{
  "id": 100161,
  "title": "How would a human score? (Human benchmark)",
  "url": "/competitions/recursion-cellular-image-classification/discussion/100161",
  "author_name": "",
  "post_date": "2019-07-16T23:10:38.200980200Z",
  "votes": 7,
  "comment_count": 1,
  "views": 0,
  "content": "<p>In other words, would a human reach 100% accuracy for this task? What is the human/traditional-way-of-performing-this-classification benchmark?   </p>\n\n<p>Thanks in advance!</p>\n\n<p>p.s. <a href=\"/chimael\">@chimael</a> <a href=\"/collinburton\">@collinburton</a> maybe you could try answering this question.   </p>",
  "messages": [
    {
      "id": "577665",
      "postDate": "07/16/2019 23:10:38",
      "content": "<p>In other words, would a human reach 100% accuracy for this task? What is the human/traditional-way-of-performing-this-classification benchmark?   </p>\n\n<p>Thanks in advance!</p>\n\n<p>p.s. <a href=\"/chimael\">@chimael</a> <a href=\"/collinburton\">@collinburton</a> maybe you could try answering this question.   </p>",
      "rawMarkdown": "In other words, would a human reach 100% accuracy for this task? What is the human/traditional-way-of-performing-this-classification benchmark?   \n\nThanks in advance!\n\np.s. @chimael @collinburton maybe you could try answering this question.",
      "votes": null
    },
    {
      "id": "577749",
      "postDate": "07/17/2019 03:24:16",
      "content": "<p>Very good question. </p>\n\n<h3>Can we guess visually a siRNA based on its secondary effects on cell morphology or proliferation?</h3>\n\n<ul>\n<li><p>Markers\nWhen performing siRNA experiments, we usually have a clue about which molecular pathway may be affected by the down-regulated protein. Either because it is a known protein with many published studies (i.e. <a href=\"https://www.ncbi.nlm.nih.gov/pubmed/?term=p53\">p53</a>) or it is an unknown protein with motifs found in other known proteins. Based on this knowledge, we would probe the involved molecular pathway with a specific marker (transcription of certain genes, protein translation, cell proliferation, cell death, cell differentiation, cell migration, ...). Here we only have access to generic markers of the cell (nucleus, nucleoli, Endoplasmic reticulum, Golgi, mitochondria, and cytoplasm), thus we cannot use the same approach.</p></li>\n<li><p>Structure and counts\nWe can still visualize the morphologies of the cells and count the cells. When I observe side by side the untreated control and other positive controls, I can see differences for some positive controls. The tricky part is to quantify these changes. How many cells are changed and by how much compared to untreated control? If you can do this on positive controls, then you would analyze the same features on experimental siRNAs. Since we have replicates of the same experiment, we can evaluate if the observed difference compared to untreated control is significant or not. </p></li>\n<li><p>Visual features extraction is best with CNN\nAs you may guess, feature extraction is very subjective and error-prone. However, even if you choose a simple visual feature with a significant difference between negative control and one positive control, you got a proper dimension to score experimental siRNA on. The hope is to use CNN to automatically extract visual features. </p></li>\n<li><p>My answer to your question: \nI don't think that a human can perform this task in a reasonable amount of time (hours). </p></li>\n</ul>",
      "rawMarkdown": "Very good question. \n\n### Can we guess visually a siRNA based on its secondary effects on cell morphology or proliferation?\n  \n* Markers\nWhen performing siRNA experiments, we usually have a clue about which molecular pathway may be affected by the down-regulated protein. Either because it is a known protein with many published studies (i.e. [p53](https://www.ncbi.nlm.nih.gov/pubmed/?term=p53)) or it is an unknown protein with motifs found in other known proteins. Based on this knowledge, we would probe the involved molecular pathway with a specific marker (transcription of certain genes, protein translation, cell proliferation, cell death, cell differentiation, cell migration, ...). Here we only have access to generic markers of the cell (nucleus, nucleoli, Endoplasmic reticulum, Golgi, mitochondria, and cytoplasm), thus we cannot use the same approach.\n  \n* Structure and counts\nWe can still visualize the morphologies of the cells and count the cells. When I observe side by side the untreated control and other positive controls, I can see differences for some positive controls. The tricky part is to quantify these changes. How many cells are changed and by how much compared to untreated control? If you can do this on positive controls, then you would analyze the same features on experimental siRNAs. Since we have replicates of the same experiment, we can evaluate if the observed difference compared to untreated control is significant or not. \n  \n* Visual features extraction is best with CNN\nAs you may guess, feature extraction is very subjective and error-prone. However, even if you choose a simple visual feature with a significant difference between negative control and one positive control, you got a proper dimension to score experimental siRNA on. The hope is to use CNN to automatically extract visual features. \n  \n* My answer to your question: \nI don't think that a human can perform this task in a reasonable amount of time (hours).",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 577749,
      "author_name": "chimael",
      "author_url": "",
      "post_date": "07/17/2019 03:24:16",
      "content": "<p>Very good question. </p>\n\n<h3>Can we guess visually a siRNA based on its secondary effects on cell morphology or proliferation?</h3>\n\n<ul>\n<li><p>Markers\nWhen performing siRNA experiments, we usually have a clue about which molecular pathway may be affected by the down-regulated protein. Either because it is a known protein with many published studies (i.e. <a href=\"https://www.ncbi.nlm.nih.gov/pubmed/?term=p53\">p53</a>) or it is an unknown protein with motifs found in other known proteins. Based on this knowledge, we would probe the involved molecular pathway with a specific marker (transcription of certain genes, protein translation, cell proliferation, cell death, cell differentiation, cell migration, ...). Here we only have access to generic markers of the cell (nucleus, nucleoli, Endoplasmic reticulum, Golgi, mitochondria, and cytoplasm), thus we cannot use the same approach.</p></li>\n<li><p>Structure and counts\nWe can still visualize the morphologies of the cells and count the cells. When I observe side by side the untreated control and other positive controls, I can see differences for some positive controls. The tricky part is to quantify these changes. How many cells are changed and by how much compared to untreated control? If you can do this on positive controls, then you would analyze the same features on experimental siRNAs. Since we have replicates of the same experiment, we can evaluate if the observed difference compared to untreated control is significant or not. </p></li>\n<li><p>Visual features extraction is best with CNN\nAs you may guess, feature extraction is very subjective and error-prone. However, even if you choose a simple visual feature with a significant difference between negative control and one positive control, you got a proper dimension to score experimental siRNA on. The hope is to use CNN to automatically extract visual features. </p></li>\n<li><p>My answer to your question: \nI don't think that a human can perform this task in a reasonable amount of time (hours). </p></li>\n</ul>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "577665": "In other words, would a human reach 100% accuracy for this task? What is the human/traditional-way-of-performing-this-classification benchmark?   \n\nThanks in advance!\n\np.s. @chimael @collinburton maybe you could try answering this question.",
    "577749": "Very good question. \n\n### Can we guess visually a siRNA based on its secondary effects on cell morphology or proliferation?\n  \n* Markers\nWhen performing siRNA experiments, we usually have a clue about which molecular pathway may be affected by the down-regulated protein. Either because it is a known protein with many published studies (i.e. [p53](https://www.ncbi.nlm.nih.gov/pubmed/?term=p53)) or it is an unknown protein with motifs found in other known proteins. Based on this knowledge, we would probe the involved molecular pathway with a specific marker (transcription of certain genes, protein translation, cell proliferation, cell death, cell differentiation, cell migration, ...). Here we only have access to generic markers of the cell (nucleus, nucleoli, Endoplasmic reticulum, Golgi, mitochondria, and cytoplasm), thus we cannot use the same approach.\n  \n* Structure and counts\nWe can still visualize the morphologies of the cells and count the cells. When I observe side by side the untreated control and other positive controls, I can see differences for some positive controls. The tricky part is to quantify these changes. How many cells are changed and by how much compared to untreated control? If you can do this on positive controls, then you would analyze the same features on experimental siRNAs. Since we have replicates of the same experiment, we can evaluate if the observed difference compared to untreated control is significant or not. \n  \n* Visual features extraction is best with CNN\nAs you may guess, feature extraction is very subjective and error-prone. However, even if you choose a simple visual feature with a significant difference between negative control and one positive control, you got a proper dimension to score experimental siRNA on. The hope is to use CNN to automatically extract visual features. \n  \n* My answer to your question: \nI don't think that a human can perform this task in a reasonable amount of time (hours)."
  },
  "source": "meta"
}