Europe’s artificial intelligence race is often told through computing infrastructure and language models. In Vilnius, a different proposition is taking shape: AI can help manufacture the molecular tools that make medicines, diagnostics and industrial chemicals possible.
Lithuanian startup Biomatter closed an oversubscribed $18 million Series A on November 12, led by Astanor and joined by existing investors UVC Partners, Inventure and Metaplanet. More revealing than the funding is the company’s reported delivery of over 50 custom enzymes to more than 25 enterprise customers, including AstraZeneca, BASF, Sanofi, Kirin Holdings and Tereos.
The question is whether AI-designed biology can become a repeatable industrial service rather than a succession of expensive scientific experiments.
Designing the Reaction Before the Protein
Enzymes are protein catalysts: they accelerate chemical reactions without being consumed in the process. Their amino-acid sequences influence how they fold, while the three-dimensional arrangement of residues shapes the sites where molecules bind and react.
Traditional enzyme engineering often starts with a naturally occurring protein and modifies it for greater activity, stability or selectivity. Biomatter says its Intelligent Architecture platform starts instead with the required chemistry, then designs candidate sequences and structures for laboratory testing.
That approach is scientifically difficult. A sequence may appear plausible to a model but fail to fold correctly, remain soluble, catalyse the intended reaction or survive manufacturing conditions. Prediction is not the same as a functional enzyme.
Biomatter’s founders supplied early experimental evidence. In a 2021 Nature Machine Intelligence study, Donatas Repecka, Vykintas Jauniškis, Laurynas Karpus and colleagues developed ProteinGAN, a generative model trained on natural enzyme sequences.
Of 55 designed malate dehydrogenase variants tested, 13, or 24%, were soluble and catalytically active under the study’s laboratory conditions. One functional variant differed by 106 amino-acid substitutions.
The experiment demonstrated that generative design could produce working sequences with substantial changes. It did not establish that every generated sequence would function or that industrial performance was guaranteed.
The Competitive Asset Is the Laboratory Loop
Biomatter’s commercial premise depends on something less conspicuous than an algorithm: experiments. Each designed enzyme is tested, and the results contribute to a proprietary dataset connecting protein sequences with experimentally measured performance.
According to Invest Lithuania, the company says development programs that traditionally take months or years can now take weeks.
Separate research demonstrates why experimental feedback matters.
A 2023 Nature Biotechnology study led by Ali Madani trained the ProGen language model on 280 million protein sequences. Experimentally tested artificial lysozymes showed catalytic efficiencies comparable to natural counterparts, despite substantial sequence differences.
A 2025 Nature Communications study combined language-model predictions with automated construction and testing. Researchers improved a model enzyme’s activity by up to 2.4-fold through four experimental rounds completed within ten days.
Another 2025 study integrated machine learning, language models and laboratory automation, reporting a 90-fold improvement in one enzyme’s substrate preference and a 26-fold activity improvement in a phytase at neutral pH.
Meanwhile, a 2026 enzyme-specific language-model study reported improvements in catalytic-property prediction and experimental enrichment of beneficial variants.
These findings come from independent research teams, not Biomatter’s commercial programmes. Together, they indicate why experimental data and repeated validation are central to AI-assisted enzyme engineering.
From Model Output to Commercial Reagent
One concrete commercial example is Thermo Fisher Scientific’s Invitrogen SuperScript V reverse transcriptase, which converts RNA into complementary DNA during laboratory workflows. Biomatter says it helped improve the enzyme’s resistance to oxidation.
Thermo Fisher’s product documentation confirms that the standalone enzyme supports ambient shipping and room-temperature storage for up to six months.
That property can simplify reagent logistics and reduce dependence on refrigerated distribution, although it does not establish equivalent durability across Biomatter’s other designs.
The company also reports developing a methylase for mRNA vaccine manufacturing that operates without S-adenosylmethionine, or SAM, a molecular donor normally required by this enzyme class.
That could remove a manufacturing constraint, but the company’s announcement does not independently establish performance at industrial production scale.
Why This Matters Beyond Lithuania
Biomatter plans to use its financing for a 2,000-square-metre design-and-testing facility in Vilnius, scheduled to open in 2027. It aims to shorten development cycles from weeks to days and expand its US presence, where roughly half its customers are based. The faster turnaround remains a target.
Competition is emerging across European biotechnology. Amsterdam-based Cradle collaborates with pharmaceutical companies on AI-guided protein engineering, including antibody development with Bayer.
Different business models are forming around predictive software, proprietary experiments and finished biological products.
Biomatter’s challenge is demonstrating repeatable commercial demand for enzymes that maintain their performance beyond laboratory experiments. Its funding announcement does not disclose revenue, margins or project success rates.
For European founders, that is the larger story: competitive AI businesses may be built not only on models, but on specialized datasets, laboratory infrastructure and products that must function in the physical world.
