\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F86fb50733cfe403269f890516f09a7bf%2FSelection_999(7739).png?generation=1741221549647530&alt=media\")
https://www.biorxiv.org/content/10.1101/2024.06.13.598780v1.full.pdf
\nHas AlphaFold 3 reached its success for RNAs? - 2024
a good paper:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F572406e24bbe7acba69a5b98e6bcaeda%2FSelection_999(7740).png?generation=1741221636244020&alt=media\")
\nalphafold3 results (TM-score)
https://www.biorxiv.org/content/10.1101/2024.06.13.598780v1.full.pdf
\nHas AlphaFold 3 reached its success for RNAs? - 2024
a good paper:
\n
\nalphafold3 results (TM-score)
Thanks for the paper @hengck23
\nThe performance seems to depend (a lot) on the test dataset. Is it due to leakage in the test set of some benchmarks ? AF3 seems the way-to-go though, based on the 1st plot, as it performs on-par or better than the 9 other methods
Thank you so much!
", "rawMarkdown": "Thank you so much!", "votes": null }, { "id": "3147217", "postDate": "03/11/2025 19:10:22", "content": "https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012715
\nSystematic benchmarking of deep-learning methods for tertiary RNA structure prediction
https://www.cell.com/structure/pdf/S0969-2126(24)00332-0.pdf
\nAdvances in the field of RNA 3D structure prediction and modeling, with purely theoretical approaches, and with the use of experimental data
Thanks @hengck23! The conclusion of that paper is interesting
\n\n\nIts new architecture allows the prediction of wide molecules but remains limited and hardly predicts non-Watson-Crick interactions. The predictions of AlphaFold 3 remain of competitive quality, as it outperforms most of the existing solutions and returns more physically plausible structures than ab initio methods. It outclasses existing deep-learning approaches but does not generalize well on orphan structures or long RNAs. It also returns predictions very quickly compared to ab initio or current template-based approaches. The prediction of atom coordinates instead of base frames in AlphaFold 2 allows the extension of predictions for a wide range of molecules but prevents the generalisation of RNAspecific interactions. The lack of data is also a limitation that prevents the robustness of deep learning methods in general, and so is AlphaFold 3.
\n
Takeaways:
\nThanks for the paper @hengck23
\nThe performance seems to depend (a lot) on the test dataset. Is it due to leakage in the test set of some benchmarks ? AF3 seems the way-to-go though, based on the 1st plot, as it performs on-par or better than the 9 other methods
Thank you so much!
", "votes": null, "replies": [] }, { "id": 3147217, "author_name": "hengck23", "author_url": "", "post_date": "03/11/2025 19:10:22", "content": "https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012715
\nSystematic benchmarking of deep-learning methods for tertiary RNA structure prediction
https://www.cell.com/structure/pdf/S0969-2126(24)00332-0.pdf
\nAdvances in the field of RNA 3D structure prediction and modeling, with purely theoretical approaches, and with the use of experimental data
Thanks @hengck23! The conclusion of that paper is interesting
\n\n\nIts new architecture allows the prediction of wide molecules but remains limited and hardly predicts non-Watson-Crick interactions. The predictions of AlphaFold 3 remain of competitive quality, as it outperforms most of the existing solutions and returns more physically plausible structures than ab initio methods. It outclasses existing deep-learning approaches but does not generalize well on orphan structures or long RNAs. It also returns predictions very quickly compared to ab initio or current template-based approaches. The prediction of atom coordinates instead of base frames in AlphaFold 2 allows the extension of predictions for a wide range of molecules but prevents the generalisation of RNAspecific interactions. The lack of data is also a limitation that prevents the robustness of deep learning methods in general, and so is AlphaFold 3.
\n
Takeaways:
\n