{
  "id": 568743,
  "title": "Why temporal cutoff and not normal train test split",
  "url": "/competitions/stanford-rna-3d-folding/discussion/568743",
  "author_name": "",
  "post_date": "2025-03-17T18:40:05.683016300Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>For a problem like RNA 3D folding prediction, I do not understand why we are doing train / test split based on temporal cutoff. I saw the tutorial and it says that how it's done in structural biology. But I did not got the intuition. I mean would't giving an incomplete sequence would produce incomplete results? </p>\n<p>And why we can't just do simple train test split. I mean, let's say take the first 100 rows as train and last 10 as the test out of 110 total samples. Why is't this the case?</p>",
  "messages": [
    {
      "id": "3152382",
      "postDate": "03/17/2025 18:40:05",
      "content": "<p>For a problem like RNA 3D folding prediction, I do not understand why we are doing train / test split based on temporal cutoff. I saw the tutorial and it says that how it's done in structural biology. But I did not got the intuition. I mean would't giving an incomplete sequence would produce incomplete results? </p>\n<p>And why we can't just do simple train test split. I mean, let's say take the first 100 rows as train and last 10 as the test out of 110 total samples. Why is't this the case?</p>",
      "rawMarkdown": "For a problem like RNA 3D folding prediction, I do not understand why we are doing train / test split based on temporal cutoff. I saw the tutorial and it says that how it's done in structural biology. But I did not got the intuition. I mean would't giving an incomplete sequence would produce incomplete results? \n\nAnd why we can't just do simple train test split. I mean, let's say take the first 100 rows as train and last 10 as the test out of 110 total samples. Why is't this the case?",
      "votes": null
    },
    {
      "id": "3152412",
      "postDate": "03/17/2025 19:17:21",
      "content": "<p>The temporal cutoff isn't truncating the sequence, it's cutting off the data to separate sequences for which the 3d structure has been experimentally determined before/after a given date. It makes sense to do this given that the purpose of the competition is to generate a model that can generalize to unseen sequences so their 3d structure doesn't have to be experimentally obtained <em>in the future</em>.</p>",
      "rawMarkdown": "The temporal cutoff isn't truncating the sequence, it's cutting off the data to separate sequences for which the 3d structure has been experimentally determined before/after a given date. It makes sense to do this given that the purpose of the competition is to generate a model that can generalize to unseen sequences so their 3d structure doesn't have to be experimentally obtained *in the future*.",
      "votes": null
    },
    {
      "id": "3152423",
      "postDate": "03/17/2025 19:34:13",
      "content": "<p>Ahh I see, sorry my bad then, thanks for clarification </p>",
      "rawMarkdown": "Ahh I see, sorry my bad then, thanks for clarification",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3152412,
      "author_name": "asarvazyan",
      "author_url": "",
      "post_date": "03/17/2025 19:17:21",
      "content": "<p>The temporal cutoff isn't truncating the sequence, it's cutting off the data to separate sequences for which the 3d structure has been experimentally determined before/after a given date. It makes sense to do this given that the purpose of the competition is to generate a model that can generalize to unseen sequences so their 3d structure doesn't have to be experimentally obtained <em>in the future</em>.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3152423,
          "author_name": "beatsonflow",
          "author_url": "",
          "post_date": "03/17/2025 19:34:13",
          "content": "<p>Ahh I see, sorry my bad then, thanks for clarification </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3152382": "For a problem like RNA 3D folding prediction, I do not understand why we are doing train / test split based on temporal cutoff. I saw the tutorial and it says that how it's done in structural biology. But I did not got the intuition. I mean would't giving an incomplete sequence would produce incomplete results? \n\nAnd why we can't just do simple train test split. I mean, let's say take the first 100 rows as train and last 10 as the test out of 110 total samples. Why is't this the case?",
    "3152412": "The temporal cutoff isn't truncating the sequence, it's cutting off the data to separate sequences for which the 3d structure has been experimentally determined before/after a given date. It makes sense to do this given that the purpose of the competition is to generate a model that can generalize to unseen sequences so their 3d structure doesn't have to be experimentally obtained *in the future*.",
    "3152423": "Ahh I see, sorry my bad then, thanks for clarification"
  },
  "source": "meta"
}