Several of the structures for both the training and test data are corrupt. Have the organizers published any suggestions on what to do when links don't exist between nuclei?
", "messages": [ { "id": "553037", "postDate": "06/15/2019 03:05:54", "content": "Several of the structures for both the training and test data are corrupt. Have the organizers published any suggestions on what to do when links don't exist between nuclei?
", "rawMarkdown": "Several of the structures for both the training and test data are corrupt. Have the organizers published any suggestions on what to do when links don't exist between nuclei?", "votes": null }, { "id": "553101", "postDate": "06/15/2019 05:44:23", "content": "Phil - might help if you identified the structure files that you have found to be corrupt.
", "rawMarkdown": "Phil - might help if you identified the structure files that you have found to be corrupt.", "votes": null }, { "id": "553111", "postDate": "06/15/2019 05:52:17", "content": "All of these structures are from theoretical databases such as gdb-11 or gdb-13 or versions thereof (http://quantum-machine.org/datasets/). They were later optimized by DFT. So it is quite normal that for some strained or rather unusual examples (see Bredt's rule) the bond length are extremes or compounds are not \"allowed\" or non-existing. Because this is a rather large dataset it is OK to disregard those few structures and exclude them.
", "rawMarkdown": "All of these structures are from theoretical databases such as gdb-11 or gdb-13 or versions thereof (http://quantum-machine.org/datasets/). They were later optimized by DFT. So it is quite normal that for some strained or rather unusual examples (see Bredt's rule) the bond length are extremes or compounds are not \"allowed\" or non-existing. Because this is a rather large dataset it is OK to disregard those few structures and exclude them.", "votes": null }, { "id": "553841", "postDate": "06/16/2019 12:49:42", "content": "One example is dsgdb9nsd_005804, C5H11NO2, arranged as NH3CH2CCH3CH3COO. It’s representative of a bunch of molecules.\nOn one end of the molecule you have a nitrogen bonded to three hydrogen (N_3, H_4, H_5, H_6) with H-N-H bond angles of 108° to 109°. These bond angles are larger than the 107d of real NH3, and larger still than the 105.5° of your theoretical NH3 (dsgdb9nsd_000002). If the NH3 were bonded out, I'd think the bond angles would be steeper because the tip of the pyramid would be more electronegative, but it's been over 40 years since I took any chem, so I'm probably wrong about that relationship. Regardless, since the nitrogen has used up all its valences, it cannot bond to the nearby carbon (C_1), C-N length 158pm, which is long, but not outside the range of your C-N bonds. \nOn the other end of the molecule are two terminating oxygens bonded to a carbon (C_8, O_9, O_10) with a O-C-O bond angle of 133°. This bond angle, while rather huge, doesn't suggest that the oxygens are double bonded to the carbon. Although, the C-O bond lengths are 124pm, which is closer to your mean C=O (120pm) than your mean C-O (140pm). Regardless, in absence of a hydrogen near either of the oxygen (to form a carboxylic acid) the only option is to bond them to each other, with an O-O bond length of 227pm, which again seems way too long, considering peroxides have O-O bond lengths of 145pm. I think it would be a very weak bond. Bonding the oxygens leaves two carbons with valence 1 (C_1 & C_8), which would require a C-C bond length of 264pm, which again seems way too long. \nI could simply ignore these troublesome structures in the training data, but they're in the test data as well (dsgdb9nsd_001670). Perhaps, bonding the oxygens together and forming a ring (C_1, C_7, C_8) with a long C-C bond would be theoretically possible, but it doesn't look pretty. \nThe show stopper is the NH3. How should we link the three hydrogens (H_4, H_5, H_6) to the rest of the molecule? Should we promote the nitrogen to a carbon? The bond angles and the bond length to C_1 would suggest N_3 could be carbon. But, promotion WILL require adding a hydrogen somewhere. The obvious place would be on one of the oxygens. \nI did a quick search of the database for O-C=O structures and found bond angles from 123.5° to 126°, larger than the 120° of text book carboxylic acid, but nothing close to 133° found in this example. And the bond lengths in these other structures are not identical, typically 135pm for the C-O bond and 120pm for the C=O bond.\nThat's the dilemma.
", "rawMarkdown": "One example is dsgdb9nsd\\_005804, C5H11NO2, arranged as NH3CH2CCH3CH3COO. It’s representative of a bunch of molecules.\nOn one end of the molecule you have a nitrogen bonded to three hydrogen (N\\_3, H\\_4, H\\_5, H\\_6) with H-N-H bond angles of 108° to 109°. These bond angles are larger than the 107d of real NH3, and larger still than the 105.5° of your theoretical NH3 (dsgdb9nsd\\_000002). If the NH3 were bonded out, I'd think the bond angles would be steeper because the tip of the pyramid would be more electronegative, but it's been over 40 years since I took any chem, so I'm probably wrong about that relationship. Regardless, since the nitrogen has used up all its valences, it cannot bond to the nearby carbon (C\\_1), C-N length 158pm, which is long, but not outside the range of your C-N bonds. \nOn the other end of the molecule are two terminating oxygens bonded to a carbon (C\\_8, O\\_9, O\\_10) with a O-C-O bond angle of 133°. This bond angle, while rather huge, doesn't suggest that the oxygens are double bonded to the carbon. Although, the C-O bond lengths are 124pm, which is closer to your mean C=O (120pm) than your mean C-O (140pm). Regardless, in absence of a hydrogen near either of the oxygen (to form a carboxylic acid) the only option is to bond them to each other, with an O-O bond length of 227pm, which again seems way too long, considering peroxides have O-O bond lengths of 145pm. I think it would be a very weak bond. Bonding the oxygens leaves two carbons with valence 1 (C\\_1 & C\\_8), which would require a C-C bond length of 264pm, which again seems way too long. \nI could simply ignore these troublesome structures in the training data, but they're in the test data as well (dsgdb9nsd_001670). Perhaps, bonding the oxygens together and forming a ring (C\\_1, C\\_7, C\\_8) with a long C-C bond would be theoretically possible, but it doesn't look pretty. \nThe show stopper is the NH3. How should we link the three hydrogens (H\\_4, H\\_5, H\\_6) to the rest of the molecule? Should we promote the nitrogen to a carbon? The bond angles and the bond length to C\\_1 would suggest N\\_3 could be carbon. But, promotion *WILL* require adding a hydrogen somewhere. The obvious place would be on one of the oxygens. \nI did a quick search of the database for O-C=O structures and found bond angles from 123.5° to 126°, larger than the 120° of text book carboxylic acid, but nothing close to 133° found in this example. And the bond lengths in these other structures are not identical, typically 135pm for the C-O bond and 120pm for the C=O bond.\nThat's the dilemma.", "votes": null }, { "id": "554086", "postDate": "06/17/2019 00:13:27", "content": "Look at an analogous molecule Valine in Wiki . I don't know if the image shows. It is a zwitterion where the ends are oppositely charged: -NH3(+) and -COO(-)\nValine
", "rawMarkdown": "Look at an analogous molecule [Valine](https://en.wikipedia.org/wiki/Valine) in Wiki . I don't know if the image shows. It is a zwitterion where the ends are oppositely charged: -NH3(+) and -COO(-)\n[Valine](https://upload.wikimedia.org/wikipedia/commons/thumb/9/97/Valine_at_7.4_pH.png/165px-Valine_at_7.4_pH.png)", "votes": null }, { "id": "554133", "postDate": "06/17/2019 03:23:56", "content": "@philwebb59 \nthere is nothing wrong with the structure, its a zwitterion and \"people\" have used NMR and performed NMR shift predictions with those molecules using quantum chemistry for a long time. See the figure\nin https://pubs.acs.org/doi/abs/10.1021/ja0552343 you have NH3[+] and COO[-] similar to dsgdb9nsd_00580. Nothing wrong, just a special exception. Can be excluded or treated with special rules if exclusion is not wanted.
", "rawMarkdown": "philwebb59 \nthere is nothing wrong with the structure, its a zwitterion and \"people\" have used NMR and performed NMR shift predictions with those molecules using quantum chemistry for a long time. See the figure\nin https://pubs.acs.org/doi/abs/10.1021/ja0552343 you have NH3[+] and COO[-] similar to dsgdb9nsd_00580. Nothing wrong, just a special exception. Can be excluded or treated with special rules if exclusion is not wanted.", "votes": null }, { "id": "554140", "postDate": "06/17/2019 04:07:10", "content": "Oops! My bad. I thought about allowing an exception for the carboxylate anion, but I didn't consider it sharing its extra electron with an amonia cation through a chain of three carbons. That solves about 400 bad bond errors. Thanks!
", "rawMarkdown": "Oops! My bad. I thought about allowing an exception for the carboxylate anion, but I didn't consider it sharing its extra electron with an amonia cation through a chain of three carbons. That solves about 400 bad bond errors. Thanks!", "votes": null } ], "comments": [ { "id": 553101, "author_name": "pcjimmmy", "author_url": "", "post_date": "06/15/2019 05:44:23", "content": "Phil - might help if you identified the structure files that you have found to be corrupt.
", "votes": null, "replies": [] }, { "id": 553111, "author_name": "tobikaggle", "author_url": "", "post_date": "06/15/2019 05:52:17", "content": "All of these structures are from theoretical databases such as gdb-11 or gdb-13 or versions thereof (http://quantum-machine.org/datasets/). They were later optimized by DFT. So it is quite normal that for some strained or rather unusual examples (see Bredt's rule) the bond length are extremes or compounds are not \"allowed\" or non-existing. Because this is a rather large dataset it is OK to disregard those few structures and exclude them.
", "votes": null, "replies": [ { "id": 553841, "author_name": "philwebb59", "author_url": "", "post_date": "06/16/2019 12:49:42", "content": "One example is dsgdb9nsd_005804, C5H11NO2, arranged as NH3CH2CCH3CH3COO. It’s representative of a bunch of molecules.\nOn one end of the molecule you have a nitrogen bonded to three hydrogen (N_3, H_4, H_5, H_6) with H-N-H bond angles of 108° to 109°. These bond angles are larger than the 107d of real NH3, and larger still than the 105.5° of your theoretical NH3 (dsgdb9nsd_000002). If the NH3 were bonded out, I'd think the bond angles would be steeper because the tip of the pyramid would be more electronegative, but it's been over 40 years since I took any chem, so I'm probably wrong about that relationship. Regardless, since the nitrogen has used up all its valences, it cannot bond to the nearby carbon (C_1), C-N length 158pm, which is long, but not outside the range of your C-N bonds. \nOn the other end of the molecule are two terminating oxygens bonded to a carbon (C_8, O_9, O_10) with a O-C-O bond angle of 133°. This bond angle, while rather huge, doesn't suggest that the oxygens are double bonded to the carbon. Although, the C-O bond lengths are 124pm, which is closer to your mean C=O (120pm) than your mean C-O (140pm). Regardless, in absence of a hydrogen near either of the oxygen (to form a carboxylic acid) the only option is to bond them to each other, with an O-O bond length of 227pm, which again seems way too long, considering peroxides have O-O bond lengths of 145pm. I think it would be a very weak bond. Bonding the oxygens leaves two carbons with valence 1 (C_1 & C_8), which would require a C-C bond length of 264pm, which again seems way too long. \nI could simply ignore these troublesome structures in the training data, but they're in the test data as well (dsgdb9nsd_001670). Perhaps, bonding the oxygens together and forming a ring (C_1, C_7, C_8) with a long C-C bond would be theoretically possible, but it doesn't look pretty. \nThe show stopper is the NH3. How should we link the three hydrogens (H_4, H_5, H_6) to the rest of the molecule? Should we promote the nitrogen to a carbon? The bond angles and the bond length to C_1 would suggest N_3 could be carbon. But, promotion WILL require adding a hydrogen somewhere. The obvious place would be on one of the oxygens. \nI did a quick search of the database for O-C=O structures and found bond angles from 123.5° to 126°, larger than the 120° of text book carboxylic acid, but nothing close to 133° found in this example. And the bond lengths in these other structures are not identical, typically 135pm for the C-O bond and 120pm for the C=O bond.\nThat's the dilemma.
", "votes": null, "replies": [] }, { "id": 554086, "author_name": "stocks", "author_url": "", "post_date": "06/17/2019 00:13:27", "content": "Look at an analogous molecule Valine in Wiki . I don't know if the image shows. It is a zwitterion where the ends are oppositely charged: -NH3(+) and -COO(-)\nValine
", "votes": null, "replies": [] }, { "id": 554133, "author_name": "tobikaggle", "author_url": "", "post_date": "06/17/2019 03:23:56", "content": "@philwebb59 \nthere is nothing wrong with the structure, its a zwitterion and \"people\" have used NMR and performed NMR shift predictions with those molecules using quantum chemistry for a long time. See the figure\nin https://pubs.acs.org/doi/abs/10.1021/ja0552343 you have NH3[+] and COO[-] similar to dsgdb9nsd_00580. Nothing wrong, just a special exception. Can be excluded or treated with special rules if exclusion is not wanted.
", "votes": null, "replies": [] }, { "id": 554140, "author_name": "philwebb59", "author_url": "", "post_date": "06/17/2019 04:07:10", "content": "Oops! My bad. I thought about allowing an exception for the carboxylate anion, but I didn't consider it sharing its extra electron with an amonia cation through a chain of three carbons. That solves about 400 bad bond errors. Thanks!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "553037": "Several of the structures for both the training and test data are corrupt. Have the organizers published any suggestions on what to do when links don't exist between nuclei?", "553101": "Phil - might help if you identified the structure files that you have found to be corrupt.", "553111": "All of these structures are from theoretical databases such as gdb-11 or gdb-13 or versions thereof (http://quantum-machine.org/datasets/). They were later optimized by DFT. So it is quite normal that for some strained or rather unusual examples (see Bredt's rule) the bond length are extremes or compounds are not \"allowed\" or non-existing. Because this is a rather large dataset it is OK to disregard those few structures and exclude them.", "553841": "One example is dsgdb9nsd\\_005804, C5H11NO2, arranged as NH3CH2CCH3CH3COO. It’s representative of a bunch of molecules.\nOn one end of the molecule you have a nitrogen bonded to three hydrogen (N\\_3, H\\_4, H\\_5, H\\_6) with H-N-H bond angles of 108° to 109°. These bond angles are larger than the 107d of real NH3, and larger still than the 105.5° of your theoretical NH3 (dsgdb9nsd\\_000002). If the NH3 were bonded out, I'd think the bond angles would be steeper because the tip of the pyramid would be more electronegative, but it's been over 40 years since I took any chem, so I'm probably wrong about that relationship. Regardless, since the nitrogen has used up all its valences, it cannot bond to the nearby carbon (C\\_1), C-N length 158pm, which is long, but not outside the range of your C-N bonds. \nOn the other end of the molecule are two terminating oxygens bonded to a carbon (C\\_8, O\\_9, O\\_10) with a O-C-O bond angle of 133°. This bond angle, while rather huge, doesn't suggest that the oxygens are double bonded to the carbon. Although, the C-O bond lengths are 124pm, which is closer to your mean C=O (120pm) than your mean C-O (140pm). Regardless, in absence of a hydrogen near either of the oxygen (to form a carboxylic acid) the only option is to bond them to each other, with an O-O bond length of 227pm, which again seems way too long, considering peroxides have O-O bond lengths of 145pm. I think it would be a very weak bond. Bonding the oxygens leaves two carbons with valence 1 (C\\_1 & C\\_8), which would require a C-C bond length of 264pm, which again seems way too long. \nI could simply ignore these troublesome structures in the training data, but they're in the test data as well (dsgdb9nsd_001670). Perhaps, bonding the oxygens together and forming a ring (C\\_1, C\\_7, C\\_8) with a long C-C bond would be theoretically possible, but it doesn't look pretty. \nThe show stopper is the NH3. How should we link the three hydrogens (H\\_4, H\\_5, H\\_6) to the rest of the molecule? Should we promote the nitrogen to a carbon? The bond angles and the bond length to C\\_1 would suggest N\\_3 could be carbon. But, promotion *WILL* require adding a hydrogen somewhere. The obvious place would be on one of the oxygens. \nI did a quick search of the database for O-C=O structures and found bond angles from 123.5° to 126°, larger than the 120° of text book carboxylic acid, but nothing close to 133° found in this example. And the bond lengths in these other structures are not identical, typically 135pm for the C-O bond and 120pm for the C=O bond.\nThat's the dilemma.", "554086": "Look at an analogous molecule [Valine](https://en.wikipedia.org/wiki/Valine) in Wiki . I don't know if the image shows. It is a zwitterion where the ends are oppositely charged: -NH3(+) and -COO(-)\n[Valine](https://upload.wikimedia.org/wikipedia/commons/thumb/9/97/Valine_at_7.4_pH.png/165px-Valine_at_7.4_pH.png)", "554133": "philwebb59 \nthere is nothing wrong with the structure, its a zwitterion and \"people\" have used NMR and performed NMR shift predictions with those molecules using quantum chemistry for a long time. See the figure\nin https://pubs.acs.org/doi/abs/10.1021/ja0552343 you have NH3[+] and COO[-] similar to dsgdb9nsd_00580. Nothing wrong, just a special exception. Can be excluded or treated with special rules if exclusion is not wanted.", "554140": "Oops! My bad. I thought about allowing an exception for the carboxylate anion, but I didn't consider it sharing its extra electron with an amonia cation through a chain of three carbons. That solves about 400 bad bond errors. Thanks!" }, "source": "meta" }![Valine](\"https://upload.wikimedia.org/wikipedia/commons/thumb/9/97/Valine_at_7.4_pH.png/165px-Valine_at_7.4_pH.png\"){kind=link}