{
  "id": 414985,
  "title": "Extracting Useful Features from Protein Descriptions for Training?",
  "url": "/competitions/cafa-5-protein-function-prediction/discussion/414985",
  "author_name": "",
  "post_date": "2023-06-04T12:16:47.451929200Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello, I'm wondering if it will be good to try and there are any strategies for extracting valuable features from FASTA descriptions and incorporating them into training. For example, the information about the species from which a protein originates? </p>\n<p>Are <a href=\"https://www.kaggle.com/datasets/sergeifironov/t5embeds\" target=\"_blank\">T5_embeds</a> created only from amino acid sequence or any other features included?</p>\n<p>Additionally, I'm curious if these features will be available in the final test data or should we strictly train on amino acid sequence only. Thank you for your insights. -Nara</p>",
  "messages": [
    {
      "id": "2287398",
      "postDate": "06/04/2023 12:16:47",
      "content": "<p>Hello, I'm wondering if it will be good to try and there are any strategies for extracting valuable features from FASTA descriptions and incorporating them into training. For example, the information about the species from which a protein originates? </p>\n<p>Are <a href=\"https://www.kaggle.com/datasets/sergeifironov/t5embeds\" target=\"_blank\">T5_embeds</a> created only from amino acid sequence or any other features included?</p>\n<p>Additionally, I'm curious if these features will be available in the final test data or should we strictly train on amino acid sequence only. Thank you for your insights. -Nara</p>",
      "rawMarkdown": "Hello, I'm wondering if it will be good to try and there are any strategies for extracting valuable features from FASTA descriptions and incorporating them into training. For example, the information about the species from which a protein originates? \n\nAre [T5_embeds](https://www.kaggle.com/datasets/sergeifironov/t5embeds) created only from amino acid sequence or any other features included?\n\nAdditionally, I'm curious if these features will be available in the final test data or should we strictly train on amino acid sequence only. Thank you for your insights. -Nara",
      "votes": null
    },
    {
      "id": "2288884",
      "postDate": "06/05/2023 18:28:40",
      "content": "<p>Thats a good question, I think there are many informations to consider but pretty hard to mix them with embeddings from pre-trained models. However I recommend to see this discussion <a href=\"https://www.kaggle.com/competitions/cafa-5-protein-function-prediction/discussion/413349\" target=\"_blank\">https://www.kaggle.com/competitions/cafa-5-protein-function-prediction/discussion/413349</a> that deals with the way of selecting a subset of GO terms. As you talk about \"species from which a protein originates\", it would be useful to select top 500 most frequent GO terms of each of the 3 species in order to get a final group of 1500 classes. </p>",
      "rawMarkdown": "Thats a good question, I think there are many informations to consider but pretty hard to mix them with embeddings from pre-trained models. However I recommend to see this discussion https://www.kaggle.com/competitions/cafa-5-protein-function-prediction/discussion/413349 that deals with the way of selecting a subset of GO terms. As you talk about \"species from which a protein originates\", it would be useful to select top 500 most frequent GO terms of each of the 3 species in order to get a final group of 1500 classes.",
      "votes": null
    },
    {
      "id": "2291799",
      "postDate": "06/07/2023 20:29:49",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/henriupton\" target=\"_blank\">@henriupton</a> that additional insight definitely helps!</p>",
      "rawMarkdown": "Thanks @henriupton that additional insight definitely helps!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2288884,
      "author_name": "henriupton",
      "author_url": "",
      "post_date": "06/05/2023 18:28:40",
      "content": "<p>Thats a good question, I think there are many informations to consider but pretty hard to mix them with embeddings from pre-trained models. However I recommend to see this discussion <a href=\"https://www.kaggle.com/competitions/cafa-5-protein-function-prediction/discussion/413349\" target=\"_blank\">https://www.kaggle.com/competitions/cafa-5-protein-function-prediction/discussion/413349</a> that deals with the way of selecting a subset of GO terms. As you talk about \"species from which a protein originates\", it would be useful to select top 500 most frequent GO terms of each of the 3 species in order to get a final group of 1500 classes. </p>",
      "votes": null,
      "replies": [
        {
          "id": 2291799,
          "author_name": "narareddy",
          "author_url": "",
          "post_date": "06/07/2023 20:29:49",
          "content": "<p>Thanks <a href=\"https://www.kaggle.com/henriupton\" target=\"_blank\">@henriupton</a> that additional insight definitely helps!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2287398": "Hello, I'm wondering if it will be good to try and there are any strategies for extracting valuable features from FASTA descriptions and incorporating them into training. For example, the information about the species from which a protein originates? \n\nAre [T5_embeds](https://www.kaggle.com/datasets/sergeifironov/t5embeds) created only from amino acid sequence or any other features included?\n\nAdditionally, I'm curious if these features will be available in the final test data or should we strictly train on amino acid sequence only. Thank you for your insights. -Nara",
    "2288884": "Thats a good question, I think there are many informations to consider but pretty hard to mix them with embeddings from pre-trained models. However I recommend to see this discussion https://www.kaggle.com/competitions/cafa-5-protein-function-prediction/discussion/413349 that deals with the way of selecting a subset of GO terms. As you talk about \"species from which a protein originates\", it would be useful to select top 500 most frequent GO terms of each of the 3 species in order to get a final group of 1500 classes.",
    "2291799": "Thanks @henriupton that additional insight definitely helps!"
  },
  "source": "meta"
}