{
  "id": 566232,
  "title": "how to handle different rna conformational states in pdb training data?",
  "url": "/competitions/stanford-rna-3d-folding/discussion/566232",
  "author_name": "",
  "post_date": "2025-03-04T13:36:36.445476600Z",
  "votes": 5,
  "comment_count": 3,
  "views": 0,
  "content": "<p>i am biomedical expert, so my knowledge come from the internet and chatgpt (who may hallucinate).<br>\ni want to get more data from pdb bank, so i download the 1scl.pdb. there are 6 models(structure) in this pdb file. after comparing with kaggle train_labels.csv, i find that kaggle 1SCL_A label correspond to MODEL 6.</p>\n<p>I thought the lowest energy conformation is MODEL 1? <br>\nand i though most deep learning model (like alphafold) will predict  lowest energy conformation?</p>",
  "messages": [
    {
      "id": "3140310",
      "postDate": "03/04/2025 13:36:36",
      "content": "<p>i am biomedical expert, so my knowledge come from the internet and chatgpt (who may hallucinate).<br>\ni want to get more data from pdb bank, so i download the 1scl.pdb. there are 6 models(structure) in this pdb file. after comparing with kaggle train_labels.csv, i find that kaggle 1SCL_A label correspond to MODEL 6.</p>\n<p>I thought the lowest energy conformation is MODEL 1? <br>\nand i though most deep learning model (like alphafold) will predict  lowest energy conformation?</p>",
      "rawMarkdown": "i am biomedical expert, so my knowledge come from the internet and chatgpt (who may hallucinate).\ni want to get more data from pdb bank, so i download the 1scl.pdb. there are 6 models(structure) in this pdb file. after comparing with kaggle train\\_labels.csv, i find that kaggle 1SCL\\_A label correspond to MODEL 6.\n\nI thought the lowest energy conformation is MODEL 1? \nand i though most deep learning model (like alphafold) will predict  lowest energy conformation?",
      "votes": null
    },
    {
      "id": "3140593",
      "postDate": "03/04/2025 18:48:42",
      "content": "<blockquote>\n  <p>I thought the lowest energy conformation is MODEL 1?</p>\n</blockquote>\n<p>In most NMR structure files model #1 will be the one with lowest energy, fewest restraint violations, etc. Yet that's done out of convenience rather than because of some rule. The best way to figure this out is to read through the <code>REMARK</code> lines.</p>\n<p>For this structure (1SCL) there is a <code>REMARK</code> line that says:</p>\n<pre><code>\n</code></pre>\n<p>Either they picked a random model, or maybe they always extract the last conformer in an ensemble.</p>",
      "rawMarkdown": "> I thought the lowest energy conformation is MODEL 1?\n\nIn most NMR structure files model #1 will be the one with lowest energy, fewest restraint violations, etc. Yet that's done out of convenience rather than because of some rule. The best way to figure this out is to read through the `REMARK` lines.\n\nFor this structure (1SCL) there is a `REMARK` line that says:\n\n    REMARK 210 BEST REPRESENTATIVE CONFORMER IN THIS ENSEMBLE : NULL\n\nEither they picked a random model, or maybe they always extract the last conformer in an ensemble.",
      "votes": null
    },
    {
      "id": "3140598",
      "postDate": "03/04/2025 18:52:19",
      "content": "<blockquote>\n  <p>and i though most deep learning model (like alphafold) will predict lowest energy conformation?</p>\n</blockquote>\n<p>They will pick the lowest energy (or highest average pLDDT score) from a set of different trajectories. There is a stochastic component in modeling which makes it impossible to know whether the first model (or any other) will be the best. By repeating the modeling several times and picking the best one we increase the chance of getting a model close to the global energy minimum. Still no guarantee, though.</p>",
      "rawMarkdown": "> and i though most deep learning model (like alphafold) will predict lowest energy conformation?\n\nThey will pick the lowest energy (or highest average pLDDT score) from a set of different trajectories. There is a stochastic component in modeling which makes it impossible to know whether the first model (or any other) will be the best. By repeating the modeling several times and picking the best one we increase the chance of getting a model close to the global energy minimum. Still no guarantee, though.",
      "votes": null
    },
    {
      "id": "3143556",
      "postDate": "03/07/2025 10:31:15",
      "content": "<p>thanks for the answer.</p>",
      "rawMarkdown": "thanks for the answer.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3140593,
      "author_name": "tilii7",
      "author_url": "",
      "post_date": "03/04/2025 18:48:42",
      "content": "<blockquote>\n  <p>I thought the lowest energy conformation is MODEL 1?</p>\n</blockquote>\n<p>In most NMR structure files model #1 will be the one with lowest energy, fewest restraint violations, etc. Yet that's done out of convenience rather than because of some rule. The best way to figure this out is to read through the <code>REMARK</code> lines.</p>\n<p>For this structure (1SCL) there is a <code>REMARK</code> line that says:</p>\n<pre><code>\n</code></pre>\n<p>Either they picked a random model, or maybe they always extract the last conformer in an ensemble.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3143556,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "03/07/2025 10:31:15",
          "content": "<p>thanks for the answer.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3140598,
      "author_name": "tilii7",
      "author_url": "",
      "post_date": "03/04/2025 18:52:19",
      "content": "<blockquote>\n  <p>and i though most deep learning model (like alphafold) will predict lowest energy conformation?</p>\n</blockquote>\n<p>They will pick the lowest energy (or highest average pLDDT score) from a set of different trajectories. There is a stochastic component in modeling which makes it impossible to know whether the first model (or any other) will be the best. By repeating the modeling several times and picking the best one we increase the chance of getting a model close to the global energy minimum. Still no guarantee, though.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3140310": "i am biomedical expert, so my knowledge come from the internet and chatgpt (who may hallucinate).\ni want to get more data from pdb bank, so i download the 1scl.pdb. there are 6 models(structure) in this pdb file. after comparing with kaggle train\\_labels.csv, i find that kaggle 1SCL\\_A label correspond to MODEL 6.\n\nI thought the lowest energy conformation is MODEL 1? \nand i though most deep learning model (like alphafold) will predict  lowest energy conformation?",
    "3140593": "> I thought the lowest energy conformation is MODEL 1?\n\nIn most NMR structure files model #1 will be the one with lowest energy, fewest restraint violations, etc. Yet that's done out of convenience rather than because of some rule. The best way to figure this out is to read through the `REMARK` lines.\n\nFor this structure (1SCL) there is a `REMARK` line that says:\n\n    REMARK 210 BEST REPRESENTATIVE CONFORMER IN THIS ENSEMBLE : NULL\n\nEither they picked a random model, or maybe they always extract the last conformer in an ensemble.",
    "3140598": "> and i though most deep learning model (like alphafold) will predict lowest energy conformation?\n\nThey will pick the lowest energy (or highest average pLDDT score) from a set of different trajectories. There is a stochastic component in modeling which makes it impossible to know whether the first model (or any other) will be the best. By repeating the modeling several times and picking the best one we increase the chance of getting a model close to the global energy minimum. Still no guarantee, though.",
    "3143556": "thanks for the answer."
  },
  "source": "meta"
}