The Factory Belongs in the Design

In December 2025, researchers reported designing active enzymes for three distinct catalytic mechanisms with RFdiffusion2, identifying working candidates after experimentally testing fewer than 96 sequences in each case. The authors also noted that the designed enzymes remained less active than native enzymes. Design capability had advanced; performance still needed work. 1

Earlier that year, researchers studying enzymes for PET plastics recycling raised a different concern. Published performance claims were difficult to compare because testing materials and conditions were inconsistent—and often disconnected from industrial operation. 2

Together, these findings expose a gap between what biology can be designed to do and what it can reliably deliver.

Industrial enzymes offer one route into programmable biology: turning a specified function into a repeatable commercial capability. We see a company-building opportunity in connecting molecular design to the experiments and production conditions that make that possible.

At Aldea, we are interested in companies that make biological performance increasingly repeatable—and retain the learning that makes the next batch, process or product better.

Design for the whole process

For an industrial enzyme, we would ask how it handles variable inputs, how long it remains active, how much preparation the material requires, and what it costs to recover a usable product.

The PET researchers make those dependencies explicit. They identify material preparation, temperature, acidity and the amount of plastic in the reactor as important to evaluating performance under commercially relevant conditions. 2

Sometimes the advance is not in the enzyme itself. A separate 2025 study in PNAS showed that reactively mixing certain difficult polyester wastes made them much more accessible to enzymatic breakdown. Changing the material expanded what the biology could accomplish in the tested system. 3

Our read is that the factory should help define the molecule, rather than merely receive it. If material preparation drives cost, the next design target might be an enzyme that tolerates less processing. If recovery is expensive, the priority might be a cleaner product stream rather than a faster reaction.

That is more than a route to better laboratory scores. It is a way to design toward a product customers could not previously obtain at the required cost or quality.

Epoch Biodesign illustrates this approach in a different polymer: nylon 6,6. The company describes a platform combining enzyme design, laboratory screening, process optimization and manufacturing. Its intended output is usable material, not a protein sequence. 4

The commercial test is whether improvements in biology make the process better, while experience with the process improves the next biological design.

The experiment becomes part of the product

The investment case does not rest on feedback being new. It rests on more flexible molecular design and better choices about what to test next.

RFdiffusion2 expands design flexibility by generating enzyme structures around the chemical groups required for a reaction, without prescribing where those groups must sit in the protein sequence. That broadens the possible starting points for experimentation. 1

In August 2026, researchers reported that REAP, a system combining machine learning with robotic experimentation, improved the targeted activity of a P450 enzyme 57-fold over five cycles relative to its starting variant. Experimental results informed subsequent predictions and the choice of variants to test. 5

A different approach targets production conditions. OpenAI and Ginkgo Bioworks reported a 40% reduction in reaction cost per gram for a test protein against the previous best laboratory baseline. The work used model-guided experiments in one cell-free production system; transfer to other proteins and larger-scale operations remained unproven. 6

These are laboratory results, not industrial cost curves. But they demonstrate progress in both selecting biological designs and improving the conditions in which biology performs.

For us, a useful test of programmability is whether a biological capability can increasingly be specified, measured, iterated and reproduced under relevant conditions. The opportunity is to connect those steps so that experimental and operating evidence shapes the next design—not simply validates the last one.

Where a company earns its advantage

Our interest is in businesses that turn this feedback into an advantage customers pay for.

For a product company, that could mean a proprietary enzyme and process that together deliver reliable output at a competitive cost. Performance across real input conditions could create operating knowledge that a rival cannot reproduce simply by accessing the same design tools.

For an infrastructure company, it could mean becoming embedded in a valuable development workflow: connecting design choices, experimental results and production requirements, with the rights to reuse the resulting knowledge. The product must improve the customer’s development economics, not merely coordinate laboratory activity.

A more defensible dataset may not be a catalogue of molecules. It may be a record of which molecules work, under which conditions, and at what cost.

We would look for evidence that this record changes outcomes: fewer failed development cycles, more predictable performance, lower production costs or faster customer qualification.

For a startup, integration alone is not enough. The opportunity is to own a biological capability or application-specific development system that existing providers cannot readily reproduce—and turn it into a product customers value.

That does not require owning every laboratory or factory. Partners can provide capacity. What matters is controlling the distinctive technology, relevant knowledge or customer relationship through which the company captures value.

Two ways for learning to compound

Learning can compound within a product or across products.

A product company may deepen its advantage by making one valuable process more reliable and economical. Better consistency, lower processing costs and expanding customer adoption can support a substantial business without requiring a new application every year.

A platform makes an additional claim: the tools and knowledge from one program make the next program easier. That requires evidence of transfer, not just success repeated through another round of bespoke work.

Both paths are interesting. The evidence should match the claim.

For a product business, we would follow how performance and economics improve with operating experience. For a platform, we would also ask whether development becomes faster or more predictable across programs. A company should not need to call itself a platform to justify an ambitious outcome.

A loop is not automatically a moat

The countercase for a development platform is a laboratory full of one-off projects.

If each new target requires an entirely different test, customer data cannot be reused, and successful programs leave little transferable knowledge, the company may be building a capable research service rather than a compounding platform.

The advantage could also accrue to an established manufacturer with the relevant customer relationships and operating knowledge. Or broadly available models and laboratory providers could make coordination easier to replicate. A startup needs a distinctive capability, not simply a connected workflow.

Owning physical infrastructure does not resolve this. A factory can add capital requirements without creating a learning advantage. The test is whether operating it improves the product and strengthens the economics enough to justify the capital committed.

A differentiated enzyme or design method can still be the central asset. Closed-loop learning gives us another place to look for durable advantage—not permission to assume it exists.

Watch what carries forward

Three things would strengthen our conviction.

First, performance survives relevant conditions. The biology remains useful across the input variation, operating demands and quality requirements that customers actually bring.

Second, the economics survive scale. The measure is the cost of usable, customer-qualified output, including preparation, processing and recovery—not an isolated improvement in laboratory activity.

Third, learning becomes an economic advantage. Within a product, operating experience improves consistency, cost or customer value. Across a platform, one program makes the next faster or more predictable. In either case, the company must retain a meaningful share of that value.

Those are signs that a company is accumulating a capability, not simply completing projects.

The opportunity in programmable biology is larger than proposing molecules. It is making useful biological functions increasingly possible to specify and deliver—and building businesses that improve through the act of doing so.

The lab and the factory need not sit downstream of intelligence. They can be where the company’s intelligence compounds.

Disclosure: Aldea is an investor in Epoch Biodesign.