AI selected sequences for complex RNA pseudoknots, interwoven regions within a single molecule; experiments confirmed their structures
AI selected sequences for complex RNA pseudoknots, interwoven regions within a single molecule; experiments confirmed their structures
On 27 August, Science published a study on the OpenKnot challenge, in which algorithms and experienced members of the Eterna online community designed sequences for RNA pseudoknots. In each of the two new rounds, both AI methods and Eterna participants achieved a score of 90 out of 100 for 19 of 20 targets. The resulting molecular structures were then tested using several independent methods.
RNA consists of chemical building blocks called nucleotides, and its structure affects how the molecule functions. In a pseudoknot, a loop in the RNA strand pairs with a distant region of the same strand, causing the pairing lines to cross in a two-dimensional diagram. Producing the intended structure is difficult because the same sequence can form a different set of base pairs in solution.
Eterna is an online community whose participants propose RNA sequences. A remote laboratory synthesizes selected designs and reports which regions of each molecule are flexible and which are base-paired. These measurements were used to train RNet, a model that predicts this type of chemical map for a new sequence. RNet helped eliminate weak designs before laboratory testing.
First, six AI methods and Eterna participants worked on 17 targets. Most algorithms performed worse than the human participants. In the second round, the researchers presented the same tasks again. Participants gained access to RNet through the interface, while the AI methods were updated to incorporate its predictions. The researchers then introduced two new sets of 20 targets, including one set containing RNA molecules up to 240 nucleotides long. In both sets, AI methods and Eterna participants found solutions scoring 90 out of 100 for 19 targets. These rounds tested whether the methods could generalize to unfamiliar RNA structures.
SHAPE provided the first test. A chemical reagent preferentially labels flexible nucleotides, while base-paired nucleotides remain protected. The authors then tested specific base pairs using compensatory mutations. They changed one RNA nucleotide to disrupt a pair, then changed both nucleotides so that the pair could form again. Restoration of the original chemical map supported the presence of that specific interaction. For 19 of 20 targets in the third round, at least one design restored at least 80% of the intended stems, which are regions formed by consecutive nucleotide pairs.
Three molecules with sequences proposed by AI methods were examined using cryo-electron microscopy, a method that reveals the structure of frozen molecules. All three contained the seven intended stems. The two highest-resolution maps also showed additional base pairs and contacts that the programs had not specified in advance.
The model proposes a sequence, the laboratory tests its structure, and the resulting data guide the next round of design.