{
  "id": 627748,
  "title": "Clarification on the meaning and behavior of the score column in IA.tsv (Gene Ontology)",
  "url": "/competitions/cafa-6-protein-function-prediction/discussion/627748",
  "author_name": "",
  "post_date": "2025-11-17T03:41:26.970066400Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Dear colleagues,</p>\n<p>Could anyone clarify the meaning of a score value of zero in the IA.tsv file for Gene Ontology terms? Additionally, is the score identical for all proteins linked to a given GO term, or is it protein-specific?</p>\n<p>Any guidance or documentation would be appreciated.</p>",
  "messages": [
    {
      "id": "3332983",
      "postDate": "11/17/2025 03:41:26",
      "content": "<p>Dear colleagues,</p>\n<p>Could anyone clarify the meaning of a score value of zero in the IA.tsv file for Gene Ontology terms? Additionally, is the score identical for all proteins linked to a given GO term, or is it protein-specific?</p>\n<p>Any guidance or documentation would be appreciated.</p>",
      "rawMarkdown": "Dear colleagues,\n\nCould anyone clarify the meaning of a score value of zero in the IA.tsv file for Gene Ontology terms? Additionally, is the score identical for all proteins linked to a given GO term, or is it protein-specific?\n\nAny guidance or documentation would be appreciated.",
      "votes": null
    },
    {
      "id": "3334489",
      "postDate": "11/17/2025 14:54:32",
      "content": "<p>The scores, or information accretion (IA) values, are not protein-specific but linked to the given GO term. </p>\n<p>An IA value of zero can occur when a child term is annotated to the exact same proteins as the parent term. Based on this <a href=\"https://pmc.ncbi.nlm.nih.gov/articles/PMC3694662/\" target=\"_blank\">paper</a>, </p>\n<p><strong>ia(v) = -log Pr(v|P(v))</strong>, </p>\n<p>where <strong>P(v)</strong> is a set of parent nodes of <strong>v</strong>. When you have <strong>Pr(v|P(v)) = 1</strong>, i.e., all proteins annotated to <strong>P(v)</strong> are also annotated to term <strong>v</strong>, the score/information accretion of term <strong>v</strong> is 0.</p>\n<p>Most of the times, only a subset of proteins annotated to <strong>P(v)</strong> are annotated to <strong>v</strong> (since <strong>v</strong> is a child term), so most information accretion scores are not zero.</p>",
      "rawMarkdown": "The scores, or information accretion (IA) values, are not protein-specific but linked to the given GO term. \n\nAn IA value of zero can occur when a child term is annotated to the exact same proteins as the parent term. Based on this [paper](https://pmc.ncbi.nlm.nih.gov/articles/PMC3694662/), \n\n**ia(v) = -log Pr(v|P(v))**, \n\nwhere **P(v)** is a set of parent nodes of **v**. When you have **Pr(v|P(v)) = 1**, i.e., all proteins annotated to **P(v)** are also annotated to term **v**, the score/information accretion of term **v** is 0.\n\nMost of the times, only a subset of proteins annotated to **P(v)** are annotated to **v** (since **v** is a child term), so most information accretion scores are not zero.",
      "votes": null
    },
    {
      "id": "3343721",
      "postDate": "11/22/2025 00:24:37",
      "content": "<p>I assume we better delete the Go terms with an IA value of zero before labeling the training data. Because it only increases overfitting and make the model worse in predicting rare Go terms. I am training my final models now, if they fail to get weighted F1 of 0.3 or more than I should use other data than the sequences, I guess for the subcellular localization only the 100 AA N-terminal and 100 AA C-terminal are sufficient since my analysis of post translational tagging for protein transport in cells shows these tags are located overthere. I am thinking about doing an analysis for protein domains too but it seems much harder to cut the sequences for training dependent from protein domain analysis for predicting function and process. I am going on holliday so I hope the more gifted can accomplish this task and get scores close to 0.99 if the information is present on the globe to do this, otherwise superintelligent AI will also not. Lower scores are fine for giving direction to lab experiments. I am wondering about the COVID vaccin design and simulations, if it is so hard to predict with limited information, how accurate are these simulations, and how far advanced are big tech companies like google calico  in simulating the human body on cellular/protein resolution starting from DNA sequences.</p>",
      "rawMarkdown": "I assume we better delete the Go terms with an IA value of zero before labeling the training data. Because it only increases overfitting and make the model worse in predicting rare Go terms. I am training my final models now, if they fail to get weighted F1 of 0.3 or more than I should use other data than the sequences, I guess for the subcellular localization only the 100 AA N-terminal and 100 AA C-terminal are sufficient since my analysis of post translational tagging for protein transport in cells shows these tags are located overthere. I am thinking about doing an analysis for protein domains too but it seems much harder to cut the sequences for training dependent from protein domain analysis for predicting function and process. I am going on holliday so I hope the more gifted can accomplish this task and get scores close to 0.99 if the information is present on the globe to do this, otherwise superintelligent AI will also not. Lower scores are fine for giving direction to lab experiments. I am wondering about the COVID vaccin design and simulations, if it is so hard to predict with limited information, how accurate are these simulations, and how far advanced are big tech companies like google calico  in simulating the human body on cellular/protein resolution starting from DNA sequences.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3334489,
      "author_name": "ahphan",
      "author_url": "",
      "post_date": "11/17/2025 14:54:32",
      "content": "<p>The scores, or information accretion (IA) values, are not protein-specific but linked to the given GO term. </p>\n<p>An IA value of zero can occur when a child term is annotated to the exact same proteins as the parent term. Based on this <a href=\"https://pmc.ncbi.nlm.nih.gov/articles/PMC3694662/\" target=\"_blank\">paper</a>, </p>\n<p><strong>ia(v) = -log Pr(v|P(v))</strong>, </p>\n<p>where <strong>P(v)</strong> is a set of parent nodes of <strong>v</strong>. When you have <strong>Pr(v|P(v)) = 1</strong>, i.e., all proteins annotated to <strong>P(v)</strong> are also annotated to term <strong>v</strong>, the score/information accretion of term <strong>v</strong> is 0.</p>\n<p>Most of the times, only a subset of proteins annotated to <strong>P(v)</strong> are annotated to <strong>v</strong> (since <strong>v</strong> is a child term), so most information accretion scores are not zero.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3343721,
          "author_name": "joostertryckx",
          "author_url": "",
          "post_date": "11/22/2025 00:24:37",
          "content": "<p>I assume we better delete the Go terms with an IA value of zero before labeling the training data. Because it only increases overfitting and make the model worse in predicting rare Go terms. I am training my final models now, if they fail to get weighted F1 of 0.3 or more than I should use other data than the sequences, I guess for the subcellular localization only the 100 AA N-terminal and 100 AA C-terminal are sufficient since my analysis of post translational tagging for protein transport in cells shows these tags are located overthere. I am thinking about doing an analysis for protein domains too but it seems much harder to cut the sequences for training dependent from protein domain analysis for predicting function and process. I am going on holliday so I hope the more gifted can accomplish this task and get scores close to 0.99 if the information is present on the globe to do this, otherwise superintelligent AI will also not. Lower scores are fine for giving direction to lab experiments. I am wondering about the COVID vaccin design and simulations, if it is so hard to predict with limited information, how accurate are these simulations, and how far advanced are big tech companies like google calico  in simulating the human body on cellular/protein resolution starting from DNA sequences.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3332983": "Dear colleagues,\n\nCould anyone clarify the meaning of a score value of zero in the IA.tsv file for Gene Ontology terms? Additionally, is the score identical for all proteins linked to a given GO term, or is it protein-specific?\n\nAny guidance or documentation would be appreciated.",
    "3334489": "The scores, or information accretion (IA) values, are not protein-specific but linked to the given GO term. \n\nAn IA value of zero can occur when a child term is annotated to the exact same proteins as the parent term. Based on this [paper](https://pmc.ncbi.nlm.nih.gov/articles/PMC3694662/), \n\n**ia(v) = -log Pr(v|P(v))**, \n\nwhere **P(v)** is a set of parent nodes of **v**. When you have **Pr(v|P(v)) = 1**, i.e., all proteins annotated to **P(v)** are also annotated to term **v**, the score/information accretion of term **v** is 0.\n\nMost of the times, only a subset of proteins annotated to **P(v)** are annotated to **v** (since **v** is a child term), so most information accretion scores are not zero.",
    "3343721": "I assume we better delete the Go terms with an IA value of zero before labeling the training data. Because it only increases overfitting and make the model worse in predicting rare Go terms. I am training my final models now, if they fail to get weighted F1 of 0.3 or more than I should use other data than the sequences, I guess for the subcellular localization only the 100 AA N-terminal and 100 AA C-terminal are sufficient since my analysis of post translational tagging for protein transport in cells shows these tags are located overthere. I am thinking about doing an analysis for protein domains too but it seems much harder to cut the sequences for training dependent from protein domain analysis for predicting function and process. I am going on holliday so I hope the more gifted can accomplish this task and get scores close to 0.99 if the information is present on the globe to do this, otherwise superintelligent AI will also not. Lower scores are fine for giving direction to lab experiments. I am wondering about the COVID vaccin design and simulations, if it is so hard to predict with limited information, how accurate are these simulations, and how far advanced are big tech companies like google calico  in simulating the human body on cellular/protein resolution starting from DNA sequences."
  },
  "source": "meta"
}