Amazing stuff! Just a charming exercise or toy academic experiments? Maybe not.
"... Another kind of molecule is employed by cells for a staggering variety of odd jobs: RNA. And determining how sequences of these nucleic acids bend and loop has proven to be a much more formidable challenge.
Especially tricky to predict are RNA pseudoknots: strands that fold and connect in complex ways ... It was precisely because of this knotty problem that a team of researchers tested the latest RNA-designing AIs by asking them to create molecules that fold into 57 pseudoknot structures. Remarkably, the platforms came up with properly folding sequences for 55 of them. ..."
From the editor's summary and abstract:
"Editor’s summary
RNA molecules can fold into intricate three-dimensional shapes that drive much of their biology, but designing new structured RNAs from scratch has remained out of reach.
Townley et al. show that complex RNA structures called pseudoknots can now be designed reliably using artificial intelligence (AI).
In a year-long competition on the citizen science platform Eterna, AI methods solved more than 95% of 57 design challenges, matching the performance of expert human players ...
Cryo–electron microscopy revealed that the molecules folded into entirely new three-dimensional architectures, sometimes featuring intricate interactions that the AI had not been instructed to build. RNA design has thus entered the deep-learning era. ...
Abstract
RNA design has been hindered by the limited accuracy of three-dimensional (3D) structure prediction. In this study, we show that intricate RNA structures can be generated with current deep learning tools through accurate de novo design of pseudoknot secondary structures.
In an Eterna competition involving 57 pseudoknots, generative artificial intelligence (AI) methods matched experienced human designers in solving most blind challenges, evaluated by single nucleotide–resolution chemical mapping, compensatory mutagenesis, and cryo–electron microscopy.
AI-generated molecules with accurate secondary structures formed well-ordered 3D folds stabilized by noncanonical tertiary interactions not modeled during design.
Success was guided by an RNet foundation model trained on prior chemical mapping data, suggesting that some difficult RNA design tasks may be tractable without first solving RNA 3D structure prediction."
OpenKnot (Eterna competition) "Many important biological processes depend on RNAs that form pseudoknots, and they are among the most conserved structures in evolutionary history. However, scientists still have much to learn about their properties, structure, and functions."
De novo design of RNA pseudoknots with deep learning (no public access)
De novo design of RNA pseudoknots with deep learning (preprint, open access)
Figure 2 The workflow of RNet.
(A) RNA native structures or
(B) simulation trajectories are transformed into networks.
(C) A machine learning-based algorithm decomposes the local and global network properties to identify binding sites.
(D) A distance-based dynamical graph algorithm can accurately describe the binding dynamical motions.
(E, F) Local and Global network properties.
(G) The diagram of the DDNC.
Some of the RNA pseudoknots
Fig 4 Cryo-electron microscopy of AI-designed pseudoknotted RNA.
(A) Secondary structure of Kissing Multiloops (target P20 in Round 3), colored by stem.
(B) AlphaFold 3 3D informed secondary structure and predicted model, colored by stem.
(C-E) For the tested designs from (C) Struct2SeQ-SHAPE,
(D) MPNN-fixbb, and
(E) gRNAde, cryo-EM derived secondary structures (top), cryo-EM maps (unsharpened) and fitted coordinates, colored by stem (bottom), show high accuracy in recovering the target pseudoknot secondary structure while also highlighting distinct topologies from AlphaFold 3 prediction and noncanonical interactions (insets under (D) and (E)).
No comments:
Post a Comment