and they are listed so a reader can check them rather than take them on trust. Where a count depends on what you choose to count, that is said rather than smoothed over.
| Claim | Source |
|---|---|
| There are around 7,000 known rare diseases, and roughly 95 percent have no approved treatment. | The commonly cited public figures, as stated in the peer-reviewed literature and echoed in a 2025 US Government Accountability Office report on rare-disease drug development.Narrative review, PMCUS GAO 2025 |
| About 2,500 approved medicines carry a usable chemical structure. | Open drug-database releases. This count depends entirely on what is counted: every marketed package runs to six figures, every approved active ingredient of any kind to roughly 4,500 to 6,500, and approved small molecules with a structure a model can take to roughly 2,200 to 3,000. The last of those is the number quoted.DrugBankWWAD, Front. Pharmacol. 2024 |
| About 90 percent of patients transfer off insulin, and 75 of 81 in an international cohort remained on the tablet alone after a median of ten years. | Bowman et al., Lancet Diabetes & Endocrinology 2018;6:637-46. A published correction to that paper fixed a contributor's name and changed none of the results.Bowman et al. 2018Correction |
| A sovereign hospital deployed a variant-interpretation platform on-premises so genomic data would not leave the country. | King Faisal Specialist Hospital and Research Centre with Saphetor's VarSome, announced September 2022. Somebody else's contract, cited as market evidence and not as a relationship of ours.Announcement, Sept 2022 |
| The prediction layer is already free: drug-disease scores, zero-shot repurposing, and proteome-wide variant pathogenicity. | Every Cure's MATRIX scores, published under a permissive licence; TxGNN, Nature Medicine 2024, with code on GitHub; AlphaMissense, integrated into Ensembl, UniProt and DECIPHER by EMBL-EBI.MATRIXTxGNN, Nat. Med. 2024AlphaMissense at EMBL-EBI |
| The best-funded company running this thesis has taken roughly 115 million dollars and about ten years to reach a single Phase 2 study. | Healx, from its disclosed financing rounds through the 2024 Series C and its published pipeline. Named here because a source a reader cannot follow is not a source.Series C coveragePipeline |
| The figures and the second variant in the walkthrough use solved structures. | Protein Data Bank entries 2HYY for the kinase domain with the drug bound, and 8P0X and 8P0V for the sixteen-subunit complex. Coordinates as deposited.2HYY8P0X8P0V |
Checked on 4 September 2026. Each source is linked. If any of these drifts, the figure on the page changes rather than the source being dropped.
What is not proven yet
The step that makes this worth having, working out how a specific variant changes a specific interaction, is a model head with no public checkpoint. We are training it. Until a known-answer case is marked and published, everything above is a designed position rather than a demonstrated one, and the walkthrough is where the line between the two is drawn.