Isomorphic Labs applies predictive AI to drug design and research partnerships

Isomorphic Labs develops AI systems for drug design and applies them through its own research and pharmaceutical collaborations. Its Isomorphic Labs Drug Design Engine, or IsoDDE, combines predictive capabilities relevant to molecular interactions. The useful reader decision is whether a research partnership can improve a defined discovery problem, with experimental evidence and clear program responsibilities, rather than whether a convincing molecular prediction is already a medicine.

01 / ProductA drug-design engine extends beyond one structure prediction

Isomorphic’s technology page describes a unified engine spanning therapeutic areas and drug modalities, alongside an internal pipeline focused on oncology and immunology. The company was launched from DeepMind. Its May 2026 funding announcement names Alphabet among existing backers. This is a distinct drug-design business, not a separate product subscription within Google’s Gemini offering.

The IsoDDE research update distinguishes three important questions: what interacting molecules look like, how strongly a small molecule may bind, and where a binding pocket may exist. These are related but different outputs. A plausible pose can inform a hypothesis without establishing the strength of binding or the behavior of a candidate in a biological system.

That distinction makes the platform relevant to medicinal and computational chemistry. A program may have a target but limited structural evidence, or chemical series that need prioritization. A predictive system can help decide which uncertainty to investigate next. The value of the prediction depends on whether it changes a useful research decision and survives an appropriate experimental check.

The company’s technical report evaluates selected predictive capabilities on defined benchmark sets. It examines generalization beyond familiar structures as well as affinity and binding-pocket tasks. These are author-reported computational results. They should not be rewritten as a measured increase in clinical success or a guarantee that the engine handles any new target equally well.

02 / AudienceThe relevant reader owns a discovery program and its experimental decisions

A pharmaceutical or biotechnology research team can evaluate Isomorphic when it has a clearly framed target or molecular design problem and the expertise to judge evidence. Computational chemistry alone is not the entire program. Biological context, assay quality, synthesis feasibility and the later development path all shape which designs are worth pursuing.

The current partnership page names collaborations with Novartis, Lilly and Johnson & Johnson. Their scopes are not identical: the page distinguishes small-molecule programs from a multi-target, cross-modality collaboration. Their existence demonstrates a concrete research route, but it does not establish that every prospective partner can obtain the same access or terms.

The Chai Discovery blueprint is useful when comparing biomolecular modeling approaches. The Recursion blueprint provides context for a different discovery organization connecting computation with experimental programs. Compare the actual research job, access route and evidence produced; a broad AI-for-biology label is too vague to determine which approach fits.

03 / WorkflowA proposed collaboration tests whether predictions improve a research decision

Consider a proposed early discovery collaboration around a target for which the team has several competing binding hypotheses. The starting package should state the biological rationale, existing measurements and the decision that new evidence must change. The aim is to distinguish productive hypotheses, not merely to generate a large collection of visually persuasive structures.

Define the uncertainty before requesting a prediction

Separate uncertainty about the binding location from uncertainty about the pose and expected interaction strength. A model output addressing one question should not silently answer the others. The program team would agree what constitutes useful evidence, how uncertain outputs will be handled and which experimental observations can discriminate between the remaining explanations.

Reserve a suitable set of observations for evaluation rather than revealing every known result during model adaptation. Record how similar the evaluation cases are to the material used in development. This matters because a model can appear strong on familiar chemical space while offering less help on the novel systems that motivated the collaboration.

Connect computational ranking to experimental learning

In the proposed program, chemists would review prioritized candidates for feasibility and scientific relevance before selecting a bounded experimental set. Preserve why each candidate was chosen, including whether it tests a promising design or reduces uncertainty. An informative negative result can change the next iteration even if it does not advance a lead compound.

Evaluate the model and the research process separately. Did the predicted ordering agree with the selected measurements, and did acting on that ordering improve the program’s next decision? A useful structure can still fail to produce a worthwhile candidate. Conversely, discovering that a hypothesis is weak can be a valuable outcome when it prevents a much larger commitment.

This workflow is an editorial proposal, not an account of using IsoDDE or a laboratory protocol. Isomorphic’s published materials do not give this desk access to its private models, partner datasets or program records. Any real collaboration would need its own scientific review, experimental plan and responsibilities before relying on predictions.

04 / PricingPartnership economics are different from an API price list

The January 2024 collaboration announcement describes upfront and milestone payments for the initial Lilly and Novartis agreements. These are historical negotiated transactions, not current list prices for using IsoDDE. A prospective partner should frame commercial discussions around program scope, rights and responsibilities rather than divide an announced deal value by an assumed number of predictions.

Potential milestone totals are conditional on future events. They should not be presented as cash already received, a standard price or a promise that programs will succeed. Likewise, financing supports a company’s activities but is not evidence that a particular candidate has cleared a development stage. Keeping these categories separate makes the business model easier to understand.

A meaningful collaboration proposal should identify the output being purchased or jointly developed. A set of predictions, a research program and rights to a potential medicine have very different economic boundaries. Ask how background data, newly generated observations and resulting inventions will be handled; this article does not infer confidential terms from the public announcements.

05 / DistinctionsThe difficult step is generalization to the program that matters

IsoDDE’s research framing emphasizes systems unlike the training examples. That is a relevant distinction for discovery, where the reason to use a new method may be that familiar approaches have limited evidence. A prospective collaborator should connect this generalization claim to the novelty of its own target and chemistry, rather than rely on one headline comparison.

Binding-pocket identification also differs from optimizing a molecule around a known site. The former can broaden the hypotheses worth considering; the latter works within a more defined starting point. The evidence required to move forward changes accordingly. An unexpected pocket prediction is an invitation to investigate, not proof that it creates a clinically useful intervention.

The platform’s broader ambition is to connect predictions inside a drug-design organization. That offers a different research relationship from obtaining a downloadable model and building the surrounding workflow internally. The tradeoff is between access to an integrated collaboration and the independence, transparency and operational responsibility of a self-managed modeling approach.

06 / QuestionsRead computational results within their actual scope

Does better structure prediction mean a better medicine?

It can contribute useful evidence but does not establish the full outcome. Molecular interactions are one part of a much longer research and development process. This blueprint makes no claim that IsoDDE predictions alone demonstrate safety, effectiveness or approval for any candidate.

Is IsoDDE the same access route as AlphaFold?

No such equivalence is established here. The research update describes work beyond AlphaFold 3, while Isomorphic’s public commercial material describes collaborations. Access to another model or public resource does not confer access to the proprietary engine or a partner program.

What evidence would strengthen a new partnership decision?

A relevant evaluation with documented data boundaries, experimental checks and clear failure analysis would be more useful than an unqualified benchmark rank. The parties should agree what the experiment can establish before interpreting its results as support for a larger program.

07 / DecisionChoose the collaboration around a specific scientific uncertainty

Isomorphic Labs is a prominent AI drug-design company with a substantive predictive research program and named pharmaceutical collaborations. The practical next step is a well-scoped scientific discussion. Keep the capability claim, the experimental evidence and the eventual therapeutic outcome distinct throughout that discussion.