{
  "id": 679904,
  "title": "Three things I learned optimising RNA structure prediction this week ",
  "url": "/competitions/stanford-rna-3d-folding-2/discussion/679904",
  "author_name": "Craig_Parker",
  "post_date": "2026-03-04T15:26:51.356000",
  "votes": 6,
  "comment_count": 1,
  "views": 0,
  "content": "<ol>\n<li>More geometry correction doesn't mean better score\nAdded one extra pass of bond-length/clash correction to my predictions. Score dropped from 0.295 → 0.282. Turns out Protenix already bakes in physical constraints , adding more on top just distorts\nthe conformations it worked hard to predict.</li>\n<li>The winner used no deep learning at all\nWhile I've been wrestling with Protenix configs, the competition leaderboard paper says team \"john\" won Phase 1 with pure template matching (TM-align 0.593). A well-aligned homolog beats a confused neural net every time. Humbling.</li>\n<li>Finetuned checkpoints need surgery to swap in\nTried plugging in an RNA3DB-finetuned checkpoint. Score collapsed to 0.220 (basically pure TBM fallback). The InferenceRunner locks checkpoint loading to {dir}/{model_name}.pt with no override path- so architecture/name mismatches are completely silent failures.\nQuestion for the community: Anyone successfully gotten the zoushuxian/protenix-finetuned-rna3db-all-1599 checkpoint working with InferenceRunner? Does it need different model configs than the base model, or is a symlink sufficient?</li>\n</ol>",
  "messages": [
    {
      "id": 3417087,
      "postDate": "2026-03-04T15:26:51.357Z",
      "content": "<ol>\n<li>More geometry correction doesn't mean better score\nAdded one extra pass of bond-length/clash correction to my predictions. Score dropped from 0.295 → 0.282. Turns out Protenix already bakes in physical constraints , adding more on top just distorts\nthe conformations it worked hard to predict.</li>\n<li>The winner used no deep learning at all\nWhile I've been wrestling with Protenix configs, the competition leaderboard paper says team \"john\" won Phase 1 with pure template matching (TM-align 0.593). A well-aligned homolog beats a confused neural net every time. Humbling.</li>\n<li>Finetuned checkpoints need surgery to swap in\nTried plugging in an RNA3DB-finetuned checkpoint. Score collapsed to 0.220 (basically pure TBM fallback). The InferenceRunner locks checkpoint loading to {dir}/{model_name}.pt with no override path- so architecture/name mismatches are completely silent failures.\nQuestion for the community: Anyone successfully gotten the zoushuxian/protenix-finetuned-rna3db-all-1599 checkpoint working with InferenceRunner? Does it need different model configs than the base model, or is a symlink sufficient?</li>\n</ol>",
      "rawMarkdown": " 1. More geometry correction doesn't mean better score\n  Added one extra pass of bond-length/clash correction to my predictions. Score dropped from 0.295 → 0.282. Turns out Protenix already bakes in physical constraints , adding more on top just distorts\n   the conformations it worked hard to predict.\n\n2. The winner used no deep learning at all\n  While I've been wrestling with Protenix configs, the competition leaderboard paper says team \"john\" won Phase 1 with pure template matching (TM-align 0.593). A well-aligned homolog beats a confused neural net every time. Humbling.\n\n3. Finetuned checkpoints need surgery to swap in\nTried plugging in an RNA3DB-finetuned checkpoint. Score collapsed to 0.220 (basically pure TBM fallback). The InferenceRunner locks checkpoint loading to {dir}/{model_name}.pt with no override path- so architecture/name mismatches are completely silent failures.\n\nQuestion for the community: Anyone successfully gotten the zoushuxian/protenix-finetuned-rna3db-all-1599 checkpoint working with InferenceRunner? Does it need different model configs than the base model, or is a symlink sufficient?",
      "votes": 6
    },
    {
      "id": 3417800,
      "postDate": "2026-03-06T09:02:44.513Z",
      "content": "<p>Yes but the samples which do not have a matching template can be solved via protenix. i think this does improve the score, see the score difference between:\n<a href=\"https://www.kaggle.com/code/nihilisticneuralnet/stanford-rna-folding-2-template-based-approach\" target=\"_blank\">https://www.kaggle.com/code/nihilisticneuralnet/stanford-rna-folding-2-template-based-approach</a>\n<a href=\"https://www.kaggle.com/code/nihilisticneuralnet/0-409-stanford-rna-folding-2-protenix-template\" target=\"_blank\">https://www.kaggle.com/code/nihilisticneuralnet/0-409-stanford-rna-folding-2-protenix-template</a></p>\n<p>However a very large portion of samples do have a template, so we should concentrate on TBM :D</p>\n<p>I am using TBM combined with another notebook to estimate the resulting validation score (since we know the coordinates of the validation sequences). Using this I will now try to improve the TBM approach since I also think that there is a lot to improve, including:</p>\n<p>-Base-pair foldings are just left as-is, instead of moving the neighbours as well to get a proper structure bending.</p>",
      "rawMarkdown": "Yes but the samples which do not have a matching template can be solved via protenix. i think this does improve the score, see the score difference between:\nhttps://www.kaggle.com/code/nihilisticneuralnet/stanford-rna-folding-2-template-based-approach\nhttps://www.kaggle.com/code/nihilisticneuralnet/0-409-stanford-rna-folding-2-protenix-template\n\nHowever a very large portion of samples do have a template, so we should concentrate on TBM :D\n\nI am using TBM combined with another notebook to estimate the resulting validation score (since we know the coordinates of the validation sequences). Using this I will now try to improve the TBM approach since I also think that there is a lot to improve, including:\n\n-Base-pair foldings are just left as-is, instead of moving the neighbours as well to get a proper structure bending.\n\n"
    }
  ],
  "comments": [
    {
      "id": 3417800,
      "author_name": "Andres H. Zapke",
      "author_url": "",
      "post_date": "2026-03-06T09:02:44.513000",
      "content": "<p>Yes but the samples which do not have a matching template can be solved via protenix. i think this does improve the score, see the score difference between:\n<a href=\"https://www.kaggle.com/code/nihilisticneuralnet/stanford-rna-folding-2-template-based-approach\" target=\"_blank\">https://www.kaggle.com/code/nihilisticneuralnet/stanford-rna-folding-2-template-based-approach</a>\n<a href=\"https://www.kaggle.com/code/nihilisticneuralnet/0-409-stanford-rna-folding-2-protenix-template\" target=\"_blank\">https://www.kaggle.com/code/nihilisticneuralnet/0-409-stanford-rna-folding-2-protenix-template</a></p>\n<p>However a very large portion of samples do have a template, so we should concentrate on TBM :D</p>\n<p>I am using TBM combined with another notebook to estimate the resulting validation score (since we know the coordinates of the validation sequences). Using this I will now try to improve the TBM approach since I also think that there is a lot to improve, including:</p>\n<p>-Base-pair foldings are just left as-is, instead of moving the neighbours as well to get a proper structure bending.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3417087": " 1. More geometry correction doesn't mean better score\n  Added one extra pass of bond-length/clash correction to my predictions. Score dropped from 0.295 → 0.282. Turns out Protenix already bakes in physical constraints , adding more on top just distorts\n   the conformations it worked hard to predict.\n\n2. The winner used no deep learning at all\n  While I've been wrestling with Protenix configs, the competition leaderboard paper says team \"john\" won Phase 1 with pure template matching (TM-align 0.593). A well-aligned homolog beats a confused neural net every time. Humbling.\n\n3. Finetuned checkpoints need surgery to swap in\nTried plugging in an RNA3DB-finetuned checkpoint. Score collapsed to 0.220 (basically pure TBM fallback). The InferenceRunner locks checkpoint loading to {dir}/{model_name}.pt with no override path- so architecture/name mismatches are completely silent failures.\n\nQuestion for the community: Anyone successfully gotten the zoushuxian/protenix-finetuned-rna3db-all-1599 checkpoint working with InferenceRunner? Does it need different model configs than the base model, or is a symlink sufficient?",
    "3417800": "Yes but the samples which do not have a matching template can be solved via protenix. i think this does improve the score, see the score difference between:\nhttps://www.kaggle.com/code/nihilisticneuralnet/stanford-rna-folding-2-template-based-approach\nhttps://www.kaggle.com/code/nihilisticneuralnet/0-409-stanford-rna-folding-2-protenix-template\n\nHowever a very large portion of samples do have a template, so we should concentrate on TBM :D\n\nI am using TBM combined with another notebook to estimate the resulting validation score (since we know the coordinates of the validation sequences). Using this I will now try to improve the TBM approach since I also think that there is a lot to improve, including:\n\n-Base-pair foldings are just left as-is, instead of moving the neighbours as well to get a proper structure bending.\n\n"
  }
}