Insilico Medicine

AI is the mechanism end to end and is the company's actual claim to novelty: PandaOmics prioritized the target, Chemistry42 generated the molecule using generative tensorial reinforcement learning rather than screening existing libraries, and inClinico informed trial design. Rentosertib is the proof: a drug whose target, structure, and development were each AI driven.

The oversight here is real, external, and the heaviest that exists, but it does not cover every product equally and the uncovered part is the one worth asking about.

For the discovery modules the position is clear. Target rankings are hypotheses a research team selects from. Generated molecules must be synthesised and tested before they mean anything. The company's flagship programme is the demonstration: a target identified by the platform and a molecule generated by it proceeded through preclinical work, investigational new drug clearance, and human clinical trials now at Phase III, with peer reviewed publication at two stages. Every consequential output was checked by synthesis, by regulators and by patients in controlled studies. No vendor description of an internal review process would be worth as much.

The trial outcome prediction module is a different problem and the difference is important. Its output is used to decide whether a programme proceeds. A prediction that a trial will succeed is eventually checked by running the trial. A prediction that a trial will fail is usually not checked at all, because the programme is stopped and no counterfactual is ever generated. The error that matters most is therefore the one the world never observes, and no external mechanism corrects it. Ask what confidence the module expresses, on what evidence it was calibrated, and what governance the customer applies before a negative prediction is allowed to end a programme.

Held at B rather than A because nothing describes in product oversight: what uncertainty a user sees on a ranking or a prediction, what the system declines to answer, and what a reviewer is shown before acting.

Unusually specific for this index: named engines with disclosed methods (generative tensorial reinforcement learning in Chemistry42), and medicinal chemistry disclosed in peer reviewed literature. Held back from A because no model cards or technical documentation for the platform modules themselves were retrieved.

One genuine control exists and the surrounding position is not disclosed, and a competitive consideration belongs on this axis rather than being treated as commercial. The control is an offline appliance. Target discovery is sold as a self contained device running with no internet connection and chip level confidential computing, marketed precisely at hospitals, institutes and companies holding data they are unwilling or unable to place in a cloud.

For a customer taking that route the question largely dissolves, because nothing is transferred, and it is a structural answer rather than a policy one. For the hosted platform nothing equivalent was located. Customers of the target discovery module upload their own omics data, which in a hospital or translational setting is patient derived, and no statement was found on whether uploaded omics or chemical data improves the vendor's models, whether it is segregated per customer, how long it is retained or what happens at termination.

The published product policy addresses users' personal data rather than research data submitted for analysis, and permits sharing with affiliated companies including for those affiliates' own marketing, which is a broad onward permission to sit above sensitive research material. The competitive point: this vendor runs its own drug pipeline in the therapeutic areas its customers work in, so what happens to uploaded targets and chemistry is a competitive question as well as a privacy one. Put the answer in the agreement, not in a policy the vendor can amend.

Pipeline progress is the evidence that matters in this category, and it is peer reviewed. Rentosertib discovery was published in Nature Biotechnology and Phase IIa results in Nature Medicine (reported mean FVC change of +98.4 mL at 12 weeks in the 60 mg arm against -20.3 mL for placebo), and the program entered a 320 patient Phase III in July 2026. Thirteen programs have reached IND clearance. Platform speed claims (12 to 18 months to preclinical candidate, 60 to 200 molecules synthesized per program) remain vendor reported.

One genuine control exists and it is a good one. The surrounding position is not disclosed.

The control is the offline appliance. The company sells target discovery as a self contained device that runs with no internet connection and chip level confidential computing, and it markets it precisely at hospitals, institutes and companies holding data they are unwilling or unable to place in a cloud. Its chief executive has framed the rationale in those terms publicly. For a customer that takes this route the stewardship question largely dissolves, because nothing is transferred.

For the hosted platform nothing equivalent was located. Customers of the target discovery module upload their own omics data, which in a hospital or translational research setting is patient derived. No statement was found on whether customer uploaded omics or chemical data is used to improve the vendor's models, whether it is segregated per customer, how long it is retained, or what happens to it at termination. The published product privacy policy addresses personal data of users rather than the research data submitted for analysis, and it permits sharing with affiliated companies including for those affiliates' own marketing purposes, which is a broad onward permission to sit above sensitive research material.

One further consideration belongs on this axis rather than being treated as commercial. This vendor runs its own drug pipeline in the therapeutic areas its customers work in. What happens to a customer's uploaded targets and chemistry is therefore a competitive question as well as a privacy one, and the answer belongs in the agreement rather than in a policy the vendor can amend.

Not a scoping determination, and this is the distinction that matters on this record. Most vendors in this category are excused from this axis because they work only with chemistry and never touch patient data. This one markets a target discovery product directly at hospitals and research institutes on the strength of their holding large amounts of their own data, including personal medical data. The regime therefore genuinely reaches a plausible deployment, and nothing addresses it.

No business associate agreement, no availability statement, and no characterisation of the company's role under the United States health privacy rule was located. The published product privacy policy addresses personal data of platform users and names data protection concepts drawn from other regimes.

The offline appliance is a real structural answer for one class of buyer and should be recognised as such. Where an institution runs the analysis on its own premises with no data leaving, there is no disclosure to a vendor and the business associate framework is not engaged in the ordinary way. That is a legitimate route and it is the one a covered entity should prefer here on the present disclosure.

For the hosted route the position is unresolved and a United States covered entity should treat it as an open question. Note in particular that the named processing locations for the hosted chemistry module are outside the United States. Offshore processing is not prohibited, but a covered entity remains responsible for its business associate arrangements wherever processing occurs, and practical enforcement against an offshore processor is a real consideration. Establish before any patient derived material is uploaded whether the company will execute a business associate agreement, which entity would sign it, and where the processing would occur.

No SOC 2, ISO 27001 or equivalent attestation was located and no trust centre was found. Two real technical controls are described, and neither is an attestation: chip level confidential computing in the offline appliance, and named processing locations for the hosted chemistry module.

The issuer route deserves recording because it exists here and is worth revisiting rather than concluding on. This company listed on the Main Board of the Hong Kong exchange, and a Hong Kong listed issuer must publish a sustainability report alongside its annual report whose prescribed scope reaches customer data protection. That route has produced a graded disclosure for another vendor in this category. It does not yet produce one here, and the reason is timing: the listing completed on 30 December 2025, two days before the financial year end, so the first reporting period is one during almost all of which the company was private and the obligation did not apply. Nothing located establishes what the first sustainability report says on data protection, if anything.

This is a third variant of a timing problem this category keeps producing, and a distinct one. Two other vendors here are listed but have no annual report yet. This vendor has one, covering a year it spent private. Recheck this row against the annual report and sustainability report for the first full financial year as a listed company, due in 2027, which is where a substantive disclosure would appear if it appears at all.

Note for contracting that the listed entity is a Cayman incorporated holding company registered in Hong Kong, while operations span more than a dozen jurisdictions. Establish which entity signs and which entity holds any assurance that is eventually produced.

Positioning is careful and correct. Rentosertib holds FDA Orphan Drug designation for IPF (2023) and was included in China's CDE Breakthrough Therapy list (2025), and both the company and coverage state plainly that it remains investigational and is not approved by any regulatory authority. No overclaiming.

Publication is extensive and the governance question is nonetheless unanswered. Those are different things and this record is a clean illustration of the difference.

The platform modules are documented in peer reviewed venues, and the flagship programme is published in two Nature portfolio journals covering discovery and Phase IIa results. That is strong technical transparency and it is graded on the transparency axis. None of it characterises where the platform degrades.

The domain relevant bias question here is unusually concrete and unusually pressing, because of how the target discovery module works. It mines published biomedical literature and public omics data. The biomedical literature is heavily concentrated on a small subset of well studied genes, a skew that is itself well documented, and a system that ranks targets by mining that corpus inherits the concentration. The company's own strongest claim, that it found a novel target for a fibrotic disease, is a counterexample offered against exactly this concern, which shows the company understands the question. What is missing is the general answer: no reported rate at which the platform surfaces genuinely under studied targets against well studied ones, no analysis of how ranking behaves where omics coverage is thin, no statement of the disease areas or data regimes where it performs poorly.

Nothing else expected on this axis was located either: no formal governance framework, no model card for any module, and no statement of intended and unsuitable use.

What would move this row is narrow and achievable. Report the distribution of surfaced targets against a measure of prior study, and state where the platform should not be relied upon.

Named engines with disclosed methods put this above the floor. The generative chemistry approach is named specifically rather than described as artificial intelligence, and medicinal chemistry work appears in the peer reviewed literature, so a reader can evaluate the method and the resulting compounds through venues that applied their own standard. That is real disclosure and rarer in this category than the volume of announcements would suggest.

Held at C on a distinction worth drawing carefully, because it recurs across this lane. Peer reviewed medicinal chemistry establishes things about particular molecules: that a compound was made, that it bound what it was supposed to, that it behaved a certain way in a model system. It does not establish platform predictive performance, which is the claim being sold, and the two are separated by exactly the selection effect that makes the first easy to publish and the second hard.

Papers describe programmes that worked. No model cards or technical documentation for the platform modules were retrieved, no prospective success rate against an industry baseline is published, and no warranty, indemnity or remediation commitment was located. Ask for the platform's prospective hit rate with a denominator, how many programmes entered against how many progressed, and what the vendor commits to on a failed prediction.

A scoping determination, and the domain equivalent is answered.

This is discovery and development software. It does not sit in a clinical workflow, does not read or write to a patient record, and no electronic health record integration is claimed or would be meaningful. Grading the absence of an integration the product is not designed to have would misdescribe it.

The equivalent question for a discovery platform is what data it can take in and how it reaches a customer's environment, and on both the position is disclosed. The target discovery module is documented as working across transcriptomics, genomics, epigenomics, proteomics, single cell data and compound information, alongside biomedical text, which is a broad and specifically enumerated ingestion surface rather than a general claim to handle omics. Delivery is available either as hosted software or as a self contained on premise appliance, which means an institution that cannot export data can still use the platform inside its own walls.

Held at B because no programmatic interface documentation, supported file format specification or export path was located. A customer intending to wire the platform into an existing informatics pipeline, rather than working inside the vendor's interface, should confirm that an interface exists and what it accepts.

Among the most specific deployment disclosures in this category, and specific in the way that actually helps a buyer.

Two delivery models are offered and both are documented. The hosted software carries a named residency position tied to named endpoints: data for subscriptions on one hosted service are stated to be stored and processed in Seoul, and data for subscriptions on the other in Ireland. Naming the processing location per endpoint rather than gesturing at regional availability is rare in this index and it is checkable. Note for buyers weighing jurisdictional exposure that the company has substantial mainland China operations and neither named processing location is in mainland China.

The second model is stronger still for anyone who cannot use hosted software at all. The company sells an on premise hardware appliance that runs the target discovery module entirely offline, with no internet connection required and chip level confidential computing, marketed explicitly at hospitals and institutes that hold data they cannot place in a cloud. An air gapped appliance removes the transfer question rather than governing it.

Held at B because the disclosure does not cover the platform. The named residency applies to the generative chemistry module. No equivalent statement was located for target discovery as a hosted service, for trial outcome prediction, or for the newer model offerings. The applicable privacy policy also permits sharing information with affiliated companies, and states those affiliates may use it for their own marketing purposes, which is a broad onward permission sitting alongside an otherwise precise residency commitment. Ask for a per module residency statement and for the affiliate sharing position in the commercial agreement rather than the published policy.

No public pricing. Contact the vendor. The commercial surface is partnership and licensing structure rather than a platform rate card, which is the norm in drug discovery AI.

Broad and clearly stated. The platform spans the discovery chain from target identification through molecular generation to clinical trial outcome prediction, and the company's own pipeline covers fibrosis, oncology, immunology, pain, and obesity and metabolic disorders. It also states extension of the platform beyond human therapeutics into advanced materials, agriculture, nutritional products and veterinary medicine.

Held at B for two reasons a buyer should weigh rather than one.

The non pharmaceutical verticals are named without any evidence of depth. A list of adjacent industries is a statement of ambition until a customer or a result is attached to one.

More substantively, this company is simultaneously a software vendor and a clinical stage drug developer competing in the therapeutic areas it lists. A pharmaceutical customer licensing the generative chemistry or target discovery modules is placing its target selections and chemical series inside a company that runs its own pipeline in overlapping indications. That is a structural feature of the business model rather than an allegation, and the company is open about the dual identity, but it makes scope a commercial question as well as a coverage one. Establish contractually which therapeutic areas are excluded from the vendor's own programmes for the duration, and what firewall separates the platform team from the pipeline team.