A key goal of this competition is to develop models that understand 3D structures of RNA and we think it might also help to use existing 3D structure prediction models to generate 3D coordinates to model the data in Ribonanza. As a result, we have generated 3D coordinates with RhoFold for a subset (~120k sequences) of the Ribonanza data. These can be accessed at:
\nhttps://www.kaggle.com/datasets/shujun717/ribonanza-3d-coords
Each folder is named by sequence_id and unrelaxed_model.pdb has the 3D coordinates predicted by RhoFold. We hope that this will help your modeling efforts!
@rhijudas has also made a notebook showing how to install RhoFold and visualize 3D structures:
\n https://www.kaggle.com/code/rhijudas/interactive-rhofold
Video lecture from @rhijudas on 3D structure prediction and how Ribonanza can help us create an Alphafold moment for RNA:
\nhttps://www.youtube.com/watch?v=R6-MwrGkj7M&t=95s&ab_channel=CASPRNASIG
\nRhoFold paper:
\nhttps://arxiv.org/abs/2207.01586
A key goal of this competition is to develop models that understand 3D structures of RNA and we think it might also help to use existing 3D structure prediction models to generate 3D coordinates to model the data in Ribonanza. As a result, we have generated 3D coordinates with RhoFold for a subset (~120k sequences) of the Ribonanza data. These can be accessed at:
\nhttps://www.kaggle.com/datasets/shujun717/ribonanza-3d-coords
Each folder is named by sequence_id and unrelaxed_model.pdb has the 3D coordinates predicted by RhoFold. We hope that this will help your modeling efforts!
@rhijudas has also made a notebook showing how to install RhoFold and visualize 3D structures:
\n https://www.kaggle.com/code/rhijudas/interactive-rhofold
Video lecture from @rhijudas on 3D structure prediction and how Ribonanza can help us create an Alphafold moment for RNA:
\nhttps://www.youtube.com/watch?v=R6-MwrGkj7M&t=95s&ab_channel=CASPRNASIG
\nRhoFold paper:
\nhttps://arxiv.org/abs/2207.01586
Here are some additional packages with downloadable code that might be of interest:
\n• trRosettaRNA (https://yanglab.qd.sdu.edu.cn/trRosettaRNA/) has a nice server (& training sets!) as well as downloadable code. The package has a neural network to predict distograms and other '2D' information about an RNA, which may be useful for comparison to Ribonanza chemical mapping data. These features are then used to build a 3D structure model in a process that is slower and not differentiable.
\n• RhoFold (previously called E2Efold-3D) is end-to-end from sequence to 3D structure. It used to have a server and github repository, which are apparently under maintenance. You can access the previously public code through the notebook (https://www.kaggle.com/code/rhijudas/interactive-rhofold) that @shujun717 and I set up!
\n• RosettaFold2NA is another end-to-end differentiable model, and its training that includes not just RNA but also RNA-protein complexes, DNA, and more. It's available here: https://github.com/uw-ipd/RoseTTAFold2NA.
\n• OpenComplex builds on the OpenFold framework and includes RNA. Code is here: https://github.com/baaihealth/OpenComplex
\n• DeepFoldRNA is available for download and for server use here: https://zhanggroup.org/DeepFoldRNA/. Similar to trRosettaRNA, it's not quite end-to-end, with an L-BGFS based 3D structure modeling part.
\nThere are other packages for fast and differentiable folding of 3D RNA structure emerging, too -- please post in this thread as you find them!
\nSome additional notes/caveats:
\n• None of the above packages natively predict chemical mapping data from their 2D features or 3D models. So using them to learn from Ribonanza data and to make submissions will require some additional development!
\n• These packages only output a single structure, whereas many --perhaps most-- of the Ribonanza molecules actually form an ensemble of diverse 3D structures. If using the package output, modeling may benefit from using dropout at inference time and train time to generate multiple structures. The predicted chemical mapping profiles corresponding to each of these structures could perhaps be averaged for comparison to experimental profiles.
\n• Most of the above packages follow the lead of AlphaFold and RosettaFold for proteins -- they rely on the input sequence having many recognizable evolutionary homologs (so-called \"multiple sequence alignments\" or MSA's). These MSA's are not readily available for the majority of Ribonanza molecules, which are either non-natural (much of Eterna) or are not members of previously curated non-coding RNA families. To save time and enforce generality, you may wish to run the models above without the step of generating MSA's -- just input the target sequence as a single sequence MSA.
\n• These neural packages did not do as well as expert human modelers in last year's CASP15 3D structure modeling competition. But perhaps that's due to a data bottleneck that can be resolved with our efforts here in the Ribonanza challenge.
\nLooking forward to seeing what you find!
", "rawMarkdown": "Here are some additional packages with downloadable code that might be of interest:\n\n• **trRosettaRNA** (https://yanglab.qd.sdu.edu.cn/trRosettaRNA/) has a nice server (& training sets!) as well as downloadable code. The package has a neural network to predict distograms and other '2D' information about an RNA, which may be useful for comparison to Ribonanza chemical mapping data. These features are then used to build a 3D structure model in a process that is slower and not differentiable.\n\n• **RhoFold** (previously called E2Efold-3D) is end-to-end from sequence to 3D structure. It used to have a server and github repository, which are apparently under maintenance. You can access the previously public code through the notebook (https://www.kaggle.com/code/rhijudas/interactive-rhofold) that @shujun717 and I set up!\n\n• **RosettaFold2NA** is another end-to-end differentiable model, and its training that includes not just RNA but also RNA-protein complexes, DNA, and more. It's available here: https://github.com/uw-ipd/RoseTTAFold2NA. \n\n• **OpenComplex** builds on the OpenFold framework and includes RNA. Code is here: https://github.com/baaihealth/OpenComplex \n\n• **DeepFoldRNA** is available for download and for server use here: https://zhanggroup.org/DeepFoldRNA/. Similar to trRosettaRNA, it's not quite end-to-end, with an L-BGFS based 3D structure modeling part.\n\nThere are other packages for fast and differentiable folding of 3D RNA structure emerging, too -- please post in this thread as you find them!\n\nSome additional notes/caveats:\n\n• None of the above packages natively predict chemical mapping data from their 2D features or 3D models. So using them to learn from Ribonanza data and to make submissions will require some additional development!\n\n• These packages only output a single structure, whereas many --perhaps most-- of the Ribonanza molecules actually form an ensemble of diverse 3D structures. If using the package output, modeling may benefit from using dropout at inference time and train time to generate multiple structures. The predicted chemical mapping profiles corresponding to each of these structures could perhaps be averaged for comparison to experimental profiles.\n\n• Most of the above packages follow the lead of AlphaFold and RosettaFold for proteins -- they rely on the input sequence having many recognizable evolutionary homologs (so-called \"multiple sequence alignments\" or MSA's). These MSA's are *not* readily available for the majority of Ribonanza molecules, which are either non-natural (much of Eterna) or are not members of previously curated non-coding RNA families. To save time and enforce generality, you may wish to run the models above without the step of generating MSA's -- just input the target sequence as a single sequence MSA.\n\n• These neural packages did not do as well as expert human modelers in last year's [CASP15 3D structure modeling competition](https://doi.org/10.1002/prot.26602). But perhaps that's due to a data bottleneck that can be resolved with our efforts here in the Ribonanza challenge.\n\n\nLooking forward to seeing what you find!", "votes": null }, { "id": "2505760", "postDate": "10/30/2023 19:32:31", "content": "Could you check the error messages? Thank you.
\nhttps://www.kaggle.com/crimson206/interactive-rhofold-errorlog
", "rawMarkdown": "Could you check the error messages? Thank you.\n\nhttps://www.kaggle.com/crimson206/interactive-rhofold-errorlog", "votes": null }, { "id": "2505795", "postDate": "10/30/2023 20:17:26", "content": "\n\nAs a result, we have generated 3D coordinates with RhoFold for a subset (~120k sequences) of the Ribonanza data.
\n
Hi, thanks for the insightful post.
\nIs there a reason you only did 120k sequences?
Would it be feasible for us to create the coordinates for all sequences, or at least all test sequences?
", "rawMarkdown": "> As a result, we have generated 3D coordinates with RhoFold for a subset (~120k sequences) of the Ribonanza data.\n\nHi, thanks for the insightful post. \nIs there a reason you only did 120k sequences? \n\nWould it be feasible for us to create the coordinates for all sequences, or at least all test sequences?", "votes": null }, { "id": "2505971", "postDate": "10/31/2023 02:49:11", "content": "Thank you. How well do the simulated 3D structures correspond to real structures (seen in experiments)?
", "rawMarkdown": "Thank you. How well do the simulated 3D structures correspond to real structures (seen in experiments)?", "votes": null }, { "id": "2506106", "postDate": "10/31/2023 05:20:35", "content": "Interesting and timely post! Had been looking at a paper suggested elsewhere in Discussion -
\nPhysics-aware Graph Neural Network for Accurate RNA 3D Structure Prediction
\nhttps://arxiv.org/abs/2210.16392
\nand code https://github.com/zetayue/Physics-aware-Multiplex-GNN
The paper referenced ARES a neural network, the Atomic Rotationally Equivariant Scorer in Geometric deep learning of rna structure (Rhiju Das is one of the authors) https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9829186/ and links to code in the references.
\nWas considering how feasible these might be here. If available RNA 3D structures can be leveraged in an easier more efficient way, need less resources than e.g. AlphaFold which has a scaled down version in
\nhttps://colab.research.google.com/github/deepmind/alphafold/blob/main/notebooks/AlphaFold.ipynb
Thanks for providing RhoFold, hit the not found 404 error previously looking for it.
", "rawMarkdown": "Interesting and timely post! Had been looking at a paper suggested elsewhere in Discussion - \nPhysics-aware Graph Neural Network for Accurate RNA 3D Structure Prediction\nhttps://arxiv.org/abs/2210.16392\nand code https://github.com/zetayue/Physics-aware-Multiplex-GNN\n\nThe paper referenced ARES a neural network, the Atomic Rotationally Equivariant Scorer in Geometric deep learning of rna structure (Rhiju Das is one of the authors) https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9829186/ and links to code in the references.\n\nWas considering how feasible these might be here. If available RNA 3D structures can be leveraged in an easier more efficient way, need less resources than e.g. AlphaFold which has a scaled down version in \nhttps://colab.research.google.com/github/deepmind/alphafold/blob/main/notebooks/AlphaFold.ipynb\n\nThanks for providing RhoFold, hit the not found 404 error previously looking for it.", "votes": null }, { "id": "2506738", "postDate": "10/31/2023 14:09:06", "content": "Thanks for flagging. There's a new version of the RhoFold notebook (link) that should run all the way through without errors.
", "rawMarkdown": "Thanks for flagging. There's a new version of the RhoFold notebook ([link](https://www.kaggle.com/code/rhijudas/interactive-rhofold)) that should run all the way through without errors.", "votes": null }, { "id": "2506883", "postDate": "10/31/2023 15:54:20", "content": "Just out, new AlphaFold will include RNA 3D modeling, albeit not with human-competitive accuracy: https://deepmind.google/discover/blog/a-glimpse-of-the-next-generation-of-alphafold/
\nCode does not seem to be released yet, though!
@shujun717 can you please give your opinion on performance of RHA Fold with and without MSA?
", "rawMarkdown": "shujun717 can you please give your opinion on performance of RHA Fold with and without MSA?", "votes": null }, { "id": "2508724", "postDate": "11/02/2023 00:03:55", "content": "@shujun717 I ask you a favor, the dataset is uploaded as a zip within another zip so the PDBS files cannot be accessed when importing the dataset into the notebooks. Could you please upload it with a single zip level? Thank you so much !
", "rawMarkdown": "shujun717 I ask you a favor, the dataset is uploaded as a zip within another zip so the PDBS files cannot be accessed when importing the dataset into the notebooks. Could you please upload it with a single zip level? Thank you so much !", "votes": null }, { "id": "2508736", "postDate": "11/02/2023 00:29:19", "content": "I see. Will fix soon
", "rawMarkdown": "I see. Will fix soon", "votes": null }, { "id": "2508738", "postDate": "11/02/2023 00:31:01", "content": "I have used RhoFold's coordinates in some pre-testing with a very small subset of full training data and RhoFold coordinates provided some signal for modeling
", "rawMarkdown": "I have used RhoFold's coordinates in some pre-testing with a very small subset of full training data and RhoFold coordinates provided some signal for modeling", "votes": null }, { "id": "2509518", "postDate": "11/02/2023 12:26:06", "content": "@shujun717 @lauraromar
\nhave added a notebook https://www.kaggle.com/code/something4kag/ribonanza-3d-coords-prep
\nto extract the ribonanza-3d-coords dataset to a new zip with folders per sequence id using shutil
\nit also creates and saves a reduced version of train.csv for the sequence ids with PDBs
there is a follow on notebook https://www.kaggle.com/something4kag/ribonanza-3d-coords-using-prep
\nthat shows how to extract the zip in prep and an view example /unrelaxed_model.pdb for a sequence id
Did you create a contact matrix from the PDB files ?
", "rawMarkdown": "Did you create a contact matrix from the PDB files ?", "votes": null }, { "id": "2510244", "postDate": "11/02/2023 20:25:15", "content": "Thanks a lot!
", "rawMarkdown": "Thanks a lot!", "votes": null }, { "id": "2510388", "postDate": "11/03/2023 00:31:44", "content": "Yeah something similar to that
", "rawMarkdown": "Yeah something similar to that", "votes": null }, { "id": "2511755", "postDate": "11/03/2023 23:21:40", "content": "Thank you so much !
", "rawMarkdown": "Thank you so much !", "votes": null }, { "id": "2532281", "postDate": "11/20/2023 22:49:11", "content": "@shujun717 , @rhijudas : how accurate can we expect these 3D coordinates to be, given that they come from RhoFold, itself a learned model?
", "rawMarkdown": "shujun717 , @rhijudas : how accurate can we expect these 3D coordinates to be, given that they come from RhoFold, itself a learned model?", "votes": null } ], "comments": [ { "id": 2505758, "author_name": "rhijudas", "author_url": "", "post_date": "10/30/2023 19:29:57", "content": "Here are some additional packages with downloadable code that might be of interest:
\n• trRosettaRNA (https://yanglab.qd.sdu.edu.cn/trRosettaRNA/) has a nice server (& training sets!) as well as downloadable code. The package has a neural network to predict distograms and other '2D' information about an RNA, which may be useful for comparison to Ribonanza chemical mapping data. These features are then used to build a 3D structure model in a process that is slower and not differentiable.
\n• RhoFold (previously called E2Efold-3D) is end-to-end from sequence to 3D structure. It used to have a server and github repository, which are apparently under maintenance. You can access the previously public code through the notebook (https://www.kaggle.com/code/rhijudas/interactive-rhofold) that @shujun717 and I set up!
\n• RosettaFold2NA is another end-to-end differentiable model, and its training that includes not just RNA but also RNA-protein complexes, DNA, and more. It's available here: https://github.com/uw-ipd/RoseTTAFold2NA.
\n• OpenComplex builds on the OpenFold framework and includes RNA. Code is here: https://github.com/baaihealth/OpenComplex
\n• DeepFoldRNA is available for download and for server use here: https://zhanggroup.org/DeepFoldRNA/. Similar to trRosettaRNA, it's not quite end-to-end, with an L-BGFS based 3D structure modeling part.
\nThere are other packages for fast and differentiable folding of 3D RNA structure emerging, too -- please post in this thread as you find them!
\nSome additional notes/caveats:
\n• None of the above packages natively predict chemical mapping data from their 2D features or 3D models. So using them to learn from Ribonanza data and to make submissions will require some additional development!
\n• These packages only output a single structure, whereas many --perhaps most-- of the Ribonanza molecules actually form an ensemble of diverse 3D structures. If using the package output, modeling may benefit from using dropout at inference time and train time to generate multiple structures. The predicted chemical mapping profiles corresponding to each of these structures could perhaps be averaged for comparison to experimental profiles.
\n• Most of the above packages follow the lead of AlphaFold and RosettaFold for proteins -- they rely on the input sequence having many recognizable evolutionary homologs (so-called \"multiple sequence alignments\" or MSA's). These MSA's are not readily available for the majority of Ribonanza molecules, which are either non-natural (much of Eterna) or are not members of previously curated non-coding RNA families. To save time and enforce generality, you may wish to run the models above without the step of generating MSA's -- just input the target sequence as a single sequence MSA.
\n• These neural packages did not do as well as expert human modelers in last year's CASP15 3D structure modeling competition. But perhaps that's due to a data bottleneck that can be resolved with our efforts here in the Ribonanza challenge.
\nLooking forward to seeing what you find!
", "votes": null, "replies": [] }, { "id": 2505760, "author_name": "", "author_url": "", "post_date": "10/30/2023 19:32:31", "content": "Could you check the error messages? Thank you.
\nhttps://www.kaggle.com/crimson206/interactive-rhofold-errorlog
", "votes": null, "replies": [ { "id": 2506738, "author_name": "rhijudas", "author_url": "", "post_date": "10/31/2023 14:09:06", "content": "Thanks for flagging. There's a new version of the RhoFold notebook (link) that should run all the way through without errors.
", "votes": null, "replies": [ { "id": 2510244, "author_name": "", "author_url": "", "post_date": "11/02/2023 20:25:15", "content": "Thanks a lot!
", "votes": null, "replies": [] } ] } ] }, { "id": 2505795, "author_name": "fnands", "author_url": "", "post_date": "10/30/2023 20:17:26", "content": "\n\nAs a result, we have generated 3D coordinates with RhoFold for a subset (~120k sequences) of the Ribonanza data.
\n
Hi, thanks for the insightful post.
\nIs there a reason you only did 120k sequences?
Would it be feasible for us to create the coordinates for all sequences, or at least all test sequences?
", "votes": null, "replies": [] }, { "id": 2505971, "author_name": "sapr3s", "author_url": "", "post_date": "10/31/2023 02:49:11", "content": "Thank you. How well do the simulated 3D structures correspond to real structures (seen in experiments)?
", "votes": null, "replies": [] }, { "id": 2506106, "author_name": "something4kag", "author_url": "", "post_date": "10/31/2023 05:20:35", "content": "Interesting and timely post! Had been looking at a paper suggested elsewhere in Discussion -
\nPhysics-aware Graph Neural Network for Accurate RNA 3D Structure Prediction
\nhttps://arxiv.org/abs/2210.16392
\nand code https://github.com/zetayue/Physics-aware-Multiplex-GNN
The paper referenced ARES a neural network, the Atomic Rotationally Equivariant Scorer in Geometric deep learning of rna structure (Rhiju Das is one of the authors) https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9829186/ and links to code in the references.
\nWas considering how feasible these might be here. If available RNA 3D structures can be leveraged in an easier more efficient way, need less resources than e.g. AlphaFold which has a scaled down version in
\nhttps://colab.research.google.com/github/deepmind/alphafold/blob/main/notebooks/AlphaFold.ipynb
Thanks for providing RhoFold, hit the not found 404 error previously looking for it.
", "votes": null, "replies": [] }, { "id": 2506883, "author_name": "rhijudas", "author_url": "", "post_date": "10/31/2023 15:54:20", "content": "Just out, new AlphaFold will include RNA 3D modeling, albeit not with human-competitive accuracy: https://deepmind.google/discover/blog/a-glimpse-of-the-next-generation-of-alphafold/
\nCode does not seem to be released yet, though!
@shujun717 can you please give your opinion on performance of RHA Fold with and without MSA?
", "votes": null, "replies": [ { "id": 2508738, "author_name": "shujun717", "author_url": "", "post_date": "11/02/2023 00:31:01", "content": "I have used RhoFold's coordinates in some pre-testing with a very small subset of full training data and RhoFold coordinates provided some signal for modeling
", "votes": null, "replies": [ { "id": 2509799, "author_name": "mtinti", "author_url": "", "post_date": "11/02/2023 14:56:39", "content": "Did you create a contact matrix from the PDB files ?
", "votes": null, "replies": [ { "id": 2510388, "author_name": "shujun717", "author_url": "", "post_date": "11/03/2023 00:31:44", "content": "Yeah something similar to that
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2508724, "author_name": "lauraromar", "author_url": "", "post_date": "11/02/2023 00:03:55", "content": "@shujun717 I ask you a favor, the dataset is uploaded as a zip within another zip so the PDBS files cannot be accessed when importing the dataset into the notebooks. Could you please upload it with a single zip level? Thank you so much !
", "votes": null, "replies": [ { "id": 2508736, "author_name": "shujun717", "author_url": "", "post_date": "11/02/2023 00:29:19", "content": "I see. Will fix soon
", "votes": null, "replies": [ { "id": 2509518, "author_name": "something4kag", "author_url": "", "post_date": "11/02/2023 12:26:06", "content": "@shujun717 @lauraromar
\nhave added a notebook https://www.kaggle.com/code/something4kag/ribonanza-3d-coords-prep
\nto extract the ribonanza-3d-coords dataset to a new zip with folders per sequence id using shutil
\nit also creates and saves a reduced version of train.csv for the sequence ids with PDBs
there is a follow on notebook https://www.kaggle.com/something4kag/ribonanza-3d-coords-using-prep
\nthat shows how to extract the zip in prep and an view example /unrelaxed_model.pdb for a sequence id
Thank you so much !
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2532281, "author_name": "fnands", "author_url": "", "post_date": "11/20/2023 22:49:11", "content": "@shujun717 , @rhijudas : how accurate can we expect these 3D coordinates to be, given that they come from RhoFold, itself a learned model?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2505739": "A key goal of this competition is to develop models that understand 3D structures of RNA and we think it might also help to use existing 3D structure prediction models to generate 3D coordinates to model the data in Ribonanza. As a result, we have generated 3D coordinates with RhoFold for a subset (~120k sequences) of the Ribonanza data. These can be accessed at:\nhttps://www.kaggle.com/datasets/shujun717/ribonanza-3d-coords\n\nEach folder is named by `sequence_id` and `unrelaxed_model.pdb` has the 3D coordinates predicted by RhoFold. We hope that this will help your modeling efforts! \n\n@rhijudas has also made a notebook showing how to install RhoFold and visualize 3D structures: \n https://www.kaggle.com/code/rhijudas/interactive-rhofold\n\nVideo lecture from @rhijudas on 3D structure prediction and how Ribonanza can help us create an Alphafold moment for RNA:\nhttps://www.youtube.com/watch?v=R6-MwrGkj7M&t=95s&ab_channel=CASPRNASIG\nRhoFold paper:\nhttps://arxiv.org/abs/2207.01586", "2505758": "Here are some additional packages with downloadable code that might be of interest:\n\n• **trRosettaRNA** (https://yanglab.qd.sdu.edu.cn/trRosettaRNA/) has a nice server (& training sets!) as well as downloadable code. The package has a neural network to predict distograms and other '2D' information about an RNA, which may be useful for comparison to Ribonanza chemical mapping data. These features are then used to build a 3D structure model in a process that is slower and not differentiable.\n\n• **RhoFold** (previously called E2Efold-3D) is end-to-end from sequence to 3D structure. It used to have a server and github repository, which are apparently under maintenance. You can access the previously public code through the notebook (https://www.kaggle.com/code/rhijudas/interactive-rhofold) that @shujun717 and I set up!\n\n• **RosettaFold2NA** is another end-to-end differentiable model, and its training that includes not just RNA but also RNA-protein complexes, DNA, and more. It's available here: https://github.com/uw-ipd/RoseTTAFold2NA. \n\n• **OpenComplex** builds on the OpenFold framework and includes RNA. Code is here: https://github.com/baaihealth/OpenComplex \n\n• **DeepFoldRNA** is available for download and for server use here: https://zhanggroup.org/DeepFoldRNA/. Similar to trRosettaRNA, it's not quite end-to-end, with an L-BGFS based 3D structure modeling part.\n\nThere are other packages for fast and differentiable folding of 3D RNA structure emerging, too -- please post in this thread as you find them!\n\nSome additional notes/caveats:\n\n• None of the above packages natively predict chemical mapping data from their 2D features or 3D models. So using them to learn from Ribonanza data and to make submissions will require some additional development!\n\n• These packages only output a single structure, whereas many --perhaps most-- of the Ribonanza molecules actually form an ensemble of diverse 3D structures. If using the package output, modeling may benefit from using dropout at inference time and train time to generate multiple structures. The predicted chemical mapping profiles corresponding to each of these structures could perhaps be averaged for comparison to experimental profiles.\n\n• Most of the above packages follow the lead of AlphaFold and RosettaFold for proteins -- they rely on the input sequence having many recognizable evolutionary homologs (so-called \"multiple sequence alignments\" or MSA's). These MSA's are *not* readily available for the majority of Ribonanza molecules, which are either non-natural (much of Eterna) or are not members of previously curated non-coding RNA families. To save time and enforce generality, you may wish to run the models above without the step of generating MSA's -- just input the target sequence as a single sequence MSA.\n\n• These neural packages did not do as well as expert human modelers in last year's [CASP15 3D structure modeling competition](https://doi.org/10.1002/prot.26602). But perhaps that's due to a data bottleneck that can be resolved with our efforts here in the Ribonanza challenge.\n\n\nLooking forward to seeing what you find!", "2505760": "Could you check the error messages? Thank you.\n\nhttps://www.kaggle.com/crimson206/interactive-rhofold-errorlog", "2505795": "> As a result, we have generated 3D coordinates with RhoFold for a subset (~120k sequences) of the Ribonanza data.\n\nHi, thanks for the insightful post. \nIs there a reason you only did 120k sequences? \n\nWould it be feasible for us to create the coordinates for all sequences, or at least all test sequences?", "2505971": "Thank you. How well do the simulated 3D structures correspond to real structures (seen in experiments)?", "2506106": "Interesting and timely post! Had been looking at a paper suggested elsewhere in Discussion - \nPhysics-aware Graph Neural Network for Accurate RNA 3D Structure Prediction\nhttps://arxiv.org/abs/2210.16392\nand code https://github.com/zetayue/Physics-aware-Multiplex-GNN\n\nThe paper referenced ARES a neural network, the Atomic Rotationally Equivariant Scorer in Geometric deep learning of rna structure (Rhiju Das is one of the authors) https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9829186/ and links to code in the references.\n\nWas considering how feasible these might be here. If available RNA 3D structures can be leveraged in an easier more efficient way, need less resources than e.g. AlphaFold which has a scaled down version in \nhttps://colab.research.google.com/github/deepmind/alphafold/blob/main/notebooks/AlphaFold.ipynb\n\nThanks for providing RhoFold, hit the not found 404 error previously looking for it.", "2506738": "Thanks for flagging. There's a new version of the RhoFold notebook ([link](https://www.kaggle.com/code/rhijudas/interactive-rhofold)) that should run all the way through without errors.", "2506883": "Just out, new AlphaFold will include RNA 3D modeling, albeit not with human-competitive accuracy: https://deepmind.google/discover/blog/a-glimpse-of-the-next-generation-of-alphafold/ \nCode does not seem to be released yet, though!", "2508708": "shujun717 can you please give your opinion on performance of RHA Fold with and without MSA?", "2508724": "shujun717 I ask you a favor, the dataset is uploaded as a zip within another zip so the PDBS files cannot be accessed when importing the dataset into the notebooks. Could you please upload it with a single zip level? Thank you so much !", "2508736": "I see. Will fix soon", "2508738": "I have used RhoFold's coordinates in some pre-testing with a very small subset of full training data and RhoFold coordinates provided some signal for modeling", "2509518": "shujun717 @lauraromar \nhave added a notebook https://www.kaggle.com/code/something4kag/ribonanza-3d-coords-prep\nto extract the ribonanza-3d-coords dataset to a new zip with folders per sequence id using shutil\nit also creates and saves a reduced version of train.csv for the sequence ids with PDBs\n\n\nthere is a follow on notebook https://www.kaggle.com/something4kag/ribonanza-3d-coords-using-prep\nthat shows how to extract the zip in prep and an view example /unrelaxed_model.pdb for a sequence id", "2509799": "Did you create a contact matrix from the PDB files ?", "2510244": "Thanks a lot!", "2510388": "Yeah something similar to that", "2511755": "Thank you so much !", "2532281": "shujun717 , @rhijudas : how accurate can we expect these 3D coordinates to be, given that they come from RhoFold, itself a learned model?" }, "source": "meta" }