![\"vfold](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11578898%2F9d529c00a1a57653e8ccb22b99235df2%2Fvfold.png?generation=1742635795652460&alt=media\")
Anyone ran Vfold Pipeline locally?
\nI think
\n
Anyone ran Vfold Pipeline locally?
\nI think
\n
I thought Vfold was a template-based model?
", "rawMarkdown": "I thought Vfold was a template-based model?", "votes": null }, { "id": "3156736", "postDate": "03/22/2025 13:25:20", "content": "Not sure if your results is correct. You should use results downloaded as reference. Further please consult vfold casp presentation material. Think it is a combinations of methods and tool.
\nNot sure casp submission is pure server results
", "rawMarkdown": "Not sure if your results is correct. You should use results downloaded as reference. Further please consult vfold casp presentation material. Think it is a combinations of methods and tool.\n\nNot sure casp submission is pure server results", "votes": null }, { "id": "3156744", "postDate": "03/22/2025 13:34:46", "content": "\nPresentation
\nhttps://predictioncenter.org/casp16/doc/CASP16_Abstracts.pdf
\nMethod abstract
\n\"We use various methods, including Vfold-Pipeline1, IsRNA2,3,4, and RNAJP5, to generate RNA 3D
\nstructures from the sequences. If structural templates for specific motifs in the RNA targets are identified
\nin the PDB database, we incorporate them as folding constraints. For protein/RNA/DNA complexes, RNA
\nwith ligands, and RNA with solvent shell, we employ AlphaFold36, ITScore7,8, and AMBER9 to generate
\nstructural candidates.\"
there is only one automatic top method:
\nhttps://predictioncenter.org/casp16/doc/presentations/Day-3/Day3-01-Kretsch_CASP16_NA_Assessement_PuntaCana_RCK_v1.pptx
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fea3db814c1f415b383afe8815ad0790c%2FSelection_114.png?generation=1742651300067567&alt=media\")
in my experiments, i select DL automatic methods and combine them.the combined version probbaly can beat vfold (even though each of the DL methods may be worse).
\npromising combinations are from af3, drfold2, rhofold+, nufold, trRoesetta2
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fb88dce18b9b6193d5b9ef8abd985f94b%2FSelection_115.png?generation=1742651809040035&alt=media\")
Yes and no.
\n![\"Vfold](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11578898%2F738221a8df0432c29f816556fb984d48%2Fvfold-archi.png?generation=1742651830128672&alt=media\")
According to https://vfold.missouri.edu/research.html,
\n• Vfold2D [1-4] a physics-based model that employs the statistics of known structures to compute energy parameters, enabling the prediction of 2D structures with pseudoknots and non-canonical base pairs
\n• VfoldMCPX [5-6] a physics-based model designed to predict the structure of multi-strand RNA assembly.
\n• Vfold3D [1] a motif template-based model
\n• VfoldLA [2] a loop template-based model
\n• IsRNA [3-5], a model based on statistical potentials and coarse-grained molecular dynamics
", "rawMarkdown": "Yes and no.\n\n\n\nAccording to https://vfold.missouri.edu/research.html, \n\n • Vfold2D [1-4] a physics-based model that employs the statistics of known structures to compute energy parameters, enabling the prediction of 2D structures with pseudoknots and non-canonical base pairs\n\n• VfoldMCPX [5-6] a physics-based model designed to predict the structure of multi-strand RNA assembly.\n\n• Vfold3D [1] a motif template-based model\n\n• VfoldLA [2] a loop template-based model\n\n• IsRNA [3-5], a model based on statistical potentials and coarse-grained molecular dynamics", "votes": null }, { "id": "3156830", "postDate": "03/22/2025 15:53:41", "content": "Given that Vfold is much higher than the second place's Zscore, it's unfortunate that we can't distill from vfold. I wonder what kind of human intervention was used in their version.
\n![\"Rank](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11578898%2Ff23a549ea61c797e426d470f8eb6c131%2FScreenshot%202025-03-22%20at%207.48.09PM.png?generation=1742658544677290&alt=media\")
\n(source: https://predictioncenter.org/casp16/zscores_rna.cgi)
This is z score not tm score. You need to remake the ranking from tm score download at casp16
", "rawMarkdown": "This is z score not tm score. You need to remake the ranking from tm score download at casp16", "votes": null }, { "id": "3156836", "postDate": "03/22/2025 16:03:28", "content": "tm score ranking is here
\nhttps://www.kaggle.com/competitions/stanford-rna-3d-folding/discussion/566906#3155466
Did u solve the problem? I also try u using vfold to cal the tm-score of R1107, the score is similar to yours, which is around 0.210
", "rawMarkdown": "Did u solve the problem? I also try u using vfold to cal the tm-score of R1107, the score is similar to yours, which is around 0.210", "votes": null }, { "id": "3174664", "postDate": "04/09/2025 10:18:16", "content": "please read the vfold presentation slide/method abstract and video for CASP16.
\nthere are lot of hidden tricks and some are not disclosed.
it is not as simple as running their open source software or server to get results of CASP16
", "rawMarkdown": "please read the vfold presentation slide/method abstract and video for CASP16.\nthere are lot of hidden tricks and some are not disclosed.\n\nit is not as simple as running their open source software or server to get results of CASP16", "votes": null }, { "id": "3174910", "postDate": "04/09/2025 15:02:18", "content": "I have also run Vfold locally, but for RNAs longer than 100 nucleotides, the processing speed is very slow, taking over an hour.
\nFor R1108, I encountered the issue \"Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" with a **length of 69 nucleotides. **
\nFor R1126, the problem was \"ERROR: 6-way 0 0 2 1 1 2 chain A 9 A 10 A 27 A 28 A 72 A 75 A 299 A 301 A 333 A 335 A 352 A 355 sequence GC GA UGCC GCA UGG CGCC can not find template. Can not find viable templates for (at least) one of motifs,\" with a** length of 363 nucleotides.**
\nFor R1128, it seems unable to predict the secondary structure, with a length of 238 nucleotides. For R1136, the issue was \"ERROR: 6-way 1 1 2 1 1 2 chain A 21 A 23 A 42 A 44 A 66 A 69 A 91 A 93 A 328 A 330 A 351 A 354 sequence ACG UGA UGCC GCU AGG CGCU can not find template. Can not find viable templates for (at least) one of motifs,\" with a length of 374 nucleotides.
\nFor R1138, the problem is also the inability to obtain a secondary structure, with a **length of 720 nucleotides.
\n**
\n For R1189, the issue was \"ERROR: Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" with a length of 118 nucleotides, and the same applies to R1190 with the same error \"ERROR: Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" also with a** length of 118 nucleotides.**
It appears that Vfold struggles to provide secondary structures for particularly long RNAs, and for some specific RNAs, it cannot generate tertiary structures either. Has anyone else encountered similar issues? Moreover, the majority of target RNAs in CASP16 are longer than 200 nucleotides, yet Vfold managed to top the leaderboard, which seems quite different from the performance observed in local runs. Could it be that the Vfold used in CASP16 differs from the version we downloaded from https://rna.physics.missouri.edu/vfoldPipeline/index.html, possibly an updated version? When running locally, longer RNAs encounter RAM insufficiency issues, but even after allocating sufficient RAM, the final 3D structure still cannot be generated, and there are no CPU performance limitations.
", "rawMarkdown": "I have also run Vfold locally, but for RNAs longer than 100 nucleotides, the processing speed is very slow, taking over an hour. \n\n For R1108, I encountered the issue \"Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" with a **length of 69 nucleotides. **\n\n For R1126, the problem was \"ERROR: 6-way 0 0 2 1 1 2 chain A 9 A 10 A 27 A 28 A 72 A 75 A 299 A 301 A 333 A 335 A 352 A 355 sequence GC GA UGCC GCA UGG CGCC can not find template. Can not find viable templates for (at least) one of motifs,\" with a** length of 363 nucleotides.**\n\n For R1128, it seems unable to predict the secondary structure, with a length of 238 nucleotides. For R1136, the issue was \"ERROR: 6-way 1 1 2 1 1 2 chain A 21 A 23 A 42 A 44 A 66 A 69 A 91 A 93 A 328 A 330 A 351 A 354 sequence ACG UGA UGCC GCU AGG CGCU can not find template. Can not find viable templates for (at least) one of motifs,\" with a **length of 374 nucleotides.**\n\n For R1138, the problem is also the inability to obtain a secondary structure, with a **length of 720 nucleotides.\n**\n For R1189, the issue was \"ERROR: Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" with a length of 118 nucleotides, and the same applies to R1190 with the same error \"ERROR: Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" also with a** length of 118 nucleotides.**\n\n It appears that Vfold struggles to provide secondary structures for particularly long RNAs, and for some specific RNAs, it cannot generate tertiary structures either. Has anyone else encountered similar issues? Moreover, the majority of target RNAs in CASP16 are longer than 200 nucleotides, yet Vfold managed to top the leaderboard, which seems quite different from the performance observed in local runs. Could it be that the Vfold used in CASP16 differs from the version we downloaded from https://rna.physics.missouri.edu/vfoldPipeline/index.html, possibly an updated version? When running locally, longer RNAs encounter RAM insufficiency issues, but even after allocating sufficient RAM, the final 3D structure still cannot be generated, and there are no CPU performance limitations.", "votes": null }, { "id": "3175439", "postDate": "04/10/2025 06:20:02", "content": "Thanks for your update, did you use the 2d structure generated by vfold-pipeline and then use it in the RNAJP or IsRNA to generate 3d coor?
", "rawMarkdown": "Thanks for your update, did you use the 2d structure generated by vfold-pipeline and then use it in the RNAJP or IsRNA to generate 3d coor?", "votes": null } ], "comments": [ { "id": 3156616, "author_name": "ignasialemany", "author_url": "", "post_date": "03/22/2025 10:36:43", "content": "I thought Vfold was a template-based model?
", "votes": null, "replies": [ { "id": 3156759, "author_name": "deepauxilary", "author_url": "", "post_date": "03/22/2025 13:59:28", "content": "Yes and no.
\n
According to https://vfold.missouri.edu/research.html,
\n• Vfold2D [1-4] a physics-based model that employs the statistics of known structures to compute energy parameters, enabling the prediction of 2D structures with pseudoknots and non-canonical base pairs
\n• VfoldMCPX [5-6] a physics-based model designed to predict the structure of multi-strand RNA assembly.
\n• Vfold3D [1] a motif template-based model
\n• VfoldLA [2] a loop template-based model
\n• IsRNA [3-5], a model based on statistical potentials and coarse-grained molecular dynamics
", "votes": null, "replies": [] } ] }, { "id": 3156736, "author_name": "hengck23", "author_url": "", "post_date": "03/22/2025 13:25:20", "content": "Not sure if your results is correct. You should use results downloaded as reference. Further please consult vfold casp presentation material. Think it is a combinations of methods and tool.
\nNot sure casp submission is pure server results
", "votes": null, "replies": [ { "id": 3156744, "author_name": "hengck23", "author_url": "", "post_date": "03/22/2025 13:34:46", "content": "\nPresentation
\nhttps://predictioncenter.org/casp16/doc/CASP16_Abstracts.pdf
\nMethod abstract
\n\"We use various methods, including Vfold-Pipeline1, IsRNA2,3,4, and RNAJP5, to generate RNA 3D
\nstructures from the sequences. If structural templates for specific motifs in the RNA targets are identified
\nin the PDB database, we incorporate them as folding constraints. For protein/RNA/DNA complexes, RNA
\nwith ligands, and RNA with solvent shell, we employ AlphaFold36, ITScore7,8, and AMBER9 to generate
\nstructural candidates.\"
there is only one automatic top method:
\nhttps://predictioncenter.org/casp16/doc/presentations/Day-3/Day3-01-Kretsch_CASP16_NA_Assessement_PuntaCana_RCK_v1.pptx
in my experiments, i select DL automatic methods and combine them.the combined version probbaly can beat vfold (even though each of the DL methods may be worse).
\npromising combinations are from af3, drfold2, rhofold+, nufold, trRoesetta2
\n
Given that Vfold is much higher than the second place's Zscore, it's unfortunate that we can't distill from vfold. I wonder what kind of human intervention was used in their version.
\n
\n(source: https://predictioncenter.org/casp16/zscores_rna.cgi)
This is z score not tm score. You need to remake the ranking from tm score download at casp16
", "votes": null, "replies": [ { "id": 3156836, "author_name": "hengck23", "author_url": "", "post_date": "03/22/2025 16:03:28", "content": "tm score ranking is here
\nhttps://www.kaggle.com/competitions/stanford-rna-3d-folding/discussion/566906#3155466
Did u solve the problem? I also try u using vfold to cal the tm-score of R1107, the score is similar to yours, which is around 0.210
", "votes": null, "replies": [] }, { "id": 3174664, "author_name": "hengck23", "author_url": "", "post_date": "04/09/2025 10:18:16", "content": "please read the vfold presentation slide/method abstract and video for CASP16.
\nthere are lot of hidden tricks and some are not disclosed.
it is not as simple as running their open source software or server to get results of CASP16
", "votes": null, "replies": [ { "id": 3175439, "author_name": "akomwins", "author_url": "", "post_date": "04/10/2025 06:20:02", "content": "Thanks for your update, did you use the 2d structure generated by vfold-pipeline and then use it in the RNAJP or IsRNA to generate 3d coor?
", "votes": null, "replies": [] } ] }, { "id": 3174910, "author_name": "yuanlao153", "author_url": "", "post_date": "04/09/2025 15:02:18", "content": "I have also run Vfold locally, but for RNAs longer than 100 nucleotides, the processing speed is very slow, taking over an hour.
\nFor R1108, I encountered the issue \"Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" with a **length of 69 nucleotides. **
\nFor R1126, the problem was \"ERROR: 6-way 0 0 2 1 1 2 chain A 9 A 10 A 27 A 28 A 72 A 75 A 299 A 301 A 333 A 335 A 352 A 355 sequence GC GA UGCC GCA UGG CGCC can not find template. Can not find viable templates for (at least) one of motifs,\" with a** length of 363 nucleotides.**
\nFor R1128, it seems unable to predict the secondary structure, with a length of 238 nucleotides. For R1136, the issue was \"ERROR: 6-way 1 1 2 1 1 2 chain A 21 A 23 A 42 A 44 A 66 A 69 A 91 A 93 A 328 A 330 A 351 A 354 sequence ACG UGA UGCC GCU AGG CGCU can not find template. Can not find viable templates for (at least) one of motifs,\" with a length of 374 nucleotides.
\nFor R1138, the problem is also the inability to obtain a secondary structure, with a **length of 720 nucleotides.
\n**
\n For R1189, the issue was \"ERROR: Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" with a length of 118 nucleotides, and the same applies to R1190 with the same error \"ERROR: Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" also with a** length of 118 nucleotides.**
It appears that Vfold struggles to provide secondary structures for particularly long RNAs, and for some specific RNAs, it cannot generate tertiary structures either. Has anyone else encountered similar issues? Moreover, the majority of target RNAs in CASP16 are longer than 200 nucleotides, yet Vfold managed to top the leaderboard, which seems quite different from the performance observed in local runs. Could it be that the Vfold used in CASP16 differs from the version we downloaded from https://rna.physics.missouri.edu/vfoldPipeline/index.html, possibly an updated version? When running locally, longer RNAs encounter RAM insufficiency issues, but even after allocating sufficient RAM, the final 3D structure still cannot be generated, and there are no CPU performance limitations.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3156574": "Anyone ran [Vfold Pipeline](https://rna.physics.missouri.edu/vfoldPipeline/) locally?\n\nI think\n- The gap between my local vfold and the vfold expert on the leader board is so large.\n- It's very slow to run. Some targets need 2h. Is there a GPU version of Vfold?\n- Some targets are failed. It looks related to the sequence length but I am not sure.\n\n\n", "3156616": "I thought Vfold was a template-based model?", "3156736": "Not sure if your results is correct. You should use results downloaded as reference. Further please consult vfold casp presentation material. Think it is a combinations of methods and tool.\n\nNot sure casp submission is pure server results", "3156744": "https://predictioncenter.org/casp16/doc/presentations/Day-3/Day3-03-Chen-Vfold-RNA-Predictor-Talk1_Redacted.pdf\n\nPresentation \n\nhttps://predictioncenter.org/casp16/doc/CASP16_Abstracts.pdf\n\nMethod abstract \n\n\"We use various methods, including Vfold-Pipeline1, IsRNA2,3,4, and RNAJP5, to generate RNA 3D\nstructures from the sequences. If structural templates for specific motifs in the RNA targets are identified\nin the PDB database, we incorporate them as folding constraints. For protein/RNA/DNA complexes, RNA\nwith ligands, and RNA with solvent shell, we employ AlphaFold36, ITScore7,8, and AMBER9 to generate\nstructural candidates.\"", "3156751": "there is only one automatic top method:\nhttps://predictioncenter.org/casp16/doc/presentations/Day-3/Day3-01-Kretsch_CASP16_NA_Assessement_PuntaCana_RCK_v1.pptx\n\n\n\nin my experiments, i select DL automatic methods and combine them.the combined version probbaly can beat vfold (even though each of the DL methods may be worse).\n\npromising combinations are from af3, drfold2, rhofold+, nufold, trRoesetta2\n\n", "3156759": "Yes and no.\n\n\n\nAccording to https://vfold.missouri.edu/research.html, \n\n • Vfold2D [1-4] a physics-based model that employs the statistics of known structures to compute energy parameters, enabling the prediction of 2D structures with pseudoknots and non-canonical base pairs\n\n• VfoldMCPX [5-6] a physics-based model designed to predict the structure of multi-strand RNA assembly.\n\n• Vfold3D [1] a motif template-based model\n\n• VfoldLA [2] a loop template-based model\n\n• IsRNA [3-5], a model based on statistical potentials and coarse-grained molecular dynamics", "3156830": "Given that Vfold is much higher than the second place's Zscore, it's unfortunate that we can't distill from vfold. I wonder what kind of human intervention was used in their version.\n\n\n(source: https://predictioncenter.org/casp16/zscores_rna.cgi)", "3156832": "This is z score not tm score. You need to remake the ranking from tm score download at casp16", "3156836": "tm score ranking is here\nhttps://www.kaggle.com/competitions/stanford-rna-3d-folding/discussion/566906#3155466", "3174649": "Did u solve the problem? I also try u using vfold to cal the tm-score of R1107, the score is similar to yours, which is around 0.210", "3174664": "please read the vfold presentation slide/method abstract and video for CASP16.\nthere are lot of hidden tricks and some are not disclosed.\n\nit is not as simple as running their open source software or server to get results of CASP16", "3174910": "I have also run Vfold locally, but for RNAs longer than 100 nucleotides, the processing speed is very slow, taking over an hour. \n\n For R1108, I encountered the issue \"Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" with a **length of 69 nucleotides. **\n\n For R1126, the problem was \"ERROR: 6-way 0 0 2 1 1 2 chain A 9 A 10 A 27 A 28 A 72 A 75 A 299 A 301 A 333 A 335 A 352 A 355 sequence GC GA UGCC GCA UGG CGCC can not find template. Can not find viable templates for (at least) one of motifs,\" with a** length of 363 nucleotides.**\n\n For R1128, it seems unable to predict the secondary structure, with a length of 238 nucleotides. For R1136, the issue was \"ERROR: 6-way 1 1 2 1 1 2 chain A 21 A 23 A 42 A 44 A 66 A 69 A 91 A 93 A 328 A 330 A 351 A 354 sequence ACG UGA UGCC GCU AGG CGCU can not find template. Can not find viable templates for (at least) one of motifs,\" with a **length of 374 nucleotides.**\n\n For R1138, the problem is also the inability to obtain a secondary structure, with a **length of 720 nucleotides.\n**\n For R1189, the issue was \"ERROR: Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" with a length of 118 nucleotides, and the same applies to R1190 with the same error \"ERROR: Input 2D structure contains (Vfold3D) undefined motifs: A open loop,\" also with a** length of 118 nucleotides.**\n\n It appears that Vfold struggles to provide secondary structures for particularly long RNAs, and for some specific RNAs, it cannot generate tertiary structures either. Has anyone else encountered similar issues? Moreover, the majority of target RNAs in CASP16 are longer than 200 nucleotides, yet Vfold managed to top the leaderboard, which seems quite different from the performance observed in local runs. Could it be that the Vfold used in CASP16 differs from the version we downloaded from https://rna.physics.missouri.edu/vfoldPipeline/index.html, possibly an updated version? When running locally, longer RNAs encounter RAM insufficiency issues, but even after allocating sufficient RAM, the final 3D structure still cannot be generated, and there are no CPU performance limitations.", "3175439": "Thanks for your update, did you use the 2d structure generated by vfold-pipeline and then use it in the RNAJP or IsRNA to generate 3d coor?" }, "source": "meta" }