Iktos

Generative deep learning is the product across all three named systems. Makya generates de novo molecules under multi parametric optimization constraints, Spaya performs retrosynthesis to convert target compounds into commercially available starting materials, and Ilaka orchestrates the resulting workflow.

The design claim is specific rather than generic: the generative model is trained on millions of reactions and builds synthetic accessibility into generation by working from commercial building blocks and organic reactions, rather than generating freely and filtering for feasibility afterwards. The robotics layer exists to execute what the models design.

Meaningful autonomy exists at the execution layer rather than the decision layer. Ilaka is described as taking over the workflow from ordering raw materials through directing synthesis campaigns and robotic chemistry, which is genuine hands off operation, but the objects being automated are laboratory tasks rather than decisions affecting a patient. Design direction stays with chemists, and the platform's stated positioning as built by chemists for chemists reinforces that.

Held at B because no confidence thresholds, failure rates or documented human approval gates were located, so the boundary between what the system decides and what a chemist decides is described rather than specified.

More specific than marketing but short of independently verifiable. Disclosed detail includes the generative deep learning basis, training on millions of reactions, an explored chemical space stated at the order of ten to the twenty seventh, synthesis constraints imposed during generation, and published descriptions of newer generative models that design molecules and their three dimensional poses directly from a known active ligand without docking, rigid superposition or conformer enumeration.

Held at B because no peer reviewed methods paper establishing core platform performance was located in this review and no model cards or benchmark documentation were found, so the accuracy claims cannot be independently checked.

The stewardship question that matters in this lane is protection of customer chemistry rather than patient data, and it is answered here architecturally rather than contractually, which is the more durable form. A structural arrangement that keeps a partner's compounds and targets out of a shared estate holds regardless of what a later contract says or who owns the company next, where a confidentiality clause depends on both.

That is worth crediting because most of this category answers the same question with a clause and nothing else. The clinical form of this axis does not reach the workflow, which operates on chemical structures, reaction data and assay results with no patient data involved, and that is a fact about the product rather than a claim about controls. What is absent is enumeration and the residual this index now applies to every architectural separation claim.

No hosting arrangement, sub processor list or retention position was located, and nothing states whether anything learned from one partner's chemistry, including model parameters or general representations rather than the compounds themselves, informs the system another partner uses. Keeping compounds separate and keeping learning separate are different commitments. Ask which is guaranteed, how the separation is implemented, and what happens to partner derived artefacts at the end of an engagement.

Adoption is substantial and evidence is thin, which is precisely the pattern this index grades down. The company reports validation through more than 50 and by some accounts more than 60 academic and industrial collaborations, with named partners including Pfizer, Janssen, Merck KGaA, Servier, Sanofi, Ono, Teijin, Galapagos, Chiesi and UCB. That is real commercial traction and a meaningful signal that sophisticated medicinal chemistry organizations find the tools useful.

It is not evidence of benefit. No published outcome study, controlled comparison against conventional design, or disclosed clinical stage asset attributable to the platform was located, and the claim that the integrated approach shortens discovery to under two years is vendor stated. Applying the index precedent that scale of use does not substitute for evidence, this sits at C until an outcome is published.

Not applicable in the provider sense and rated accordingly rather than penalized. The platform operates on chemical structures, reaction data and assay results, with no patient data in the workflow. The stewardship question that does matter here is protection of customer chemistry, and unusually for this category it is answered architecturally rather than contractually.

Not applicable. Customers are pharmaceutical and biotechnology medicinal chemistry organizations licensing design software or entering discovery collaborations, not covered entities transferring protected health information.

No SOC 2, ISO 27001 or equivalent attestation was located and no trust centre was found, across two differently phrased searches. For a French company selling licensed software into European and Asian pharmaceutical organisations, a data processing agreement will exist contractually, but nothing about the security programme is published.

The mitigating factor is architectural rather than attested, and it is now well corroborated. The design software is offered three ways: as a hosted service, installed on customer premises, or deployed inside the customer's own virtual private cloud, and it can also be driven as a Python package from a notebook. That is not a marketing claim taken on trust; multiple named customers, including large pharmaceutical companies and a contract research organisation, announced their own deployments describing the same options. A security conscious buyer can therefore remove the vendor from the trust boundary entirely rather than relying on a certification, which is an option the partnership only platforms in this category cannot offer at all.

One consequence is worth stating plainly, because it changes who the missing attestation actually affects. This vendor's customer base sits in two different risk positions. A buyer taking the hosted option is fully exposed to the gap: proprietary target and chemistry data goes into an environment with no published, independently examined controls. A buyer taking the on premise or private cloud option has substituted architecture for assurance, and the attestation matters far less because the vendor never holds the data. Same vendor, same product, two entirely different questions at contract.

So the ask depends on the deployment. Hosted buyers should press for the attestation and its scope. In boundary buyers should press instead on what the software transmits outbound, what telemetry or licensing calls leave their environment, and what the vendor can see during support.

Nothing to assess rather than something assessed poorly. The platform is design software and is not a regulated medical device, correctly, and the company does not maintain a disclosed proprietary clinical pipeline of its own. Regulatory standing for molecules designed with these tools sits with the partner that owns the asset, and partner assets are not publicly attributed to the platform, so no asset level regulatory record is available to grade.

No formal governance framework or model card exists, but a publications search overturns the earlier finding that the central tension here is unaddressed. This company has published on it, in a peer reviewed journal, with code released.

The domain relevant question is whether a generative model constrained toward synthesisable chemistry collapses back toward familiar space, producing molecules that are easy to make and close to what already exists rather than genuinely novel. The company published exactly that experiment. It introduced a synthetic accessibility score derived from full retrosynthetic analysis, compared it against three established scores from the literature, and validated all of them against a binary judgement made by human chemists on a bench of generated molecules. It then compared generator outputs across a range of constraints and conditions, and reported that constraining generation on its score produced more synthesisable solutions with higher diversity rather than less. That is the tension measured rather than asserted, and the direction of the result is not the one a sceptic would assume. The code for the scores and the experiments was released openly.

A second publication carries an explicit limitation statement, which is rarer still. Describing its reaction based generative models, the company states that regio and chemo selectivity are deliberately omitted during generation to accelerate exploration, and that those considerations should be handled by dedicated retrosynthesis tooling instead. That is a plain published account of what the generator does not do and where the user must go for it.

Held at B on the same basis as the strongest peers in this category. There is no governance framework, no model card and no statement of intended and unsuitable use in product documentation, so the disclosure is a byproduct of scientific practice and lives in the literature rather than anywhere a buyer would look. Ask for novelty measured against the training distribution specifically, since diversity within a generated set and distance from prior art are different quantities.

The method is described with a level of technical specificity most of this lane avoids, and one part of it is a negative scope statement of an unusual kind. Disclosed detail includes the generative basis, training on millions of reactions, an explored chemical space stated by order of magnitude, and synthesis constraints imposed during generation rather than filtered afterwards, which is the difference between designing molecules that can be made and designing molecules that must then be triaged for whether anyone could make them.

Newer generative models are described as designing molecules and their three dimensional poses directly from a known active ligand without docking, rigid superposition or conformer enumeration, and naming the steps the method does not require tells a reader with the relevant training exactly where it departs from standard practice and therefore where it might fail differently. That is a more useful disclosure than a capability list. Held at C because none of it is independently checkable.

No peer reviewed methods paper establishing core platform performance was located, no model cards or benchmark documentation were found, and no warranty, indemnity or remediation commitment attaches, so the accuracy claims rest on the company's own account. The synthesis constraint claim is the one most worth testing, since it is directly measurable. Ask what proportion of generated molecules are successfully synthesised on first attempt, and against what baseline.

Not applicable in the provider sense, with no EHR touchpoint or clinical workflow surface. The research equivalent is better than most peers: Makya can be driven through a graphical interface designed for medicinal chemists or operated as a Python package through a Jupyter notebook interface, which allows it to sit inside an existing computational chemistry toolchain rather than beside it.

The best answer in this category to the question that decides many pharmaceutical software purchases, which is where the chemistry goes. Makya is offered as a SaaS platform, for implementation on customer premises, or inside the customer's own virtual private cloud, with Python package access alongside the interface.

For an organization whose undisclosed structures are its most valuable asset, the on premise and private cloud options mean proprietary chemistry never has to leave the customer's control, which is a materially different risk posture from platforms reachable only by sending targets to a vendor. Explicit regional residency commitments were not located, but the deployment choice itself resolves most of what residency terms exist to address.

A genuine asymmetry sits here. Iktos is one of the few companies in this category selling a licensable SaaS product where a published rate card would be both possible and useful, yet no pricing was located, so the gap is a choice rather than a structural inevitability as it is for partnership only peers.

What is disclosed is deal shape: a strategic collaboration with Servier reported at potential total value over 1 billion euros, more than 50 named collaborations, and a strategic investor base including M Ventures and the Debiopharm Innovation Fund. Dual pricing paths, software licence and discovery collaboration, are stated openly, which is more than most disclose.

Broad across therapeutic areas through more than 50 partner programmes spanning oncology, dermatology, immunology and other areas via named collaborators, but deliberately confined on modality to small molecule medicinal chemistry. The design make test integration including proprietary robotics is built for small molecule synthesis specifically. Buyers working in biologics, antibodies or nucleic acid modalities should treat this platform as out of scope rather than adaptable.