As an experiment, I produced distance matrices from some of the labels via scipy.spatial.distance.cdist and reconstructed an alternative set of coordinates via a Multi-Dimensional Scaling algorithm. However, while the labels (understandably) score a perfect 1.0 when evaluated with themselves, the alternative coordinates were often scored < 0.3 by Usalign compared to the original labels.
Is there an explanation for this behaviour? I don't know if MDS occasionally produces mirror-image coordinate sets that would affect the results.
", "messages": [ { "id": "3210645", "postDate": "05/27/2025 13:56:58", "content": "As an experiment, I produced distance matrices from some of the labels via scipy.spatial.distance.cdist and reconstructed an alternative set of coordinates via a Multi-Dimensional Scaling algorithm. However, while the labels (understandably) score a perfect 1.0 when evaluated with themselves, the alternative coordinates were often scored < 0.3 by Usalign compared to the original labels.
Is there an explanation for this behaviour? I don't know if MDS occasionally produces mirror-image coordinate sets that would affect the results.
", "rawMarkdown": "As an experiment, I produced distance matrices from some of the labels via `scipy.spatial.distance.cdist` and reconstructed an alternative set of coordinates via a Multi-Dimensional Scaling algorithm. However, while the labels (understandably) score a perfect 1.0 when evaluated with themselves, the alternative coordinates were often scored < 0.3 by Usalign compared to the original labels.\n\nIs there an explanation for this behaviour? I don't know if MDS occasionally produces mirror-image coordinate sets that would affect the results.", "votes": null }, { "id": "3210895", "postDate": "05/27/2025 20:30:58", "content": "If USalign gives even slightly different scores (beyond numerical imprecision) for two identical coordinates differing only in rotation/translation, then this would be a very problematic bug as far as this competition goes.
\nForget about third-party algorithm softwares. Can you check your before and after coordinates to make absolutely sure they are really identical up to translation/rotation?
", "rawMarkdown": "If USalign gives even slightly different scores (beyond numerical imprecision) for two identical coordinates differing only in rotation/translation, then this would be a very problematic bug as far as this competition goes.\n\nForget about third-party algorithm softwares. Can you check your before and after coordinates to make absolutely sure they are really identical up to translation/rotation?", "votes": null }, { "id": "3210899", "postDate": "05/27/2025 20:36:10", "content": "\n\nI don't know if MDS occasionally produces mirror-image coordinate sets that would affect the results.
\n
Yes, it does. Since DM has not any information about that.
\nPD: No early sharing prize hero yet?
", "rawMarkdown": "> I don't know if MDS occasionally produces mirror-image coordinate sets that would affect the results.\n\nYes, it does. Since DM has not any information about that.\n\nPD: No early sharing prize hero yet?", "votes": null }, { "id": "3211109", "postDate": "05/28/2025 05:03:09", "content": "Thank you for the response; trivially multiplying the coordinates by -1 yields the correct answer in these cases. I thought it might be useful for anyone considering the use of distance-matrix targets rather than raw coordinates, though notably AlphaFold and its variants eschew this approach.
", "rawMarkdown": "Thank you for the response; trivially multiplying the coordinates by -1 yields the correct answer in these cases. I thought it might be useful for anyone considering the use of distance-matrix targets rather than raw coordinates, though notably AlphaFold and its variants eschew this approach.", "votes": null }, { "id": "3211309", "postDate": "05/28/2025 09:47:00", "content": "You're wellcome. Just FYI, only one axis actually needs to be mirrored.
", "rawMarkdown": "You're wellcome. Just FYI, only one axis actually needs to be mirrored.", "votes": null }, { "id": "3395271", "postDate": "01/22/2026 15:40:54", "content": "It is abit too late, but can you share your example?
", "rawMarkdown": "It is abit too late, but can you share your example?", "votes": null } ], "comments": [ { "id": 3210895, "author_name": "revealer", "author_url": "", "post_date": "05/27/2025 20:30:58", "content": "If USalign gives even slightly different scores (beyond numerical imprecision) for two identical coordinates differing only in rotation/translation, then this would be a very problematic bug as far as this competition goes.
\nForget about third-party algorithm softwares. Can you check your before and after coordinates to make absolutely sure they are really identical up to translation/rotation?
", "votes": null, "replies": [] }, { "id": 3210899, "author_name": "sacuscreed", "author_url": "", "post_date": "05/27/2025 20:36:10", "content": "\n\nI don't know if MDS occasionally produces mirror-image coordinate sets that would affect the results.
\n
Yes, it does. Since DM has not any information about that.
\nPD: No early sharing prize hero yet?
", "votes": null, "replies": [ { "id": 3211109, "author_name": "bigironsphere", "author_url": "", "post_date": "05/28/2025 05:03:09", "content": "Thank you for the response; trivially multiplying the coordinates by -1 yields the correct answer in these cases. I thought it might be useful for anyone considering the use of distance-matrix targets rather than raw coordinates, though notably AlphaFold and its variants eschew this approach.
", "votes": null, "replies": [ { "id": 3211309, "author_name": "sacuscreed", "author_url": "", "post_date": "05/28/2025 09:47:00", "content": "You're wellcome. Just FYI, only one axis actually needs to be mirrored.
", "votes": null, "replies": [] } ] } ] }, { "id": 3395271, "author_name": "bridgeport", "author_url": "", "post_date": "01/22/2026 15:40:54", "content": "It is abit too late, but can you share your example?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3210645": "As an experiment, I produced distance matrices from some of the labels via `scipy.spatial.distance.cdist` and reconstructed an alternative set of coordinates via a Multi-Dimensional Scaling algorithm. However, while the labels (understandably) score a perfect 1.0 when evaluated with themselves, the alternative coordinates were often scored < 0.3 by Usalign compared to the original labels.\n\nIs there an explanation for this behaviour? I don't know if MDS occasionally produces mirror-image coordinate sets that would affect the results.", "3210895": "If USalign gives even slightly different scores (beyond numerical imprecision) for two identical coordinates differing only in rotation/translation, then this would be a very problematic bug as far as this competition goes.\n\nForget about third-party algorithm softwares. Can you check your before and after coordinates to make absolutely sure they are really identical up to translation/rotation?", "3210899": "> I don't know if MDS occasionally produces mirror-image coordinate sets that would affect the results.\n\nYes, it does. Since DM has not any information about that.\n\nPD: No early sharing prize hero yet?", "3211109": "Thank you for the response; trivially multiplying the coordinates by -1 yields the correct answer in these cases. I thought it might be useful for anyone considering the use of distance-matrix targets rather than raw coordinates, though notably AlphaFold and its variants eschew this approach.", "3211309": "You're wellcome. Just FYI, only one axis actually needs to be mirrored.", "3395271": "It is abit too late, but can you share your example?" }, "source": "meta" }