Hi, I am not an expert on NMR in particular... Still although I understand you do not want to use QM calculations for this task, it brings me to the source of the data. ie if you create a model based on xyz coordinates, you have to precompute the coordinates. And I guess you do that using some quantum chemical method. Moreover, geometry optimization is typically significantly more costly than any single point calculation (though I can imagine using some lowerend method for optimization and higher end for nmr calculations). So the way I would see this is the following:\n1) you optimize coordinates using QM\n2) you calculate NMR properties using some machine learning method. \nSince step 1 can be cost defining, please tell me what is the benefit of using ML instead of QM for step 2. Or is there something I do not see?
", "messages": [ { "id": "568968", "postDate": "07/05/2019 19:16:33", "content": "Hi, I am not an expert on NMR in particular... Still although I understand you do not want to use QM calculations for this task, it brings me to the source of the data. ie if you create a model based on xyz coordinates, you have to precompute the coordinates. And I guess you do that using some quantum chemical method. Moreover, geometry optimization is typically significantly more costly than any single point calculation (though I can imagine using some lowerend method for optimization and higher end for nmr calculations). So the way I would see this is the following:\n1) you optimize coordinates using QM\n2) you calculate NMR properties using some machine learning method. \nSince step 1 can be cost defining, please tell me what is the benefit of using ML instead of QM for step 2. Or is there something I do not see?
", "rawMarkdown": "Hi, I am not an expert on NMR in particular... Still although I understand you do not want to use QM calculations for this task, it brings me to the source of the data. ie if you create a model based on xyz coordinates, you have to precompute the coordinates. And I guess you do that using some quantum chemical method. Moreover, geometry optimization is typically significantly more costly than any single point calculation (though I can imagine using some lowerend method for optimization and higher end for nmr calculations). So the way I would see this is the following:\n1) you optimize coordinates using QM\n2) you calculate NMR properties using some machine learning method. \nSince step 1 can be cost defining, please tell me what is the benefit of using ML instead of QM for step 2. Or is there something I do not see?", "votes": null }, { "id": "569000", "postDate": "07/05/2019 20:39:13", "content": "quantum chemist speaking here... \n1) you cannot extract coordinates from smiles. you may get some first guess, but this would be completely unsuitable for use because it does not represent the global minimum on the potential energy surface. \nYou could potentially get the geometry from some diffraction experiments though. Still provided you acquire coordinates using QM, this would very likely be more time consuming operation. \n2) yes I agree this would be worthwhile, no doubt about it. \n3) what would be most interesting of all though would be to try to compute this using SMILES as input as opposed to xyz coordinates.
", "rawMarkdown": "quantum chemist speaking here... \n1) you cannot extract coordinates from smiles. you may get some first guess, but this would be completely unsuitable for use because it does not represent the global minimum on the potential energy surface. \nYou could potentially get the geometry from some diffraction experiments though. Still provided you acquire coordinates using QM, this would very likely be more time consuming operation. \n2) yes I agree this would be worthwhile, no doubt about it. \n3) what would be most interesting of all though would be to try to compute this using SMILES as input as opposed to xyz coordinates.", "votes": null }, { "id": "579159", "postDate": "07/18/2019 14:45:46", "content": "@svobodnik86 I believe many people are already using the SMILES name as a feature. You just have to invent a smart way to translate the name into features that describe the molecule.
", "rawMarkdown": "svobodnik86 I believe many people are already using the SMILES name as a feature. You just have to invent a smart way to translate the name into features that describe the molecule.", "votes": null }, { "id": "579620", "postDate": "07/19/2019 03:27:58", "content": "You are saying that step 2 using QM is less computationally expensive than step 1?
", "rawMarkdown": "You are saying that step 2 using QM is less computationally expensive than step 1?", "votes": null }, { "id": "597526", "postDate": "08/12/2019 13:10:22", "content": "In general it should be. It is of course possible that you have everything already pre-computed as @returnofsputnik suggests. Simple explanation here. If you want to achieve step one, you need to perform an optimization on a very complex hypersuface with many local minima. Pick something else than proper global minimum and you fail (see for instance this https://chem.libretexts.org/Courses/University_of_California_Davis/UCD_Chem_107B%3A_Physical_Chemistry_for_Life_Scientists/Chapters/2%3A_Chemical_Kinetics/2.06%3A_Potential_Energy_Surfaces and just look at pics and variables to get quick intution). So you run the similation and generally calculate something between tens and thousands of energies. \nFor step 2 on the other hand, you typically need lesser number of energy calculations to get it.
", "rawMarkdown": "In general it should be. It is of course possible that you have everything already pre-computed as @returnofsputnik suggests. Simple explanation here. If you want to achieve step one, you need to perform an optimization on a very complex hypersuface with many local minima. Pick something else than proper global minimum and you fail (see for instance this https://chem.libretexts.org/Courses/University_of_California_Davis/UCD_Chem_107B%3A_Physical_Chemistry_for_Life_Scientists/Chapters/2%3A_Chemical_Kinetics/2.06%3A_Potential_Energy_Surfaces and just look at pics and variables to get quick intution). So you run the similation and generally calculate something between tens and thousands of energies. \nFor step 2 on the other hand, you typically need lesser number of energy calculations to get it.", "votes": null }, { "id": "597804", "postDate": "08/12/2019 18:48:03", "content": "Thanks. I do know a bit about optimization of non convex functions, but I had absolutely no clue about the relative complexity of step 1 and step 2.
", "rawMarkdown": "Thanks. I do know a bit about optimization of non convex functions, but I had absolutely no clue about the relative complexity of step 1 and step 2.", "votes": null } ], "comments": [ { "id": 569000, "author_name": "returnofsputnik", "author_url": "", "post_date": "07/05/2019 20:39:13", "content": "quantum chemist speaking here... \n1) you cannot extract coordinates from smiles. you may get some first guess, but this would be completely unsuitable for use because it does not represent the global minimum on the potential energy surface. \nYou could potentially get the geometry from some diffraction experiments though. Still provided you acquire coordinates using QM, this would very likely be more time consuming operation. \n2) yes I agree this would be worthwhile, no doubt about it. \n3) what would be most interesting of all though would be to try to compute this using SMILES as input as opposed to xyz coordinates.
", "votes": null, "replies": [ { "id": 579159, "author_name": "returnofsputnik", "author_url": "", "post_date": "07/18/2019 14:45:46", "content": "@svobodnik86 I believe many people are already using the SMILES name as a feature. You just have to invent a smart way to translate the name into features that describe the molecule.
", "votes": null, "replies": [] } ] }, { "id": 579620, "author_name": "cpmpml", "author_url": "", "post_date": "07/19/2019 03:27:58", "content": "You are saying that step 2 using QM is less computationally expensive than step 1?
", "votes": null, "replies": [ { "id": 597526, "author_name": "svobodnik86", "author_url": "", "post_date": "08/12/2019 13:10:22", "content": "In general it should be. It is of course possible that you have everything already pre-computed as @returnofsputnik suggests. Simple explanation here. If you want to achieve step one, you need to perform an optimization on a very complex hypersuface with many local minima. Pick something else than proper global minimum and you fail (see for instance this https://chem.libretexts.org/Courses/University_of_California_Davis/UCD_Chem_107B%3A_Physical_Chemistry_for_Life_Scientists/Chapters/2%3A_Chemical_Kinetics/2.06%3A_Potential_Energy_Surfaces and just look at pics and variables to get quick intution). So you run the similation and generally calculate something between tens and thousands of energies. \nFor step 2 on the other hand, you typically need lesser number of energy calculations to get it.
", "votes": null, "replies": [] }, { "id": 597804, "author_name": "cpmpml", "author_url": "", "post_date": "08/12/2019 18:48:03", "content": "Thanks. I do know a bit about optimization of non convex functions, but I had absolutely no clue about the relative complexity of step 1 and step 2.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "568968": "Hi, I am not an expert on NMR in particular... Still although I understand you do not want to use QM calculations for this task, it brings me to the source of the data. ie if you create a model based on xyz coordinates, you have to precompute the coordinates. And I guess you do that using some quantum chemical method. Moreover, geometry optimization is typically significantly more costly than any single point calculation (though I can imagine using some lowerend method for optimization and higher end for nmr calculations). So the way I would see this is the following:\n1) you optimize coordinates using QM\n2) you calculate NMR properties using some machine learning method. \nSince step 1 can be cost defining, please tell me what is the benefit of using ML instead of QM for step 2. Or is there something I do not see?", "569000": "1. For step 1, Maybe obtaining the coordinates/geometries does not require QM so it can be done very fast. Or, maybe they just have simply all been computed (everything in the SMILES/INCHI database has the xyz coordinates already computed).\n2. For step 2, ML is a lot faster than QM. I think for some rows, it can take days using QM to calculate the scalar coupling constant when the molecule is very large/complex. Therefore an ML solution would be extremely worthwhile", "578980": "quantum chemist speaking here... \n1) you cannot extract coordinates from smiles. you may get some first guess, but this would be completely unsuitable for use because it does not represent the global minimum on the potential energy surface. \nYou could potentially get the geometry from some diffraction experiments though. Still provided you acquire coordinates using QM, this would very likely be more time consuming operation. \n2) yes I agree this would be worthwhile, no doubt about it. \n3) what would be most interesting of all though would be to try to compute this using SMILES as input as opposed to xyz coordinates.", "579159": "svobodnik86 I believe many people are already using the SMILES name as a feature. You just have to invent a smart way to translate the name into features that describe the molecule.", "579620": "You are saying that step 2 using QM is less computationally expensive than step 1?", "597526": "In general it should be. It is of course possible that you have everything already pre-computed as @returnofsputnik suggests. Simple explanation here. If you want to achieve step one, you need to perform an optimization on a very complex hypersuface with many local minima. Pick something else than proper global minimum and you fail (see for instance this https://chem.libretexts.org/Courses/University_of_California_Davis/UCD_Chem_107B%3A_Physical_Chemistry_for_Life_Scientists/Chapters/2%3A_Chemical_Kinetics/2.06%3A_Potential_Energy_Surfaces and just look at pics and variables to get quick intution). So you run the similation and generally calculate something between tens and thousands of energies. \nFor step 2 on the other hand, you typically need lesser number of energy calculations to get it.", "597804": "Thanks. I do know a bit about optimization of non convex functions, but I had absolutely no clue about the relative complexity of step 1 and step 2." }, "source": "meta" }