{
  "id": 573317,
  "title": "Training labels / Multiple conformations",
  "url": "/competitions/stanford-rna-3d-folding/discussion/573317",
  "author_name": "",
  "post_date": "2025-04-14T20:10:23.952022800Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I've noticed that where the PDB entity has multiple conformations, the host always uses the last-indexed model. Say, for 1HLX/A - the model number 20 out of 20 available.</p>\n<p>I might be mistaken, but it's not always the \"default\" (MFE-based) model picked by PDB nor a model selected by the authors.</p>\n<p>I just wonder if there is any particular rationale behind such choice? Am I missing something?</p>",
  "messages": [
    {
      "id": "3178944",
      "postDate": "04/14/2025 20:10:23",
      "content": "<p>I've noticed that where the PDB entity has multiple conformations, the host always uses the last-indexed model. Say, for 1HLX/A - the model number 20 out of 20 available.</p>\n<p>I might be mistaken, but it's not always the \"default\" (MFE-based) model picked by PDB nor a model selected by the authors.</p>\n<p>I just wonder if there is any particular rationale behind such choice? Am I missing something?</p>",
      "rawMarkdown": "I've noticed that where the PDB entity has multiple conformations, the host always uses the last-indexed model. Say, for 1HLX/A - the model number 20 out of 20 available.\n\nI might be mistaken, but it's not always the \"default\" (MFE-based) model picked by PDB nor a model selected by the authors.\n\nI just wonder if there is any particular rationale behind such choice? Am I missing something?",
      "votes": null
    },
    {
      "id": "3181521",
      "postDate": "04/18/2025 02:08:31",
      "content": "<p>use DETR  dynamic matching way to train your models with multiple ground truth.</p>\n<p>training loop</p>\n<ol>\n<li>forward pass : structure = model(seq, …)</li>\n<li>dynamic match: best truth = find best match (structures, a set of conformations)</li>\n<li>backward pass: loss = compaute loss (best truth,structure)</li>\n</ol>\n<p>just need to care that the model can diversity, i.e. it don't collapse to \"few SAME conformations\"</p>",
      "rawMarkdown": "use DETR  dynamic matching way to train your models with multiple ground truth.\n\ntraining loop\n1. forward pass : structure = model(seq, ...)\n2. dynamic match: best truth = find best match (structures, a set of conformations)\n3. backward pass: loss = compaute loss (best truth,structure)\n\njust need to care that the model can diversity, i.e. it don't collapse to \"few SAME conformations\"",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3181521,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "04/18/2025 02:08:31",
      "content": "<p>use DETR  dynamic matching way to train your models with multiple ground truth.</p>\n<p>training loop</p>\n<ol>\n<li>forward pass : structure = model(seq, …)</li>\n<li>dynamic match: best truth = find best match (structures, a set of conformations)</li>\n<li>backward pass: loss = compaute loss (best truth,structure)</li>\n</ol>\n<p>just need to care that the model can diversity, i.e. it don't collapse to \"few SAME conformations\"</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3178944": "I've noticed that where the PDB entity has multiple conformations, the host always uses the last-indexed model. Say, for 1HLX/A - the model number 20 out of 20 available.\n\nI might be mistaken, but it's not always the \"default\" (MFE-based) model picked by PDB nor a model selected by the authors.\n\nI just wonder if there is any particular rationale behind such choice? Am I missing something?",
    "3181521": "use DETR  dynamic matching way to train your models with multiple ground truth.\n\ntraining loop\n1. forward pass : structure = model(seq, ...)\n2. dynamic match: best truth = find best match (structures, a set of conformations)\n3. backward pass: loss = compaute loss (best truth,structure)\n\njust need to care that the model can diversity, i.e. it don't collapse to \"few SAME conformations\""
  },
  "source": "meta"
}