{
  "id": 461174,
  "title": "Our Solution write-up",
  "url": "/competitions/open-problems-single-cell-perturbations/writeups/amirasiaeet-our-solution-write-up",
  "author_name": "",
  "post_date": "2023-12-12T23:58:20.183Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p><a href=\"https://github.com/kavaryan/sc-pertb/blob/master/report.pdf\" target=\"_blank\">https://github.com/kavaryan/sc-pertb/blob/master/report.pdf</a></p>",
  "messages": [
    {
      "id": "2559575",
      "postDate": "12/12/2023 23:56:30",
      "content": "<p><a href=\"https://github.com/kavaryan/sc-pertb/blob/master/report.pdf\" target=\"_blank\">https://github.com/kavaryan/sc-pertb/blob/master/report.pdf</a></p>",
      "rawMarkdown": "https://github.com/kavaryan/sc-pertb/blob/master/report.pdf",
      "votes": null
    },
    {
      "id": "2560089",
      "postDate": "12/13/2023 09:00:42",
      "content": "<p>It's very nice to see some causal learning in the challenge! I loved it! thanks for sharing, some comments/questions that come to my mind:</p>\n<ul>\n<li><p>I always find it hard to match the assumptions of causal methods with real biology (like modelling a complex dynamical system with DAGs). Did you inspect the resulting DAGs to see if you capture some known processes?</p></li>\n<li><p>I'm curious about the target genes as well. If I understood correctly, you assume that the target genes can be derived from gene expression. However, from the drug to changes in the transcriptional program, there can be many confounding mechanisms, so you don't have a direct causal mechanism between the drug and gene.</p></li>\n<li><p>I have experimented with some similar methods in the past (like NOTEARS) and always found them hard to train, especially in achieving convergence towards a valid DAG. However, in your case, since you allow for stable cycles, this might simplify the training of the model. How difficult is to train that model in this data?</p></li>\n</ul>\n<p>Congrats on your nice modelling approach</p>",
      "rawMarkdown": "It's very nice to see some causal learning in the challenge! I loved it! thanks for sharing, some comments/questions that come to my mind:\n\n- I always find it hard to match the assumptions of causal methods with real biology (like modelling a complex dynamical system with DAGs). Did you inspect the resulting DAGs to see if you capture some known processes?\n\n- I'm curious about the target genes as well. If I understood correctly, you assume that the target genes can be derived from gene expression. However, from the drug to changes in the transcriptional program, there can be many confounding mechanisms, so you don't have a direct causal mechanism between the drug and gene.\n\n- I have experimented with some similar methods in the past (like NOTEARS) and always found them hard to train, especially in achieving convergence towards a valid DAG. However, in your case, since you allow for stable cycles, this might simplify the training of the model. How difficult is to train that model in this data?\n\nCongrats on your nice modelling approach",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2560089,
      "author_name": "pablormier",
      "author_url": "",
      "post_date": "12/13/2023 09:00:42",
      "content": "<p>It's very nice to see some causal learning in the challenge! I loved it! thanks for sharing, some comments/questions that come to my mind:</p>\n<ul>\n<li><p>I always find it hard to match the assumptions of causal methods with real biology (like modelling a complex dynamical system with DAGs). Did you inspect the resulting DAGs to see if you capture some known processes?</p></li>\n<li><p>I'm curious about the target genes as well. If I understood correctly, you assume that the target genes can be derived from gene expression. However, from the drug to changes in the transcriptional program, there can be many confounding mechanisms, so you don't have a direct causal mechanism between the drug and gene.</p></li>\n<li><p>I have experimented with some similar methods in the past (like NOTEARS) and always found them hard to train, especially in achieving convergence towards a valid DAG. However, in your case, since you allow for stable cycles, this might simplify the training of the model. How difficult is to train that model in this data?</p></li>\n</ul>\n<p>Congrats on your nice modelling approach</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2559575": "https://github.com/kavaryan/sc-pertb/blob/master/report.pdf",
    "2560089": "It's very nice to see some causal learning in the challenge! I loved it! thanks for sharing, some comments/questions that come to my mind:\n\n- I always find it hard to match the assumptions of causal methods with real biology (like modelling a complex dynamical system with DAGs). Did you inspect the resulting DAGs to see if you capture some known processes?\n\n- I'm curious about the target genes as well. If I understood correctly, you assume that the target genes can be derived from gene expression. However, from the drug to changes in the transcriptional program, there can be many confounding mechanisms, so you don't have a direct causal mechanism between the drug and gene.\n\n- I have experimented with some similar methods in the past (like NOTEARS) and always found them hard to train, especially in achieving convergence towards a valid DAG. However, in your case, since you allow for stable cycles, this might simplify the training of the model. How difficult is to train that model in this data?\n\nCongrats on your nice modelling approach"
  },
  "source": "meta"
}