Two coding sequences can specify exactly the same protein and still produce very different amounts of it. For a therapeutic antibody, that difference can affect manufacturing cost. For an mRNA therapy, it can affect whether a practical dose is achievable.

Finding a better sequence is difficult for two reasons. There are an enormous number of ways to encode a protein, and expression depends on interactions that are hard to capture with a short list of rules. AI offers a new way to search this space, but its value depends on two distinct questions: How should we evaluate a sequence? And how much biological context should the model see?

From proxy metrics to learning from sequence

Traditional gene expression optimizers use measurable properties such as codon frequency, GC content, repeated motifs, and predicted RNA structure. These metrics are useful because each captures a feature that can affect expression.

But each is a proxy, not the outcome itself. A sequence can score well against selected metrics and still perform poorly in the lab. Its features interact, and we cannot always specify those interactions in advance.

Generative models offer a different approach: they can learn patterns and relationships from biological sequences rather than relying exclusively on a predefined set of metrics. This is why AI is relevant to gene expression optimization. It gives us a way to explore a vast design space without assuming that the rules we already know describe every route to higher expression.

That does not make proxy metrics useless, or make experimental testing optional. It changes what the optimizer has an opportunity to learn. The lab remains the test of whether a design works.

A coding sequence is only part of the transcript

The second question is separate from the first. Even a model that learns effectively from sequence is limited by the sequence it sees.

Many optimizers treat the coding sequence as a self-contained design problem. The cell, however, encounters it as part of a larger mRNA transcript. The 5′ untranslated region (5′ UTR) helps shape how translation begins; the 3′ UTR can affect stability and localization. Interactions can cross the boundary between these regions and the coding sequence.

For drug developers, this does not necessarily mean redesigning the whole transcript. Many have established expression platforms with UTRs they want to keep. The useful question is therefore often not "What is the best coding sequence for this protein?" but "What is the best coding sequence for this protein within this transcript?"

Testing the importance of context

This distinction shaped the development of ConvergeGEO.

GEOv1 was trained on coding sequences. Because it had not learned their relationship with a transcript's 5′ UTR, we added a module that modifies the opening codons after generation to better match that region. In our experience, this correction often improves expression for GEOv1 designs.

GEOv2 takes a different approach: it was trained on whole transcripts, so it can design a coding sequence while accounting for the surrounding regions. That led to a prediction. If GEOv2 already accounts for the relationship between the 5′ UTR and the start of the coding sequence, applying GEOv1's correction afterwards should no longer help.

We tested that prediction with two proteins. For each protein, we generated four GEOv2 transcript designs. We expressed each design as generated and compared it head to head with a copy modified by GEOv1's module: eight paired comparisons in total. In six of the eight, the modification reduced purified yield. The median change was a 10.4% decrease.

This was a small experiment using transient expression in one host. It does not establish how GEOv2 will perform across every protein or production setting, nor does it reveal exactly what the model learned. It does offer early evidence consistent with our prediction: a correction that helps when the coding sequence is designed largely on its own can hinder a design made with transcript context in view.

Two shifts, two problems

The two shifts address different problems. Generative AI changes how we search for better sequences, moving beyond a fixed set of proxy metrics. Whole-transcript modeling changes what relationships the model can learn. Both choices matter if the goal is higher expression in the biological system where a sequence will actually be used.

The cell does not encounter a coding sequence in isolation. We should not design one in isolation, either.