If we want to use PDB files in our solution how would this work on the test set as I'm assuming they won't be available anywhere. Would we be expected to calculate these ourselves? Is this feasible computationally?
", "messages": [ { "id": "3383442", "postDate": "12/30/2025 06:48:07", "content": "If we want to use PDB files in our solution how would this work on the test set as I'm assuming they won't be available anywhere. Would we be expected to calculate these ourselves? Is this feasible computationally?
", "rawMarkdown": "If we want to use PDB files in our solution how would this work on the test set as I'm assuming they won't be available anywhere. Would we be expected to calculate these ourselves? Is this feasible computationally?", "votes": null }, { "id": "3383647", "postDate": "12/30/2025 15:02:49", "content": "If you mean adding structural information, you should distinguish between computational predictions of structure, such as those made by AlphaFold, and experimentally generated structures obtained via X-ray crystallography or NMR. The former are predictions with varying levels of confidence along the structure; the latter are true structures (with some known margin of error in the location of the atoms). So the quality of the structural data you use is important.
\nIf you want to use AlphaFold predictions, you should note that they vary in reliability along the sequence. That said, AlphaFoldDB has many SwissProt predictions. The Protein Data Bank has experimentally derived structures. You can also use simple homology modeling to derive structures using templates from AlphaFoldDB or PDB, keeping in mind the reliability of the templates in those respective databases.
\nYou can also run AlphaFold on all the sequences you want to, but that will be much more computationallly expensive, and my not result in what you actually need.
", "rawMarkdown": "If you mean adding structural information, you should distinguish between computational predictions of structure, such as those made by AlphaFold, and experimentally generated structures obtained via X-ray crystallography or NMR. The former are predictions with varying levels of confidence along the structure; the latter are true structures (with some known margin of error in the location of the atoms). So the quality of the structural data you use is important.\n\nIf you want to use AlphaFold predictions, you should note that they vary in reliability along the sequence. That said, [AlphaFoldDB ](https://alphafold.ebi.ac.uk/) has many SwissProt predictions. The [Protein Data Bank](https://www.rcsb.org/) has experimentally derived structures. You can also use simple [homology modeling](https://proteopedia.org/w/Practical_Guide_to_Homology_Modeling) to derive structures using templates from AlphaFoldDB or PDB, keeping in mind the reliability of the templates in those respective databases.\n\n You can also run AlphaFold on all the sequences you want to, but that will be much more computationallly expensive, and my not result in what you actually need.", "votes": null }, { "id": "3383698", "postDate": "12/30/2025 17:05:03", "content": "I am looking for AF2 PDBs for the test proteins since they will cover the full protein length and I can still do plddt filtering.
\nI know there are AF2 PDBs for cannonical proteins and most Swissprot ones - but I guess the spirit of the question is if these will be available for the test test. Because if I build a pipeline based on the training set where I have most PDB files it kind of becomes useless if I can't pipe structures for the new ones - assuming a simple setup.
", "rawMarkdown": "I am looking for AF2 PDBs for the test proteins since they will cover the full protein length and I can still do plddt filtering.\n\nI know there are AF2 PDBs for cannonical proteins and most Swissprot ones - but I guess the spirit of the question is if these will be available for the test test. Because if I build a pipeline based on the training set where I have most PDB files it kind of becomes useless if I can't pipe structures for the new ones - assuming a simple setup.", "votes": null }, { "id": "3383715", "postDate": "12/30/2025 17:51:39", "content": "The test set will be a subset of the proteins listed in testsuperset.fasta. You can get AF2 PDBs (or experimental structures) for all of these. That is the only way to guarantee coverage on the final test set. We won't know which proteins from the test superset will eventually be in the final test set a priori. That's the fun (and challenge) of the prospective evaluation.
", "rawMarkdown": "The test set will be a subset of the proteins listed in testsuperset.fasta. You can get AF2 PDBs (or experimental structures) for all of these. That is the only way to guarantee coverage on the final test set. We won't know which proteins from the test superset will eventually be in the final test set *a priori*. That's the fun (and challenge) of the prospective evaluation.", "votes": null }, { "id": "3384021", "postDate": "12/31/2025 10:14:45", "content": "Thanks. Helpful.
", "rawMarkdown": "Thanks. Helpful.", "votes": null }, { "id": "3385827", "postDate": "01/04/2026 05:15:53", "content": "I'm also studying this recently…
", "rawMarkdown": "I'm also studying this recently...", "votes": null } ], "comments": [ { "id": 3383647, "author_name": "iddux1", "author_url": "", "post_date": "12/30/2025 15:02:49", "content": "If you mean adding structural information, you should distinguish between computational predictions of structure, such as those made by AlphaFold, and experimentally generated structures obtained via X-ray crystallography or NMR. The former are predictions with varying levels of confidence along the structure; the latter are true structures (with some known margin of error in the location of the atoms). So the quality of the structural data you use is important.
\nIf you want to use AlphaFold predictions, you should note that they vary in reliability along the sequence. That said, AlphaFoldDB has many SwissProt predictions. The Protein Data Bank has experimentally derived structures. You can also use simple homology modeling to derive structures using templates from AlphaFoldDB or PDB, keeping in mind the reliability of the templates in those respective databases.
\nYou can also run AlphaFold on all the sequences you want to, but that will be much more computationallly expensive, and my not result in what you actually need.
", "votes": null, "replies": [ { "id": 3383698, "author_name": "glennkroegel", "author_url": "", "post_date": "12/30/2025 17:05:03", "content": "I am looking for AF2 PDBs for the test proteins since they will cover the full protein length and I can still do plddt filtering.
\nI know there are AF2 PDBs for cannonical proteins and most Swissprot ones - but I guess the spirit of the question is if these will be available for the test test. Because if I build a pipeline based on the training set where I have most PDB files it kind of becomes useless if I can't pipe structures for the new ones - assuming a simple setup.
", "votes": null, "replies": [ { "id": 3383715, "author_name": "claradepaolis", "author_url": "", "post_date": "12/30/2025 17:51:39", "content": "The test set will be a subset of the proteins listed in testsuperset.fasta. You can get AF2 PDBs (or experimental structures) for all of these. That is the only way to guarantee coverage on the final test set. We won't know which proteins from the test superset will eventually be in the final test set a priori. That's the fun (and challenge) of the prospective evaluation.
", "votes": null, "replies": [ { "id": 3384021, "author_name": "glennkroegel", "author_url": "", "post_date": "12/31/2025 10:14:45", "content": "Thanks. Helpful.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3385827, "author_name": "sixjune", "author_url": "", "post_date": "01/04/2026 05:15:53", "content": "I'm also studying this recently…
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3383442": "If we want to use PDB files in our solution how would this work on the test set as I'm assuming they won't be available anywhere. Would we be expected to calculate these ourselves? Is this feasible computationally?", "3383647": "If you mean adding structural information, you should distinguish between computational predictions of structure, such as those made by AlphaFold, and experimentally generated structures obtained via X-ray crystallography or NMR. The former are predictions with varying levels of confidence along the structure; the latter are true structures (with some known margin of error in the location of the atoms). So the quality of the structural data you use is important.\n\nIf you want to use AlphaFold predictions, you should note that they vary in reliability along the sequence. That said, [AlphaFoldDB ](https://alphafold.ebi.ac.uk/) has many SwissProt predictions. The [Protein Data Bank](https://www.rcsb.org/) has experimentally derived structures. You can also use simple [homology modeling](https://proteopedia.org/w/Practical_Guide_to_Homology_Modeling) to derive structures using templates from AlphaFoldDB or PDB, keeping in mind the reliability of the templates in those respective databases.\n\n You can also run AlphaFold on all the sequences you want to, but that will be much more computationallly expensive, and my not result in what you actually need.", "3383698": "I am looking for AF2 PDBs for the test proteins since they will cover the full protein length and I can still do plddt filtering.\n\nI know there are AF2 PDBs for cannonical proteins and most Swissprot ones - but I guess the spirit of the question is if these will be available for the test test. Because if I build a pipeline based on the training set where I have most PDB files it kind of becomes useless if I can't pipe structures for the new ones - assuming a simple setup.", "3383715": "The test set will be a subset of the proteins listed in testsuperset.fasta. You can get AF2 PDBs (or experimental structures) for all of these. That is the only way to guarantee coverage on the final test set. We won't know which proteins from the test superset will eventually be in the final test set *a priori*. That's the fun (and challenge) of the prospective evaluation.", "3384021": "Thanks. Helpful.", "3385827": "I'm also studying this recently..." }, "source": "meta" }