![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Ff7a709d18e07b875c12097828f988cda%2FSelection_148.png?generation=1716742133882024&alt=media\")
Hey Folks,
\nI had to take a break from this competition since before the scoring update. Looks like things have gotten pretty interesting!
\nI'll probably dabble a bit this weekend (it's a long weekend in the US). Are there any cheminformatics questions y'all are finding particularly troublesome? Please feel free to suggest some topics you wish there was a public notebook for and I will help out if I can.
", "messages": [ { "id": "2837318", "postDate": "05/26/2024 12:29:15", "content": "Hey Folks,
\nI had to take a break from this competition since before the scoring update. Looks like things have gotten pretty interesting!
\nI'll probably dabble a bit this weekend (it's a long weekend in the US). Are there any cheminformatics questions y'all are finding particularly troublesome? Please feel free to suggest some topics you wish there was a public notebook for and I will help out if I can.
", "rawMarkdown": "Hey Folks,\n\nI had to take a break from this competition since before the scoring update. Looks like things have gotten pretty interesting!\n\nI'll probably dabble a bit this weekend (it's a long weekend in the US). Are there any cheminformatics questions y'all are finding particularly troublesome? Please feel free to suggest some topics you wish there was a public notebook for and I will help out if I can.", "votes": null }, { "id": "2837428", "postDate": "05/26/2024 13:34:23", "content": "3d descriptors and 3d similarity questions. From some of the scientific papers rabbit trails I've followed, ROCS tanimoto combo score (shape and \"color\") is mentioned and appears to be state of the art for (speed and) 3d similarity scores between two molecules.
\nAre there any good open source alternatives that you're aware of?
\nMore general topic, but I'm not sure whether I even understand the ~10 3d descriptors that rdkit can calculate? The documentation seemed sparse on them. Many of them seem related to shape but without being sufficient to describe a shape approximation? For example, an ellipse could be defined by 3 diameter numbers. But I'm not so sure that knowing PMI1/2/3 is as descriptive of the overall shape?
", "rawMarkdown": "3d descriptors and 3d similarity questions. From some of the scientific papers rabbit trails I've followed, ROCS tanimoto combo score (shape and \"color\") is mentioned and appears to be state of the art for (speed and) 3d similarity scores between two molecules. \n\nAre there any good open source alternatives that you're aware of?\n\nMore general topic, but I'm not sure whether I even understand the ~10 3d descriptors that rdkit can calculate? The documentation seemed sparse on them. Many of them seem related to shape but without being sufficient to describe a shape approximation? For example, an ellipse could be defined by 3 diameter numbers. But I'm not so sure that knowing PMI1/2/3 is as descriptive of the overall shape?", "votes": null }, { "id": "2837504", "postDate": "05/26/2024 14:41:27", "content": "Hey @roberthatch
\nUnfortunately, I'm not aware of any open source software that can compete with the speed of ROCS or its GPU accelerated cousin, fastROCS. You're right, these tools are state of the art and commonly applied in industry (a usual workflow is to use OMEGA to get conformers, then use ROCS to score the overlay to some known or hypothesized binding mode of a reference molecule).
\nThat said, this paper may be useful to you: https://doi.org/10.1186/s13321-023-00703-1. Within, they describe using RDKit's rdShapeHelpers and 3D fingerprints to get a \"combo score\" (paper: \"the average of shape similarity and 3D fingerprint similarity\"), which to me looks conceptually similar to the \"shape and color\" score you'd get from ROCS.
\nOn the more general topic, the way I've seen the 3D descriptors RDKit provides most commonly used is in trying to determine how \"spherical\", \"disc\", or \"rod\" like a conformer is. Along these lines, everything is defined as some components of the principal moments of inertia (PM1, PM2, and PM3 with PM1 being the smallest and PM3 being the largest). See this paper: https://doi.org/10.4155/fmc-2016-0095. The plots with NPR1 and NPR2 (defined in the text as mass normalized ratios of the smaller PMIs with the larger PMI) are rather common in chemical literature concerning the \"general\" shape of a conformer. It's worth mentioning that NPR1 and NPR2 are two of the 10 descriptors available in RDKit, which exemplify how the rest of the descriptors are just built on the 3 PMIs. I hope this helps?
\nThis is some good food for thought on some notebook work, thank you! I'll try a few things and if anything interesting arises, I'll post it :).
", "rawMarkdown": "Hey @roberthatch \n\nUnfortunately, I'm not aware of any open source software that can compete with the speed of ROCS or its GPU accelerated cousin, fastROCS. You're right, these tools are state of the art and commonly applied in industry (a usual workflow is to use OMEGA to get conformers, then use ROCS to score the overlay to some known or hypothesized binding mode of a reference molecule).\n\nThat said, this paper may be useful to you: https://doi.org/10.1186/s13321-023-00703-1. Within, they describe using RDKit's rdShapeHelpers and 3D fingerprints to get a \"combo score\" (paper: \"the average of shape similarity and 3D fingerprint similarity\"), which to me looks conceptually similar to the \"shape and color\" score you'd get from ROCS.\n\nOn the more general topic, the way I've seen the 3D descriptors RDKit provides most commonly used is in trying to determine how \"spherical\", \"disc\", or \"rod\" like a conformer is. Along these lines, everything is defined as some components of the principal moments of inertia (PM1, PM2, and PM3 with PM1 being the smallest and PM3 being the largest). See this paper: https://doi.org/10.4155/fmc-2016-0095. The plots with NPR1 and NPR2 (defined in the text as mass normalized ratios of the smaller PMIs with the larger PMI) are rather common in chemical literature concerning the \"general\" shape of a conformer. It's worth mentioning that NPR1 and NPR2 are two of the 10 descriptors available in RDKit, which exemplify how the rest of the descriptors are just built on the 3 PMIs. I hope this helps?\n\nThis is some good food for thought on some notebook work, thank you! I'll try a few things and if anything interesting arises, I'll post it :).", "votes": null }, { "id": "2837756", "postDate": "05/26/2024 16:46:53", "content": "[paper] Electrostatic‑field and surface‑shape similarity for virtual screening and pose prediction
\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC6856045/pdf/10822_2019_Article_236.pdf
https://github.com/hesther/espsim
\nnot sure if this is useful or not
\n
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F041f292024c5e1e0f644425e165a425e%2FSelection_149.png?generation=1716742217297916&alt=media\")
Thanks! And welcome back :)
", "rawMarkdown": "Thanks! And welcome back :)", "votes": null }, { "id": "2837839", "postDate": "05/26/2024 17:40:22", "content": "Thank you!
", "rawMarkdown": "Thank you!", "votes": null }, { "id": "2840574", "postDate": "05/28/2024 06:32:19", "content": "PAPER Accelerates Parallel Evaluations of ROCS
\ncode: https://simtk.org/frs/?group_id=339
Sorry @hengck23 and @roberthatch I thought I was going to have more time to contribute here, but I ended up not being able to return to the competition. Thanks to both of you for all of your discussions throughout the competition. I really enjoyed sharing with you and learning from you.
", "rawMarkdown": "Sorry @hengck23 and @roberthatch I thought I was going to have more time to contribute here, but I ended up not being able to return to the competition. Thanks to both of you for all of your discussions throughout the competition. I really enjoyed sharing with you and learning from you.", "votes": null }, { "id": "2910797", "postDate": "07/07/2024 23:11:17", "content": "@chemdatafarmer I had a great time! Like you I didn't really come back to the competition - focusing instead on trying to get (and stay in) top 5 in the AutoML Grand Prix!
\nDecided to spend a last day and a half though today and tomorrow to put together a little bit of a final submission. :)
", "rawMarkdown": "chemdatafarmer I had a great time! Like you I didn't really come back to the competition - focusing instead on trying to get (and stay in) top 5 in the AutoML Grand Prix!\n\nDecided to spend a last day and a half though today and tomorrow to put together a little bit of a final submission. :)", "votes": null }, { "id": "2911527", "postDate": "07/08/2024 11:16:36", "content": "@roberthatch exciting! Best of luck in the Grand Prix and during the private leaderboard reveal here :)
", "rawMarkdown": "roberthatch exciting! Best of luck in the Grand Prix and during the private leaderboard reveal here :)", "votes": null } ], "comments": [ { "id": 2837428, "author_name": "roberthatch", "author_url": "", "post_date": "05/26/2024 13:34:23", "content": "3d descriptors and 3d similarity questions. From some of the scientific papers rabbit trails I've followed, ROCS tanimoto combo score (shape and \"color\") is mentioned and appears to be state of the art for (speed and) 3d similarity scores between two molecules.
\nAre there any good open source alternatives that you're aware of?
\nMore general topic, but I'm not sure whether I even understand the ~10 3d descriptors that rdkit can calculate? The documentation seemed sparse on them. Many of them seem related to shape but without being sufficient to describe a shape approximation? For example, an ellipse could be defined by 3 diameter numbers. But I'm not so sure that knowing PMI1/2/3 is as descriptive of the overall shape?
", "votes": null, "replies": [ { "id": 2837504, "author_name": "chemdatafarmer", "author_url": "", "post_date": "05/26/2024 14:41:27", "content": "Hey @roberthatch
\nUnfortunately, I'm not aware of any open source software that can compete with the speed of ROCS or its GPU accelerated cousin, fastROCS. You're right, these tools are state of the art and commonly applied in industry (a usual workflow is to use OMEGA to get conformers, then use ROCS to score the overlay to some known or hypothesized binding mode of a reference molecule).
\nThat said, this paper may be useful to you: https://doi.org/10.1186/s13321-023-00703-1. Within, they describe using RDKit's rdShapeHelpers and 3D fingerprints to get a \"combo score\" (paper: \"the average of shape similarity and 3D fingerprint similarity\"), which to me looks conceptually similar to the \"shape and color\" score you'd get from ROCS.
\nOn the more general topic, the way I've seen the 3D descriptors RDKit provides most commonly used is in trying to determine how \"spherical\", \"disc\", or \"rod\" like a conformer is. Along these lines, everything is defined as some components of the principal moments of inertia (PM1, PM2, and PM3 with PM1 being the smallest and PM3 being the largest). See this paper: https://doi.org/10.4155/fmc-2016-0095. The plots with NPR1 and NPR2 (defined in the text as mass normalized ratios of the smaller PMIs with the larger PMI) are rather common in chemical literature concerning the \"general\" shape of a conformer. It's worth mentioning that NPR1 and NPR2 are two of the 10 descriptors available in RDKit, which exemplify how the rest of the descriptors are just built on the 3 PMIs. I hope this helps?
\nThis is some good food for thought on some notebook work, thank you! I'll try a few things and if anything interesting arises, I'll post it :).
", "votes": null, "replies": [ { "id": 2837803, "author_name": "roberthatch", "author_url": "", "post_date": "05/26/2024 17:19:37", "content": "Thanks! And welcome back :)
", "votes": null, "replies": [ { "id": 2837839, "author_name": "chemdatafarmer", "author_url": "", "post_date": "05/26/2024 17:40:22", "content": "Thank you!
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2837756, "author_name": "hengck23", "author_url": "", "post_date": "05/26/2024 16:46:53", "content": "[paper] Electrostatic‑field and surface‑shape similarity for virtual screening and pose prediction
\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC6856045/pdf/10822_2019_Article_236.pdf
https://github.com/hesther/espsim
\nnot sure if this is useful or not
\n
PAPER Accelerates Parallel Evaluations of ROCS
\ncode: https://simtk.org/frs/?group_id=339
Sorry @hengck23 and @roberthatch I thought I was going to have more time to contribute here, but I ended up not being able to return to the competition. Thanks to both of you for all of your discussions throughout the competition. I really enjoyed sharing with you and learning from you.
", "votes": null, "replies": [ { "id": 2910797, "author_name": "roberthatch", "author_url": "", "post_date": "07/07/2024 23:11:17", "content": "@chemdatafarmer I had a great time! Like you I didn't really come back to the competition - focusing instead on trying to get (and stay in) top 5 in the AutoML Grand Prix!
\nDecided to spend a last day and a half though today and tomorrow to put together a little bit of a final submission. :)
", "votes": null, "replies": [ { "id": 2911527, "author_name": "chemdatafarmer", "author_url": "", "post_date": "07/08/2024 11:16:36", "content": "@roberthatch exciting! Best of luck in the Grand Prix and during the private leaderboard reveal here :)
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2837318": "Hey Folks,\n\nI had to take a break from this competition since before the scoring update. Looks like things have gotten pretty interesting!\n\nI'll probably dabble a bit this weekend (it's a long weekend in the US). Are there any cheminformatics questions y'all are finding particularly troublesome? Please feel free to suggest some topics you wish there was a public notebook for and I will help out if I can.", "2837428": "3d descriptors and 3d similarity questions. From some of the scientific papers rabbit trails I've followed, ROCS tanimoto combo score (shape and \"color\") is mentioned and appears to be state of the art for (speed and) 3d similarity scores between two molecules. \n\nAre there any good open source alternatives that you're aware of?\n\nMore general topic, but I'm not sure whether I even understand the ~10 3d descriptors that rdkit can calculate? The documentation seemed sparse on them. Many of them seem related to shape but without being sufficient to describe a shape approximation? For example, an ellipse could be defined by 3 diameter numbers. But I'm not so sure that knowing PMI1/2/3 is as descriptive of the overall shape?", "2837504": "Hey @roberthatch \n\nUnfortunately, I'm not aware of any open source software that can compete with the speed of ROCS or its GPU accelerated cousin, fastROCS. You're right, these tools are state of the art and commonly applied in industry (a usual workflow is to use OMEGA to get conformers, then use ROCS to score the overlay to some known or hypothesized binding mode of a reference molecule).\n\nThat said, this paper may be useful to you: https://doi.org/10.1186/s13321-023-00703-1. Within, they describe using RDKit's rdShapeHelpers and 3D fingerprints to get a \"combo score\" (paper: \"the average of shape similarity and 3D fingerprint similarity\"), which to me looks conceptually similar to the \"shape and color\" score you'd get from ROCS.\n\nOn the more general topic, the way I've seen the 3D descriptors RDKit provides most commonly used is in trying to determine how \"spherical\", \"disc\", or \"rod\" like a conformer is. Along these lines, everything is defined as some components of the principal moments of inertia (PM1, PM2, and PM3 with PM1 being the smallest and PM3 being the largest). See this paper: https://doi.org/10.4155/fmc-2016-0095. The plots with NPR1 and NPR2 (defined in the text as mass normalized ratios of the smaller PMIs with the larger PMI) are rather common in chemical literature concerning the \"general\" shape of a conformer. It's worth mentioning that NPR1 and NPR2 are two of the 10 descriptors available in RDKit, which exemplify how the rest of the descriptors are just built on the 3 PMIs. I hope this helps?\n\nThis is some good food for thought on some notebook work, thank you! I'll try a few things and if anything interesting arises, I'll post it :).", "2837756": "[paper] Electrostatic‑field and surface‑shape similarity for virtual screening and pose prediction\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC6856045/pdf/10822_2019_Article_236.pdf\n\nhttps://github.com/hesther/espsim\n\nnot sure if this is useful or not\n\n\n\n", "2837803": "Thanks! And welcome back :)", "2837839": "Thank you!", "2840574": "PAPER Accelerates Parallel Evaluations of ROCS\ncode: https://simtk.org/frs/?group_id=339", "2904693": "Sorry @hengck23 and @roberthatch I thought I was going to have more time to contribute here, but I ended up not being able to return to the competition. Thanks to both of you for all of your discussions throughout the competition. I really enjoyed sharing with you and learning from you.", "2910797": "chemdatafarmer I had a great time! Like you I didn't really come back to the competition - focusing instead on trying to get (and stay in) top 5 in the AutoML Grand Prix!\n\nDecided to spend a last day and a half though today and tomorrow to put together a little bit of a final submission. :)", "2911527": "roberthatch exciting! Best of luck in the Grand Prix and during the private leaderboard reveal here :)" }, "source": "meta" }