{
  "id": 679682,
  "title": "metric rewards luck",
  "url": "/competitions/stanford-rna-3d-folding-2/discussion/679682",
  "author_name": "Taha_Alshatiri",
  "post_date": "2026-03-03T05:45:01.308000",
  "votes": 0,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi, even though the past competition was stable I feel that the idea of the metric choosing the best of 5 predictions  reward luck and stochastic postprocessing so much, is there any better approach for generating 5 predictions rather than  just applying random transformations to each prediction?</p>",
  "messages": [
    {
      "id": 3417716,
      "postDate": "2026-03-06T01:41:53.983Z",
      "content": "<p>You can try combining different methods for the top 5 predictions based on template quality. For instance if for a sequence, the template is poor, you can fallback to a de-novo predictor like Protenix, but if it is of medium quality, you can use top 2 positions for Protenix and bottom 3 for the medium quality template and its transforms. Also, it isn't necessary to use only 1 template, you can use 2 or 3 different templates for diversifying your top 5, and based on quality, choose which ones you apply transformations on and where you place each in the top 5. </p>",
      "rawMarkdown": "You can try combining different methods for the top 5 predictions based on template quality. For instance if for a sequence, the template is poor, you can fallback to a de-novo predictor like Protenix, but if it is of medium quality, you can use top 2 positions for Protenix and bottom 3 for the medium quality template and its transforms. Also, it isn't necessary to use only 1 template, you can use 2 or 3 different templates for diversifying your top 5, and based on quality, choose which ones you apply transformations on and where you place each in the top 5. ",
      "votes": 2
    },
    {
      "id": 3416526,
      "postDate": "2026-03-03T05:45:01.307Z",
      "content": "<p>Hi, even though the past competition was stable I feel that the idea of the metric choosing the best of 5 predictions  reward luck and stochastic postprocessing so much, is there any better approach for generating 5 predictions rather than  just applying random transformations to each prediction?</p>",
      "rawMarkdown": "Hi, even though the past competition was stable I feel that the idea of the metric choosing the best of 5 predictions  reward luck and stochastic postprocessing so much, is there any better approach for generating 5 predictions rather than  just applying random transformations to each prediction?"
    }
  ],
  "comments": [
    {
      "id": 3417716,
      "author_name": "Syed Muhammad Aun Abbas",
      "author_url": "",
      "post_date": "2026-03-06T01:41:53.983000",
      "content": "<p>You can try combining different methods for the top 5 predictions based on template quality. For instance if for a sequence, the template is poor, you can fallback to a de-novo predictor like Protenix, but if it is of medium quality, you can use top 2 positions for Protenix and bottom 3 for the medium quality template and its transforms. Also, it isn't necessary to use only 1 template, you can use 2 or 3 different templates for diversifying your top 5, and based on quality, choose which ones you apply transformations on and where you place each in the top 5. </p>",
      "votes": 2,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3417716": "You can try combining different methods for the top 5 predictions based on template quality. For instance if for a sequence, the template is poor, you can fallback to a de-novo predictor like Protenix, but if it is of medium quality, you can use top 2 positions for Protenix and bottom 3 for the medium quality template and its transforms. Also, it isn't necessary to use only 1 template, you can use 2 or 3 different templates for diversifying your top 5, and based on quality, choose which ones you apply transformations on and where you place each in the top 5. ",
    "3416526": "Hi, even though the past competition was stable I feel that the idea of the metric choosing the best of 5 predictions  reward luck and stochastic postprocessing so much, is there any better approach for generating 5 predictions rather than  just applying random transformations to each prediction?"
  }
}