Why AI Needs to Be at the Center of mRNA Design
One of the biggest hurdles in mRNA development is the in vitro to in vivo gap; results that look strong in lab models often fail to translate in living systems. This gap arises not because the science is wrong, but because the environments are fundamentally different.
This is quite common. In fact, nearly 90% of drug candidates entering clinical trials eventually fail before approval, often due to lack of efficacy or inconsistent behavior when moving from in vitro to in vivo. In mRNA work, the drop often happens when a clean cell assay readout meets the messy reality of a living system. Better screening helps, but design choices decide a significant portion of the outcome.
AI provides a new way to rethink mRNA design from the start. Instead of testing one variable at a time, it can evaluate many linked factors simultaneously and prioritize design configurations more likely to express in the right tissue, with the desired half-life, and minimal immune activation. To see why this matters, it helps to look at the specific design choices that could cause promising constructs to falter in vivo.
Why mRNA behaves differently in bench experiments vs whole-organism studies
Differences between in vitro and in vivo outcomes often relate to details inside the message, such as UTRs, codon use, cap chemistry, or tail length, which can influence routing, stability, and how the immune system reads the message. These factors are not the only contributors, but they play an important role.
Let’s start with the untranslated regions (UTRs). The 5′ UTR plays a key role in how efficiently ribosomes initiate translation of the mRNA, and its structure can strongly influence overall protein output. The 3′ UTR, on the other hand, affects stability and localization of the transcript, helping to determine how long the message persists in the cell and where it is expressed. Commonly used defaults, such as hemoglobin-derived UTRs, may perform well in cell lines but could lead to unintended expression patterns in vivo.
Codon choice also shapes outcomes. Over-optimizing for “fast” codons may raise output in vitro, yet can cause unintentional drawbacks in vivo where rapid protein synthesis can result in improper protein folding, triggering cellular stress pathways, potentially leading to protein aggregation or degradation and ultimately reducing therapeutic efficacy. Cap chemistry matters as well. Certain cap types that appear suitable in vitro can, in some cases, be read as foreign in vivo, reducing translation efficiency and triggering interferon responses.
What an AI-first design actually means
An AI-first approach treats mRNA design as a multi-dimensional search problem and co-optimizes the parts that drive in vitro performance. It varies UTRs with purpose, paces codons and tests alternative caps and tails.
Start by naming the real goal. Which tissue? Which exposure window? Which dose is acceptable? And how can innate immune activation be kept as low as possible at all times? From there, the software can generate many versions of the same message that honor those constraints. The point is not to guess, it is to explore.
At Officinae Bio, we’ve built our platforms to work this way. Minerva pulls together the day-to-day lab and manufacturing data into one clean place so models can use it. Vulcanus takes a sequence and works out how best to build it, scanning thousands of fragment options across parameters. Mercurius learns from our own data, using knowledge graphs and Bayesian tools, to suggest RNA sequences with better expression and safety. Ulysses then improves how we run the process, applying multi-objective optimization and using adaptive experiments to make results consistent and scalable.
You do not need a large dataset to do this. Models can learn from existing biology and small pilot runs to produce a short list of designs that are more likely to work in a living system.
A practical blueprint for teams
If you wish to build a similar capability in-house, start by drafting the experimental workflow of a typical design and screening campaign. For each step, clearly specify:
- Design input (e.g., UTR variants, codon strategies, cap chemistries).
- Measured output (e.g., expression level, stability, immunogenicity).
Engage your data science team as early as possible to design the data structures that will capture these inputs and outputs consistently.
Next, define your most common application. State the therapeutic target clearly and specify the endpoint measurements upfront.
Work with your statisticians to review historical data. Define what “success” means, and set realistic expectations. For instance, if screening 96 mRNA variants historically yields ~25% improvement in efficacy, avoid setting targets that assume 150% gains from just a handful of tests.
Once the goals and constraints are clear, bring together lab scientists, data scientists, and statisticians to design a closed-loop workflow.
Before going live, run this workflow on historical datasets. Use old experiments to validate the framework, check assumptions, and pressure-test performance. This rehearsal helps identify weak spots before you rely on the system for new designs.
The field is already starting to shift toward a smarter, data-driven design. Just last month, researchers at MIT came up with an AI method to design nanoparticles to deliver RNA vaccines to cells more efficiently. Analysts now project the AI-powered mRNA discovery and development segment to grow at about 24.7% per year from 2025 to 2034.
Through AI, teams should approach development as a controlled design process with predictable rules. Treat mRNA as a design problem with many linked levers. If done so, the high attrition rates that weigh down the industry today may begin to ease, opening a path where more candidates advance, costs fall, and new therapies, including regenerative medicine, reach patients faster.
Subscribe to Where Tech Meets Bio for weekly roundup of most important AI in drug discovery developments.
Topic: AI in Bio
