An AI-designed protein delivers RNA better than lipid nanoparticles
Two stories about AI leaving the chat window: one where a generative model designed a delivery vehicle that outperforms the industry's standard carriers in animals, and one where ordinary investors are handing their brokerage accounts to agents and hoping for the best.
Researchers at Helmholtz Munich and the Technical University of Munich have built an RNA delivery vehicle from scratch — and its scaffold was designed by generative AI, not borrowed from nature. The team combined functional protein parts with an AI-generated structural protein, tested more than 100 variants, and found that the ones with shapes that don't occur in the natural world worked best. The winner, STV-C8, delivered RNA into target cells far more efficiently than the virus-like particles and lipid nanoparticles it was tested against, and needed substantially less RNA to produce comparable protein. The result is published in Nature.
The animal work is what makes this more than a cell-culture result. Injected intravenously into mice, STV-C8 produced expression mostly in the lungs, with no immunological or toxic side effects detected. Loaded with CRISPR/Cas9 components and injected into pig muscle, it removed a disease-relevant section of the dystrophin gene — the gene disrupted in Duchenne muscular dystrophy. That is a generative model proposing a structure no evolutionary process arrived at, then having that structure win. The system is still experimental: targeting specific cell types and understanding how the vehicles distribute through the body are unsolved, and the team says it plans to spin the technology out.
The Wall Street Journal reports that retail investors are increasingly giving AI agents discretionary authority over their stock portfolios — describing what they want in plain language and letting the agent build, run, and rebalance the strategy. Moomoo, the overseas brand of Futu, launched agent trading in April. One options trader described running a Codex agent that screens options flow for large institutional prints and scores trades against preset conditions before entering; he told the Journal it amounts to making money without lifting a finger. Robinhood keeps AI-managed portfolios in dedicated separate accounts and sends a notification for every trade.
The bull case is emotional discipline — retail investors have long been mocked as "dumb money" for buying high and panic-selling, and users say an agent removes the second-guessing. The bear case is that the agents may all read the same tea leaves: a National Bureau of Economic Research working paper found that when models were asked to build generic investment strategies, they recommended concentrated, richly valued, heavily covered stocks and did not beat a passive benchmark. Former quant trader Irene Aldridge, who now works on AI and markets, called it a genuine breakthrough for small investors while warning them to accept the risk themselves. Crowded bots piling into the same positions is how the 2007 quant quake started.
What to watch: whether AI-designed carriers like STV-C8 hold up once someone tries to aim them at a cell type on purpose, and whether brokerages start treating agent-built portfolios as a suitability problem rather than a feature.
Would you let an agent trade your money with discretionary authority — and what guardrail would you insist on first? Tell us in the comments.
Sources: Phys.org · Nature (paper) · 科技日报 (Science and Technology Daily) · Wall Street Journal · 21财经 (21 Jingji) · TechNews 科技新報
