The Brief
Science & Discovery 4 min read

12 Out of 15: How AI Is Rewriting the Rules of mRNA Delivery

NAVION

Share

Gene therapy has long faced a stubborn paradox. The tools needed to deliver genetic material into human cells, whether viral vectors or synthetic nanoparticles, tend to trigger the very immune system they are trying to work around. The body recognizes foreign components and mounts a defense. This limits how often treatments can be administered, which in turn limits their usefulness for chronic conditions. A study published in Nature Communications by researchers from the Chinese Academy of Sciences, the Chinese University of Hong Kong, Zhejiang Lab, and ShanghaiTech University proposes a different approach: use the human body’s own proteins as the delivery vehicle.

The Problem With Borrowed Parts

To understand why this matters, it helps to understand what a delivery scaffold actually does. In mRNA therapies, the genetic payload needs a carrier to reach the interior of a cell without being destroyed or flagged as a threat. Most current systems rely on lipid nanoparticles or viral components, many of which are non-human in origin. The immune system is remarkably good at distinguishing self from non-self. Non-human components, even when carefully engineered, carry a risk of triggering immune responses that make repeated dosing difficult or dangerous.

The logic behind the new research is straightforward: if the scaffold is built from human proteins, the immune system is less likely to treat it as an intruder. The challenge is identifying which human proteins are actually capable of forming stable particles and delivering mRNA cargo. The human proteome is vast. Searching it manually or through traditional experimental methods would be prohibitively slow.

What DeepDelivery Actually Did

This is where the AI platform, named DeepDelivery by the research team, enters the picture. The system performed a proteome-scale screening, evaluating candidates across the full breadth of human proteins, and nominated 512 candidates with potential as scaffold components. From that pool, 15 were selected for experimental validation. Twelve of those 15 functioned as scaffold components capable of facilitating particle formation and mRNA delivery. That is a success rate that would be difficult to achieve through conventional trial-and-error approaches.

One of the more unexpected findings involved a family of proteins called TRIM proteins, short for tripartite motif proteins. These are typically associated with antiviral immune responses, meaning the body deploys them to fight off viruses. The research found that several TRIM family proteins exhibit strong assembly and delivery activity, making them candidates for the very kind of cargo delivery they would normally work to prevent. This is not a contradiction so much as a demonstration of biological versatility: a protein’s primary role does not exhaust its functional potential.

The team did not stop at identifying candidates. Using model interpretability and perturbation analyses, they identified specific TRIM domain features that influenced the AI’s prediction scores. This allowed them to guide domain engineering for one specific protein, TRIM25, improving its delivery performance. The platform, in other words, was not just a screening tool. It became a guide for engineering.

Why This Signals a Broader Shift

The significance of this research extends beyond any single protein or delivery system. What the DeepDelivery platform represents is a methodological shift in how biological discovery can work. Traditionally, identifying a protein with a useful function required either serendipity or exhaustive experimental work. AI-driven proteome screening compresses that process, allowing researchers to move from a space of hundreds of thousands of candidates to a manageable experimental shortlist with a high success rate.

This has implications for how mRNA therapies might be designed in the future. If human-derived scaffolds can be reliably identified and engineered, the chronic treatment problem becomes more tractable. Patients who need repeated doses of a gene therapy, for conditions that are ongoing rather than one-time, face the greatest risk from immune responses to foreign delivery components. A delivery architecture built from human proteins addresses that risk at its source.

The research also illustrates something important about the relationship between AI and experimental science. The platform did not replace the laboratory work. Twelve proteins still had to be validated through physical experiments. What the AI did was make the experimental work tractable by narrowing the search space and then, through interpretability analysis, providing guidance on how to engineer specific candidates. The scientists directed the inquiry; the AI handled the volume and complexity of the initial search.

Patent applications have been filed based on the human-derived scaffold proteins described in the study, and one of the corresponding authors is a cofounder of a biotechnology company in Shenzhen. The research is moving toward application, not just publication.

In Short

An AI platform screened the human proteome and identified proteins capable of delivering mRNA into cells, with 12 out of 15 experimentally validated candidates succeeding. The approach sidesteps a core limitation of current gene therapy: immune responses to non-human delivery components. By using the body’s own proteins as scaffolds, and by using AI to find and engineer them, the research opens a path toward therapies that can be administered repeatedly without triggering the immune system. The method matters as much as the result.

Based on reporting from Nature Communications.

Written by

NAVION