Let me check just to be sure. Are the following 6 types of data included in the test data?
\nAccording to https://www.kaggle.com/code/chemdatafarmer/scaffold-exploration, data from 4~6 account for 55.4% of the test data.
", "messages": [ { "id": "2836721", "postDate": "05/26/2024 04:53:39", "content": "Let me check just to be sure. Are the following 6 types of data included in the test data?
\nAccording to https://www.kaggle.com/code/chemdatafarmer/scaffold-exploration, data from 4~6 account for 55.4% of the test data.
", "rawMarkdown": "Let me check just to be sure. Are the following 6 types of data included in the test data?\n1. three Shared BBs + triazine core\n2. three Nonshared BBs + triazine core\n3. Shared BBs + Nonshared BBs +triazine core\n4. three Shared BBs + Non-triazine core\n5. three Nonshared BBs + Non-triazine core\n6. three Shared BBs + Nonshared BBs + Non-triazine core\n\nAccording to https://www.kaggle.com/code/chemdatafarmer/scaffold-exploration, data from 4~6 account for 55.4% of the test data.", "votes": null }, { "id": "2837291", "postDate": "05/26/2024 12:13:19", "content": "@roberthatch has done a much more thorough breakdown: Bb Breakdown
\n55.4% of the unique test molecules do not have a triazine core (I used @shlomoron folded dataset for this analysis). If you unfold it so there is one experimental observation per row, @roberthatch pointed out it's about 30% of the test rows that don't have a triazine core.
\nHope this clarifies things a bit.
", "rawMarkdown": "roberthatch has done a much more thorough breakdown: [Bb Breakdown](https://www.kaggle.com/competitions/leash-BELKA/discussion/496576)\n\n55.4% of the unique test molecules do not have a triazine core (I used @shlomoron folded dataset for this analysis). If you unfold it so there is one experimental observation per row, @roberthatch pointed out it's about 30% of the test rows that don't have a triazine core. \n\nHope this clarifies things a bit.", "votes": null }, { "id": "2837388", "postDate": "05/26/2024 13:03:36", "content": "3, 4, and 6 don't exist, if I understand your categories correctly.
\nEvery row in test data uses either ONLY shared BBs from train, or NO BBs from train. Likewise, every non-triazine core BB is exclusively used in the non-triazine core data, there's no BB overlap between that data and the \"non share BB\" test data.
", "rawMarkdown": "3, 4, and 6 don't exist, if I understand your categories correctly. \n\nEvery row in test data uses either ONLY shared BBs from train, or NO BBs from train. Likewise, every non-triazine core BB is exclusively used in the non-triazine core data, there's no BB overlap between that data and the \"non share BB\" test data.", "votes": null }, { "id": "2838034", "postDate": "05/26/2024 18:55:21", "content": "@roberthatch and @chemdatafarmer -
\nDo I understand you correctly that any non-triazine core structure in the test set also has BBs that are not in the training set? Those non-triazines are 100% composed of new BBs and a new core?
", "rawMarkdown": "roberthatch and @chemdatafarmer -\n\nDo I understand you correctly that any non-triazine core structure in the test set also has BBs that are not in the training set? Those non-triazines are 100% composed of new BBs and a new core?", "votes": null }, { "id": "2838079", "postDate": "05/26/2024 19:23:25", "content": "bb1 IS the new core, based on the host's description of the \"nefarious new library\", but yes, exactly.
\nIt's technically possible for the BB to be \"different\", but the effect on the final output molecule to be the same after chemical reactions. However, from my own EDA, it seemed like even after subtracting the chemical reactions, I didn't see any BB overlap. (If I didn't make any code or methodology mistakes.) For example, all the bb2s in the \"nefarious\" set have B(O)O in the building block smiles, but not in the final molecule output smiles. Presumably that's the \"boronate\" reaction mentioned in the host description.
", "rawMarkdown": "bb1 IS the new core, based on [the host's description](https://www.kaggle.com/competitions/leash-BELKA/discussion/503232) of the \"nefarious new library\", but yes, exactly.\n\nIt's technically possible for the BB to be \"different\", but the effect on the final output molecule to be the same after chemical reactions. However, from my own EDA, it seemed like even after subtracting the chemical reactions, I didn't see any BB overlap. (If I didn't make any code or methodology mistakes.) For example, all the bb2s in the \"nefarious\" set have B(O)O in the building block smiles, but not in the final molecule output smiles. Presumably that's the \"boronate\" reaction mentioned in the host description.", "votes": null }, { "id": "2838109", "postDate": "05/26/2024 20:07:53", "content": "Gotcha! Thank you for all the analysis and the great explanations. This is a tricky set, to be sure.
", "rawMarkdown": "Gotcha! Thank you for all the analysis and the great explanations. This is a tricky set, to be sure.", "votes": null }, { "id": "2841027", "postDate": "05/28/2024 11:14:54", "content": "There is no data that mixes non-shared BBs and shared BBs, right? I understand now. Thank you very much!
", "rawMarkdown": "There is no data that mixes non-shared BBs and shared BBs, right? I understand now. Thank you very much!", "votes": null }, { "id": "2841031", "postDate": "05/28/2024 11:16:26", "content": "Thank you very much for the clear explanation! I also appreciate you sharing the notebook.
", "rawMarkdown": "Thank you very much for the clear explanation! I also appreciate you sharing the notebook.", "votes": null } ], "comments": [ { "id": 2837291, "author_name": "chemdatafarmer", "author_url": "", "post_date": "05/26/2024 12:13:19", "content": "@roberthatch has done a much more thorough breakdown: Bb Breakdown
\n55.4% of the unique test molecules do not have a triazine core (I used @shlomoron folded dataset for this analysis). If you unfold it so there is one experimental observation per row, @roberthatch pointed out it's about 30% of the test rows that don't have a triazine core.
\nHope this clarifies things a bit.
", "votes": null, "replies": [ { "id": 2841031, "author_name": "taroiwai", "author_url": "", "post_date": "05/28/2024 11:16:26", "content": "Thank you very much for the clear explanation! I also appreciate you sharing the notebook.
", "votes": null, "replies": [] } ] }, { "id": 2837388, "author_name": "roberthatch", "author_url": "", "post_date": "05/26/2024 13:03:36", "content": "3, 4, and 6 don't exist, if I understand your categories correctly.
\nEvery row in test data uses either ONLY shared BBs from train, or NO BBs from train. Likewise, every non-triazine core BB is exclusively used in the non-triazine core data, there's no BB overlap between that data and the \"non share BB\" test data.
", "votes": null, "replies": [ { "id": 2841027, "author_name": "taroiwai", "author_url": "", "post_date": "05/28/2024 11:14:54", "content": "There is no data that mixes non-shared BBs and shared BBs, right? I understand now. Thank you very much!
", "votes": null, "replies": [] } ] }, { "id": 2838034, "author_name": "kirkdco", "author_url": "", "post_date": "05/26/2024 18:55:21", "content": "@roberthatch and @chemdatafarmer -
\nDo I understand you correctly that any non-triazine core structure in the test set also has BBs that are not in the training set? Those non-triazines are 100% composed of new BBs and a new core?
", "votes": null, "replies": [ { "id": 2838079, "author_name": "roberthatch", "author_url": "", "post_date": "05/26/2024 19:23:25", "content": "bb1 IS the new core, based on the host's description of the \"nefarious new library\", but yes, exactly.
\nIt's technically possible for the BB to be \"different\", but the effect on the final output molecule to be the same after chemical reactions. However, from my own EDA, it seemed like even after subtracting the chemical reactions, I didn't see any BB overlap. (If I didn't make any code or methodology mistakes.) For example, all the bb2s in the \"nefarious\" set have B(O)O in the building block smiles, but not in the final molecule output smiles. Presumably that's the \"boronate\" reaction mentioned in the host description.
", "votes": null, "replies": [ { "id": 2838109, "author_name": "kirkdco", "author_url": "", "post_date": "05/26/2024 20:07:53", "content": "Gotcha! Thank you for all the analysis and the great explanations. This is a tricky set, to be sure.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2836721": "Let me check just to be sure. Are the following 6 types of data included in the test data?\n1. three Shared BBs + triazine core\n2. three Nonshared BBs + triazine core\n3. Shared BBs + Nonshared BBs +triazine core\n4. three Shared BBs + Non-triazine core\n5. three Nonshared BBs + Non-triazine core\n6. three Shared BBs + Nonshared BBs + Non-triazine core\n\nAccording to https://www.kaggle.com/code/chemdatafarmer/scaffold-exploration, data from 4~6 account for 55.4% of the test data.", "2837291": "roberthatch has done a much more thorough breakdown: [Bb Breakdown](https://www.kaggle.com/competitions/leash-BELKA/discussion/496576)\n\n55.4% of the unique test molecules do not have a triazine core (I used @shlomoron folded dataset for this analysis). If you unfold it so there is one experimental observation per row, @roberthatch pointed out it's about 30% of the test rows that don't have a triazine core. \n\nHope this clarifies things a bit.", "2837388": "3, 4, and 6 don't exist, if I understand your categories correctly. \n\nEvery row in test data uses either ONLY shared BBs from train, or NO BBs from train. Likewise, every non-triazine core BB is exclusively used in the non-triazine core data, there's no BB overlap between that data and the \"non share BB\" test data.", "2838034": "roberthatch and @chemdatafarmer -\n\nDo I understand you correctly that any non-triazine core structure in the test set also has BBs that are not in the training set? Those non-triazines are 100% composed of new BBs and a new core?", "2838079": "bb1 IS the new core, based on [the host's description](https://www.kaggle.com/competitions/leash-BELKA/discussion/503232) of the \"nefarious new library\", but yes, exactly.\n\nIt's technically possible for the BB to be \"different\", but the effect on the final output molecule to be the same after chemical reactions. However, from my own EDA, it seemed like even after subtracting the chemical reactions, I didn't see any BB overlap. (If I didn't make any code or methodology mistakes.) For example, all the bb2s in the \"nefarious\" set have B(O)O in the building block smiles, but not in the final molecule output smiles. Presumably that's the \"boronate\" reaction mentioned in the host description.", "2838109": "Gotcha! Thank you for all the analysis and the great explanations. This is a tricky set, to be sure.", "2841027": "There is no data that mixes non-shared BBs and shared BBs, right? I understand now. Thank you very much!", "2841031": "Thank you very much for the clear explanation! I also appreciate you sharing the notebook." }, "source": "meta" }