When AI Designs the Ingredient: Who Controls the Data, the Model and the Discovery?

AI-enabled discovery has drawn significant attention in pharmaceutical R&D, but its impact is extending into other industries, across ingredients, food, flavour, fragrance and industrial biotechnology. As AI models advance, the key developments are not just new applications of AI, but new innovation models in which different companies contribute different parts of the discovery engine.

One company may bring an AI model; another proprietary data and domain expertise; and both may contribute to experimental validation and commercial scale-up. The collaboration should therefore be viewed not simply as a route to a defined product, but as a discovery engine. It may ultimately deliver a new molecule or ingredient, while also creating new datasets, model improvements, experimental insights and process know-how (DSM-Firmenich, 2026), (Nasdaq, 2026).

That creates an important IP question: when innovation is produced by the interaction of assets, who owns, and at least as importantly, who controls, the value that results?

An effective IP strategy should identify which data, workflows and outputs a company can control, where value may reside, and where exclusivity and defensible commercial value may be obtained or controlled through a combination of patents, trade marks, confidentiality obligations, trade secrets, contractual rights, licensing arrangements, and commercial agreements.

In collaborations combining one company’s proprietary biological or sensory dataset with another company’s foundation model, simple “background IP” and “foreground IP” terminology may not tell the whole story. The collaboration may generate new datasets, fine-tuned models, predictions, experimentally validated relationships and ultimately new commercial outputs including flavour or fragrance molecules, ingredients, functional proteins or enzymes.

These different outputs may require different ownership and exploitation rules. An improved model generated using one party’s data, for example, raises different commercial questions from the final molecule generated by that model: who can use the improvement, for what purpose, and in which fields after the collaboration ends? Once the parties understand what is being created and who needs to control it, the next question is how that control should be made defensible.

Different assets may require different approaches. A final product or technical application may be suitable for patent protection; proprietary datasets, negative experimental results, workflows and optimisation knowledge may derive more value from confidentiality or trade-secret protection; rights to reuse data or model improvements may depend principally on the collaboration agreement.

Before an AI-enabled discovery collaboration begins, four questions can help identify where future IP and commercial value may arise:

A critical question is whether one of these less obvious assets might ultimately prove more valuable than the first product generated by the collaboration. That product could eventually be superseded, while a dataset could power hundreds of subsequent discoveries; an experimental insight could reveal a new class of useful products; a model improvement could enhance every subsequent programme; and optimisation know-how could shorten future development.

Ownership, particularly in relation to registrable assets is important, but considering asset ownership alone may provide an incomplete picture of commercial control. For each strategically important asset, four questions should be asked: who can USE it, who can PREVENT USE by others, who can REUSE it elsewhere, and who can use it to IMPROVE the underlying discovery capability?

Who benefits from what the collaboration learns?
In an example where Company A is an AI discovery business and Company B has data and domain expertise, Company B might quite properly negotiate ownership or exclusivity around an output of a discovery engine. However, if during the project Company A’s model learns what properties work, which predictions fail, which experimental characteristics correlate with commercial performance, even if Company B owns the output, the collaboration may have made Company A’s discovery capability materially better.

Has value generated using Company B’s data, experimental resources and expertise migrated back into a platform that can subsequently be used for Company B’s competitors? That is not necessarily undesirable. If Company B has exclusivity in food, can Company A use related learning in cosmetics or agriculture? Can it work with another food company around a different molecule or application? Should restrictions attach to a molecule, a product family, a functional property or a field of use? And how long should those restrictions last?

An AI company may reasonably need to retain the ability to improve its core technology. The important point is that the parties identify the issue, negotiate and consciously agree the boundaries before the value is created. The appropriate boundary will depend on the commercial bargain: broader exclusivity or tighter restrictions on reuse may protect one party’s advantage, but may also affect the economics of the collaboration where the other party’s business model depends on improving and redeploying its platform.

Many of these questions are easiest to answer before the first collaboration-generated data exists, and considerably harder once the parties can see which outputs have become valuable.

Success in AI-enabled discovery may not simply favour the businesses with the best AI models, the largest datasets or the strongest existing product portfolios. Increasingly, competitive advantage may emerge from combining these assets effectively through collaboration.

The challenge for each participant is to understand not only what it contributes, but what the collaboration will create: products, datasets, model improvements, experimental insights and know-how can all carry commercial value.

The businesses best placed to capture that value will be those that decide early what must be owned, what needs to remain confidential, what may be shared, and, critically, what each participant can reuse once the collaboration ends. In AI-enabled discovery, successful collaborations may not be those that create the desired molecule, ingredient or protein. It may be those that most effectively capture the learning, data and capabilities generated along the way. The central commercial question is therefore not simply who owns the output, but who benefits from everything the discovery engine learns.

This article was prepared by HGF Partner Russell Thom

References

DSM-Firmenich. (2026, September 14). dsm-firmenich partners with Boltz on AI for ingredient discovery. Retrieved from DSM-Firmenich: https://our-company.dsm-firmenich.com/en/our-company/news/trade-news/2026/dsm-firmenich-partners-with-boltz-on-ai-for-ingredient-discovery.html

Nasdaq. (2026, March 18). Ingredion and Shiru Announce Partnership to Accelerate AI-Powered Functional Protein Discovery. Retrieved from Nasdaq: https://www.nasdaq.com/press-release/ingredion-and-shiru-announce-partnership-accelerate-ai-powered-functional-protein