One year ago, we published GEMORNA in Science.
At the time, GEMORNA demonstrated a new possibility for RNA therapeutics: instead of relying on large experimental screening campaigns and iterative sequence optimization, generative AI could directly design novel mRNA sequences with substantially improved performance. Paper link: www.science.org/doi/10.1126/science.adr8470
A year later, GEMORNA has moved far beyond the original publication.
Today, it has become a growing platform that we are using with pharmaceutical companies and leading research institutions to design mRNA medicines across a broad range of therapeutic applications.
From a scientific platform to more than 30 therapeutic programs
Over the past year, we have established 20+ collaborations and pilot studies with pharmaceutical companies and world-renowned research institutions, including more than five of the world's top 20 pharma.
GEMORNA has now been applied to 30+ mRNA therapeutic programs spanning:
- in vivo CAR-T
- neoantigen cancer vaccines
- prophylactic and therapeutic vaccines
- gene editing
- immune-cell reprogramming
- gene replacement and protein replacement therapeutics
What has been particularly exciting is that the same fundamental platform can generalize across very different therapeutic problems.
That is central to the idea behind GEMORNA.
We do not believe every new mRNA program should require starting another large sequence-optimization campaign from scratch. We believe the model should already understand enough about RNA biology to begin with a much stronger design.
Data-driven, zero-shot, end-to-end
GEMORNA is built as a data-driven, zero-shot, end-to-end mRNA therapeutic design platform.
In multiple programs, GEMORNA has generated full-length mRNAs that outperform industry-leading benchmarks, and even benchmarks derived from marketed drugs, without requiring target-specific experimental data.
This is important because conventional mRNA optimization can require repeated rounds of sequence design, synthesis, screening and redesign.
Our goal is to compress that process dramatically. With GEMORNA, we are increasingly seeing programs where a small number of AI-designed sequences can move directly into experimental validation, potentially reducing what was previously a multi-round optimization process to a single design cycle completed within weeks.
These designs have now been validated in mice across multiple therapeutic programs, both internally and with our partners.
What zero-shot design looks like in practice
One recent collaboration illustrates this particularly well.
Working with one of the world's top 10 pharma, we applied GEMORNA in a zero-shot setting, without using target-specific training data. The partner already had a highly optimized internal benchmark that represented approximately a 12-fold improvement over conventional sequence design.
96% of GEMORNA-designed sequences outperformed the partner's leading internal benchmark in the cell line of interest.
The strongest candidates were then advanced into animal studies.
All three in vitro hits translated in vivo, demonstrating up to an 8-fold improvement in mice.
For us, this is one of the clearest demonstrations of what the next generation of computational drug design can look like. Not AI assisting a large screening campaign. Not an optimization model requiring hundreds or thousands of target-specific measurements. But a model that can enter a new therapeutic problem with no target-specific training data, generate a small number of candidates, and produce designs capable of outperforming an already highly optimized industry benchmark.
Closing the gap between in vitro and in vivo
One of the most persistent challenges in therapeutic development is translation. A sequence can perform extremely well in a cell-based assay and fail once it reaches an animal.
Across multiple therapeutic programs, however, GEMORNA-designed sequences have demonstrated 95%+ translation rate of in vitro hits to in vivo performance.
This gives us increasing confidence that the platform is not merely optimizing an assay. It is learning sequence features that remain relevant as programs move toward more biologically complex systems. For mRNA developers, that matters enormously. The real objective is not to win an in vitro experiment. The objective is to design a molecule that continues to perform when it reaches the next stage of development.
The GEMORNA family is becoming a model ecosystem
GEMORNA is never a single model. The platform has grown into a family of more than 20 generative and predictive models, forming an algorithmic matrix designed to address different components of mRNA therapeutic development for high expression, longer duration, cell-specific expression, circular mRNAs and all with high manufacturability.
The generative models create novel RNA sequences. The predictive models help us understand how those sequences are likely to behave before we synthesize them. We are developing predictive models for gene expression validated in both cells and mice, extending from reporter proteins to therapeutical proteins. Together, these capabilities enable something increasingly powerful:
generate in silico, predict in silico, experimentally validate only the strongest candidates.
The long-term objective is to make computational screening increasingly representative of the biological results that matter in drug development.
One year later
The Science publication was an important scientific validation. But it was never our destination.
Over the past year, GEMORNA has evolved from a research breakthrough into a platform being tested across dozens of therapeutic programs with pharmaceutical and academic partners.
The next challenge is even more exciting:
Can we continue improving the models until mRNA sequence optimization becomes increasingly computational?
Can we predict in vivo performance of full-length mRNA therapeutics accurately enough that many traditional screening experiments become unnecessary?
And ultimately, can better sequence design enable medicines that would otherwise never reach patients?
That is what we are building toward.