{
  "id": 612655,
  "title": "Can I use CAFA 5 1st place solution as my foundation?",
  "url": "/competitions/cafa-6-protein-function-prediction/discussion/612655",
  "author_name": "",
  "post_date": "2025-10-21T09:00:54.899852400Z",
  "votes": 5,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I was thinking of using the CAFA 5 1st place solution as a foundation or reference point — mainly to understand its architecture, preprocessing pipeline, and how they integrated sequence embeddings or GO terms.</p>\n<p>My idea is to start from that baseline and then modify key components (for example, try different embedding models like ESM2 or ProtT5, or fine-tune the prediction layers differently).</p>\n<p>Would this be a reasonable approach for someone new to the challenge, or would it be better to start from scratch and design the pipeline independently given that the GO graph, taxonomies, and data distributions may have evolved since CAFA 5?</p>",
  "messages": [
    {
      "id": "3304749",
      "postDate": "10/21/2025 09:00:54",
      "content": "<p>I was thinking of using the CAFA 5 1st place solution as a foundation or reference point — mainly to understand its architecture, preprocessing pipeline, and how they integrated sequence embeddings or GO terms.</p>\n<p>My idea is to start from that baseline and then modify key components (for example, try different embedding models like ESM2 or ProtT5, or fine-tune the prediction layers differently).</p>\n<p>Would this be a reasonable approach for someone new to the challenge, or would it be better to start from scratch and design the pipeline independently given that the GO graph, taxonomies, and data distributions may have evolved since CAFA 5?</p>",
      "rawMarkdown": "I was thinking of using the CAFA 5 1st place solution as a foundation or reference point — mainly to understand its architecture, preprocessing pipeline, and how they integrated sequence embeddings or GO terms.\n\nMy idea is to start from that baseline and then modify key components (for example, try different embedding models like ESM2 or ProtT5, or fine-tune the prediction layers differently).\n\nWould this be a reasonable approach for someone new to the challenge, or would it be better to start from scratch and design the pipeline independently given that the GO graph, taxonomies, and data distributions may have evolved since CAFA 5?",
      "votes": null
    },
    {
      "id": "3304788",
      "postDate": "10/21/2025 11:09:13",
      "content": "<p>Did the team that took 1st place share (or publish) the code for their solution?</p>",
      "rawMarkdown": "Did the team that took 1st place share (or publish) the code for their solution?",
      "votes": null
    },
    {
      "id": "3304970",
      "postDate": "10/21/2025 17:33:58",
      "content": "<p>ChatGPT said:</p>\n<p>Yeah, they gave a presentation and published an article on Kaggle explaining their approach. There’s also a video on Kaggle’s YouTube channel if you want more details.</p>",
      "rawMarkdown": "ChatGPT said:\n\nYeah, they gave a presentation and published an article on Kaggle explaining their approach. There’s also a video on Kaggle’s YouTube channel if you want more details.",
      "votes": null
    },
    {
      "id": "3305228",
      "postDate": "10/22/2025 10:07:33",
      "content": "<p>Assuming that you mean \"is that within the rules or spirit of Kaggle?\" Yes, that's how Kaggle works; code gets shared and reused, and anything freely available is fair game. Working that way probably levels the playing field to some extent and makes competitions more accessible. I'd argue that it usually raises that standard reached by models in competitions, \"standing on the shoulders of giants\" as it were.</p>\n<p>Of course there are people who wish that Kaggle were more restrictive, and offred more rewards for original coding and less for reusing others' work. That's an interesting and sometimes controversial debate, and not one I'm wading any deeper into right now.</p>",
      "rawMarkdown": "Assuming that you mean \"is that within the rules or spirit of Kaggle?\" Yes, that's how Kaggle works; code gets shared and reused, and anything freely available is fair game. Working that way probably levels the playing field to some extent and makes competitions more accessible. I'd argue that it usually raises that standard reached by models in competitions, \"standing on the shoulders of giants\" as it were.\n\nOf course there are people who wish that Kaggle were more restrictive, and offred more rewards for original coding and less for reusing others' work. That's an interesting and sometimes controversial debate, and not one I'm wading any deeper into right now.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3304788,
      "author_name": "sapr3s",
      "author_url": "",
      "post_date": "10/21/2025 11:09:13",
      "content": "<p>Did the team that took 1st place share (or publish) the code for their solution?</p>",
      "votes": null,
      "replies": [
        {
          "id": 3304970,
          "author_name": "rishabhjaindev",
          "author_url": "",
          "post_date": "10/21/2025 17:33:58",
          "content": "<p>ChatGPT said:</p>\n<p>Yeah, they gave a presentation and published an article on Kaggle explaining their approach. There’s also a video on Kaggle’s YouTube channel if you want more details.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3305228,
      "author_name": "jbomitchell",
      "author_url": "",
      "post_date": "10/22/2025 10:07:33",
      "content": "<p>Assuming that you mean \"is that within the rules or spirit of Kaggle?\" Yes, that's how Kaggle works; code gets shared and reused, and anything freely available is fair game. Working that way probably levels the playing field to some extent and makes competitions more accessible. I'd argue that it usually raises that standard reached by models in competitions, \"standing on the shoulders of giants\" as it were.</p>\n<p>Of course there are people who wish that Kaggle were more restrictive, and offred more rewards for original coding and less for reusing others' work. That's an interesting and sometimes controversial debate, and not one I'm wading any deeper into right now.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3304749": "I was thinking of using the CAFA 5 1st place solution as a foundation or reference point — mainly to understand its architecture, preprocessing pipeline, and how they integrated sequence embeddings or GO terms.\n\nMy idea is to start from that baseline and then modify key components (for example, try different embedding models like ESM2 or ProtT5, or fine-tune the prediction layers differently).\n\nWould this be a reasonable approach for someone new to the challenge, or would it be better to start from scratch and design the pipeline independently given that the GO graph, taxonomies, and data distributions may have evolved since CAFA 5?",
    "3304788": "Did the team that took 1st place share (or publish) the code for their solution?",
    "3304970": "ChatGPT said:\n\nYeah, they gave a presentation and published an article on Kaggle explaining their approach. There’s also a video on Kaggle’s YouTube channel if you want more details.",
    "3305228": "Assuming that you mean \"is that within the rules or spirit of Kaggle?\" Yes, that's how Kaggle works; code gets shared and reused, and anything freely available is fair game. Working that way probably levels the playing field to some extent and makes competitions more accessible. I'd argue that it usually raises that standard reached by models in competitions, \"standing on the shoulders of giants\" as it were.\n\nOf course there are people who wish that Kaggle were more restrictive, and offred more rewards for original coding and less for reusing others' work. That's an interesting and sometimes controversial debate, and not one I'm wading any deeper into right now."
  },
  "source": "meta"
}