Nice overview in @naturemethods of #RNA structure approaches featuring our very own @RachaelKretsch and lots of amazing colleagues across the world

At Final Submission Deadline, numerous @kaggle notebooks outperform VFold human expert baseline in #RNA 3D structure prediction! Will these codes sustain performance on new targets released between now and Sep 2025? Join the discussion: https://t.co/6wMz2E1UK4

#RNA 3D structure prediction remains an open problem! Come test new ideas against SOTA & >1000 teams in Kaggle RNA 3D Folding competition, co-hosted by #CASP16 and @rnapuzzles, 2.5 weeks left. https://t.co/ODfoePDZ93 N/N

#CASP16 enabled by 46 predictor labs, 22 experimental #RNA and #DNA groups, co-assessors @rachaelkretsch, @alissahummer, @ShujunHe0717, @RongqingY, Jing Zhang, Thomas Karagianes, @qiancong, @AKryshtafovych Many thanks! 7/N

Upon accounting for available templates, no clear improvement in #CASP16 #RNA compared to prior blind challenges #CASP15 @rnapuzzles, except automated servers are less poor than they used to be. 6/N

#AlphaFold 3 server outperformed by 8 groups — including another #CASP16 server. Human expert group Vfold is SOTA. 5/N

Poor results from all predictors for #RNA and/or #DNA complexed with proteins and small molecules — even human experts can’t predict these 3D structures unless they’ve been experimentally resolved before. 4/N

#CASP16 global #RNA secondary structure accuracies were quite good. But hallmarks of functional tertiary structure –pseudoknots, singlet pairs, non-canonicals, and tertiary motifs like A-minor interactions– remain difficult to predict. 3/N

There’s one notable #CASP16 #RNA fold prediction without previously available template, the ornate large extremophilic (OLE) RNA, which has an unusually deep multiple sequence alignment. But it’s still not high accuracy. 2/N

Nucleic acid assessment for #CASP16 blind competition is out. As in prior challenges #deeplearning for 3D #RNA structure in #CASP16 prediction remains disappointing. https://t.co/V5AZwux2zx 1/N

Pretty new discovery! 💎 SLAC and @Stanford researchers using cryogenic electron microscopy showed for the first time that large RNA complexes can assemble without the help of proteins, expanding our understanding of RNA folding and function: https://t.co/f8Z59zDOc7

What are these ornate RNA structures doing in phages and bacteria? Can we use them for biotechnology? Great collab led by @rachaelkretsch, with Vivian Wu, Svetlana Shabalina, Hyunbin Lee, Grace Nye, Eugene Koonin, @algao8 and @chiulab 2/2

Our #RNA multimer paper out now in @nature! #cryoEM of 2 distinct nanocages GOLLD and ROOL and membrane-associated OLE. Multimers at ultra-low concentrations, special covariance for cross-molecule kissing loops, and speculation on function https://t.co/ZihCKHIIcX. 1/2

Stanford #RNA 3D Folding: With data refresh done, the first two superhuman submissions that make public their documented, no-cheat notebooks will get the Early Sharing Prizes. Will EIGEN be the first? https://t.co/3SiTmewCQc @kaggle @HHMIJanelia @HHMINEWS @StanfordMed #CASP16

And many thanks to @LenaSteckelberg @HartmannLab @esandersen @chiulab @HoebartnerLab @cpricejones @KoiralaLab @cdmackereth @Marcia_lincRlab Joe Piccirilli @PhoebeARice @SuZhaoming and labs for their insightful analyses. Reality check on #RNA structure prediction! 2/2

Are RNA 3D models accurate enough to be biologically useful? #CASP16 experimentalists say: not so much. https://t.co/rUlFyDaiUJ Thanks for leading this paper @rachaelkretsch @AKryshtafovych! 1/2

@shujun @nvidia @kaggle @eternagame @nvidia blog on training Rnet2 with @SJ_He via NAIRR pilot program and DGX cloud -- more coming from this effort, including RNA design AI results on @eternagame -- stay tuned and do join the competition on @kaggle by May 2025:

We see a path to scale up to a 30B param foundation model Rnet3 that should solve RNA 3D structure prediction. But #deeplearning training costs now dwarf data generation costs. Anyone have 100M GPU-h they can spare to solve one of biology’s remaining grand challenges? (N/N)

Much of the experimental tech driving Rnet2 — including ultrafast/cheap sequencing @UltimaGenomics and “AI-ready” Gentitan synthesis @GenScript — wasn’t even available last year. (6/N)

Rnet2 training data are on synthetic designs from RFdiffusion @UWproteindesign, gRNade@chaitjo, genome scans, and a new design method Shujun @SJ_He. Sequences mapped thanks to key experimental innovations by Ann Kladwang and Hamish Blair @rdaslab. (5/N)