Stanford’s Rhiju Das explains AI’s RNA-design gains—and the five-year drug gap
In a new interview, the biochemist describes experimentally observed molecular structures and the design problems still separating them from future treatments.
Loading page…
In a new interview, the biochemist describes experimentally observed molecular structures and the design problems still separating them from future treatments.
Listen to this story
Results from an Eterna RNA-design challenge suggest AI can produce pseudoknot folds without researchers first specifying or accurately predicting their full 3D structures: across 57 targets, AI teams eventually matched experienced players on most blind challenges. The evidence is preliminary: four AI designs went to cryo-electron microscopy, while four others were excluded after showing aggregation or unfolding, and only three were resolved as individual molecules. Stanford biochemist Rhiju Das sees potential therapeutic benefits in RNA stability and programmability, but estimates AI-designed RNA drugs are at least five years away.
RNet, trained on earlier chemical-mapping measurements, helped guide designs whose successful folds relied on interactions the modelers had not explicitly specified.
The approach designed patterns of paired and unpaired RNA building blocks, then used chemical mapping and other experiments to assess the resulting structures.
Das expects the RNA-and-AI community could produce proofs of concept for more design problems over the next year.
AI-designed RNA molecules formed intricate, ordered shapes that Stanford biochemist Rhiju Das says did not match known natural folds. In a November 17 Stanford Medicine interview, he explained why those results could open new paths for medicines—but estimated that AI-designed RNA therapeutics remain at least five years away, with foundational design problems still unresolved.
Das’s account concerns an RNA-design challenge run through Eterna, a community science platform that presents molecular problems as a video game. Experienced players competed against AI methods, with physical experiments used to assess whether their proposed molecules took the intended forms. The underlying OpenKnot research tested 57 structural targets.
The challenge focused on pseudoknots: RNA structures in which parts of a loop pair with sections outside that loop. These connections can support complex three-dimensional shapes. Designing new ones has been difficult because tools that predict RNA structure have limited accuracy, especially for synthetic molecules without related natural examples to guide them.
Rather than specifying every detail of a finished 3D object, the researchers designed patterns of paired and unpaired RNA building blocks. Chemical mapping then tested those patterns, identifying which positions were flexible and which appeared protected by pairing. Additional mutation experiments and cryo-electron microscopy helped assess the resulting structures.
The study found that successful AI-generated designs formed ordered 3D folds held together by interactions that had not been explicitly modeled during design. A model called RNet, trained on earlier chemical-mapping measurements, helped guide success. The authors suggest that some difficult RNA-design tasks may therefore be solvable before researchers fully solve RNA 3D-structure prediction.
The AI teams did poorly at first, Das said. After further rounds, their results collectively matched those of experienced Eterna participants. That is a narrower result than saying every AI method beat human designers: the study describes AI matching experienced people on most blind challenges.
Das said all four AI designs taken to cryo-electron microscopy could be visualized, with some showing highly ordered folds unlike known natural ones. The study’s experimental details add an important selection limit: four other designs did not advance to microscopy because purification indicated aggregation or unfolding. Three AI designs were resolved as individual molecules.
These RNA folds were elaborate, intricate interactions that I thought would take billions of years of evolution. And they came for free.
Rhiju Das, speaking to Stanford Medicine
For Das, folding is tied to a practical drug-development problem: stability. He described working with pharmaceutical companies whose RNA vaccines needed frozen storage, complicating shipping. His idea was to make RNA fold tightly, drawing on the stability of folded RNA in nature. In the interview, he said folded RNA medicines have a longer shelf life.
He also sees possibilities beyond storage. More sophisticated RNA structures might eventually sense which kind of cell they are in, process molecular signals, or control when treatment switches on and off. He presented these as potential functions of future therapeutics, not capabilities demonstrated by the challenge molecules.
The immediate work remains foundational. Das believes the Eterna and wider RNA-and-AI communities can solve several remaining design problems over the next year, producing proofs of concept for further development. He is also inviting nonscientists to participate through Eterna; the platform’s game-based approach remains part of that research effort.
Loading discussion...
Join the conversation
Explain where you would draw the line.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.