Lonza and Onava put expression data inside the AI loop. That is the signal.

What happened on October 9

Lonza and Onava announced a research evaluation collaboration using Lonza’s GS Gene Expression System with GS Ori-Go vectors. The announcement says experimental expression data will feed subsequent AI design cycles, connecting computational predictions with biological testing earlier in discovery.

That is the reported plan, not evidence that the collaboration has already achieved a particular clinical or manufacturing outcome. The signal we take from it is the deliberate connection between design and data generation. Expression is being treated as information for the next design cycle.

Our take: the loop is only as useful as the return path

It is easy to draw a cycle with design, build, test and learn around a circle. The difficult part is deciding what returns from testing. A result without its material identity, method or conditions can be hard to interpret outside the experiment that produced it. A larger dataset does not automatically resolve that ambiguity.

From a founder’s perspective, the defensible asset may include the discipline of generating comparable evidence. We would ask whether the next design cycle can distinguish an unfavorable molecular result from an execution issue, a method change or a different experimental context. Those distinctions influence what a model and a scientist should learn.

A result becomes reusable when the next person can tell what it means and where it came from.THE BIOMEDANA LENS

A hypothetical failed expression result is a useful test

Suppose a designed candidate gives an unexpectedly low expression result. One interpretation is that the design needs to change. Another is that the observation reflects a difference in the experiment or material. Before updating a ranking or feeding the observation into another cycle, the team needs enough context to decide what the result represents.

A useful record would connect the candidate and sequence version, construct identity, experimental conditions, measurement method, processing steps and review outcome. It should also preserve uncertainty when the result cannot support a clean interpretation. Discarding every ambiguous result loses information; treating every result as equally reliable can introduce a different problem.

This is our illustrative analysis of a workflow challenge. It is not a description of a failure at either company.

The founder tradeoff: throughput versus comparability

A rapidly growing discovery operation will be tempted to increase experimental throughput first. That can be the right choice, but it should come with a plan for how changes in execution are represented. Otherwise a later team may struggle to tell whether a pattern came from the molecule or from the way the evidence was generated.

We would prioritize a small, consistent set of contextual fields and a versioned account of how results are processed. This is a practical scope choice: capture what is needed to interpret and reuse an experiment, then expand when a specific scientific question requires more detail.

• Link predictions and measurements through stable candidate and material identities.
• Retain raw observations alongside derived metrics and processing versions.
• Mark method changes and exceptions rather than silently combining results.
• Keep the reviewer’s interpretation distinct from the original measurement.
• Define how uncertain or conflicting observations enter the next cycle.

How to evaluate the loop without a vanity metric

The number of proposed molecules and the number of completed experiments describe activity. They do not tell a founder whether the team is making better choices. We would look at how often a new cycle resolves a predefined uncertainty and whether the same evidence can support an independent review.

A practical pilot could choose one recurring selection question and compare decisions made with the existing handoff against decisions made with a connected evidence record. Track review time, missing-context questions and the ability to explain a changed ranking. Any performance claim should be tied to that bounded evaluation rather than generalized to an entire pipeline.

What this means for BioMedAna

We see a strong reason to connect data context, scientific modeling and reviewed decisions across discovery and development. The connection matters even when the models, assays and external partners differ. A team should be able to follow the evidence that informed a choice and understand when that choice needs revisiting.

The October collaboration points to an important question for any AI-native biologics company: what does the organization learn from an experiment, and can it prove where that learning came from? We would build the answer into the workflow before trying to scale the loop.

Why a negative result needs a careful home

A negative result can be scientifically useful without being straightforward training material. The first question is what failed: the intended molecular behavior, the execution of the experiment, or the ability to interpret the observation? These are different outcomes. Combining them under a single failed label can teach the next cycle the wrong lesson.

We would retain the observation and attach a review state rather than silently remove it or treat it as definitive. The team can then distinguish a result considered fit for comparison from one awaiting clarification. If the interpretation changes, the original record should remain available so a later reviewer can understand why.

This is also a product design issue. Scientists need a quick way to flag an exception and provide the small amount of context that explains it. If the only options are accept or discard, the software is asking users to compress a scientific judgment into a category that may be too crude.

What we would pass on to another builder

Before scaling an experimental loop, agree on what an interpretable result looks like. Select a few representative experiments, including an exception and a conflicting observation. Ask a scientist outside the original team to explain what each supports. The missing context in that exercise tells you what the return path must carry.

Version the processing as well as the model. If a metric changes because a transformation or exclusion rule changed, a later comparison should make that visible. Otherwise the apparent improvement in a design round may partly reflect a different way of measuring it. A traceable processing history lets reviewers investigate that possibility.

The commercial question is whether a team can turn new observations into better supported choices repeatedly. We would judge the loop by its ability to resolve an uncertainty and preserve that learning for the next team. More experiments are useful when the organization can explain what it learned from them.

What should return from the experiment?

This is an illustrative evidence contract for a design–test–learn workflow, rather than a description of either company’s internal data model.

The questions behind the headline

Follow the evidence

1. Lonza — Collaboration with Onava for AI-native biologics discovery (October 9, 2026) ↗

Original BioMedAna editorial analysis. The linked sources support the reported developments; examples, frameworks and product implications reflect our perspective.

Explore the underlying concepts

• What is bioprocess data contextualization?
• BioMedAna Twin