RibonanzaNet-SS live in the game
The RibonanzaNet-SS (RNNet) neural net developed by Shujun He (Texas A&M) in collaboration with the Das Lab, Kaggle, and Eterna is now implemented in the game for lab projects. Please begin using RNNet to solve the puzzle in OpenKnot Round 6. The folding calculation is not fast. The freeze button (snowflake icon) allows players to make several changes at once before engaging the folding engine.
RNNet does not generate an energy calculation. Instead, the algorithm generates eF1 and eF1,crossed-pair. These can be viewed in the spec box under the Info menu in the toolbox, or by pressing the "S" key. The plan is to display new metrics in the puzzle interface in the near future. The eF1 score indicates how confident the model is in its prediction and the eF1,crossed-pair score reflects confidence in the crossing pair (pseudoknot) portion of the predicted structure. We are looking especially for designs with eF1,crossed-pair values greater than 0.8.
RNNet could possibly be a major advance in secondary structure prediction, and the goal of OpenKnot Round 6 is to identify the strengths and weaknesses of the new model. Please submit designs that RNNet predicts will fold into the target structure. The chemical probing experiment will verify whether the design folds into that structure.
The new folding engine will continue to be improved over the next year as new training data helps fine-tune the model. Since the folding engine will continue to evolve, it cannot be used to create player puzzles in Puzzlemaker at this time. Also, most existing boosters will not work when RNNet is the active folding engine - a number of EternaScript functions will need to be changed to use their new “async” versions instead (see https://eternagame.org/scripts/docs for more details). Please post questions in the Forum or Discord for more information about scripts.
Shujun He will be kicking off Eternacon Saturday with a presentation on RNNet. Please attend to learn more about the new model and ask questions. The new model also is explained in the bioRxiv preprint.