Question to folks with chemistry knowledge - I'm using RDKit to process the molecule smiles and extract atoms and features. Given that the DNA linker is encoded as [Dy], I believe RDKit assumes it's a dysprosium atom, and the feature vector for it will be wrong. Is my understanding correct? Would it make sense to try any pre/post-processing to handle that?
", "messages": [ { "id": "2796389", "postDate": "05/06/2024 07:49:07", "content": "Question to folks with chemistry knowledge - I'm using RDKit to process the molecule smiles and extract atoms and features. Given that the DNA linker is encoded as [Dy], I believe RDKit assumes it's a dysprosium atom, and the feature vector for it will be wrong. Is my understanding correct? Would it make sense to try any pre/post-processing to handle that?
", "rawMarkdown": "Question to folks with chemistry knowledge - I'm using RDKit to process the molecule smiles and extract atoms and features. Given that the DNA linker is encoded as [Dy], I believe RDKit assumes it's a dysprosium atom, and the feature vector for it will be wrong. Is my understanding correct? Would it make sense to try any pre/post-processing to handle that?", "votes": null }, { "id": "2796432", "postDate": "05/06/2024 08:21:24", "content": "Yes, RDKit considers it as a metal atom. This would pose problems if you use physchem descriptors (molecular weight, lipophilicity, etc). One common approach to handle this is to replace Dy with a methyl group.
\n rdkit Chem\n rdkit.Chem AllChem\n ():\n\n test_molecule = Chem.MolFromSmiles(smi)\n\n Me = Chem.MolFromSmiles()\n\n \n Dy = Chem.MolFromSmiles()\n\n new_mol = AllChem.ReplaceSubstructs(test_molecule, Dy, Me)\n\n Chem.MolToSmiles(new_mol[])\nThanks a lot!
", "rawMarkdown": "Thanks a lot!", "votes": null }, { "id": "2796693", "postDate": "05/06/2024 11:28:47", "content": "Like @ruelcedeno said, if you want to use 3D structures or descriptors it definitely matters. I made a notebook about how to handle this: cheminformatics transformations
\nFor 2D descriptors like SMILES or fingerprints I think it's ok to leave it if you assume the attachment point doesn't contribute to binding much, though I haven't experimented much with this yet. I played around with ethylene glycol derived descriptors for 2D and I achieved slightly higher LB scores when I replaced it, but I haven't dug deeply into this yet in a systematic way.
", "rawMarkdown": "Like @ruelcedeno said, if you want to use 3D structures or descriptors it definitely matters. I made a notebook about how to handle this: [cheminformatics transformations](https://www.kaggle.com/code/chemdatafarmer/cheminformatics-transformations)\n\nFor 2D descriptors like SMILES or fingerprints I think it's ok to leave it if you assume the attachment point doesn't contribute to binding much, though I haven't experimented much with this yet. I played around with ethylene glycol derived descriptors for 2D and I achieved slightly higher LB scores when I replaced it, but I haven't dug deeply into this yet in a systematic way.", "votes": null } ], "comments": [ { "id": 2796432, "author_name": "ruelcedeno", "author_url": "", "post_date": "05/06/2024 08:21:24", "content": "Yes, RDKit considers it as a metal atom. This would pose problems if you use physchem descriptors (molecular weight, lipophilicity, etc). One common approach to handle this is to replace Dy with a methyl group.
\n rdkit Chem\n rdkit.Chem AllChem\n ():\n\n test_molecule = Chem.MolFromSmiles(smi)\n\n Me = Chem.MolFromSmiles()\n\n \n Dy = Chem.MolFromSmiles()\n\n new_mol = AllChem.ReplaceSubstructs(test_molecule, Dy, Me)\n\n Chem.MolToSmiles(new_mol[])\nThanks a lot!
", "votes": null, "replies": [] } ] }, { "id": 2796693, "author_name": "chemdatafarmer", "author_url": "", "post_date": "05/06/2024 11:28:47", "content": "Like @ruelcedeno said, if you want to use 3D structures or descriptors it definitely matters. I made a notebook about how to handle this: cheminformatics transformations
\nFor 2D descriptors like SMILES or fingerprints I think it's ok to leave it if you assume the attachment point doesn't contribute to binding much, though I haven't experimented much with this yet. I played around with ethylene glycol derived descriptors for 2D and I achieved slightly higher LB scores when I replaced it, but I haven't dug deeply into this yet in a systematic way.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2796389": "Question to folks with chemistry knowledge - I'm using RDKit to process the molecule smiles and extract atoms and features. Given that the DNA linker is encoded as [Dy], I believe RDKit assumes it's a dysprosium atom, and the feature vector for it will be wrong. Is my understanding correct? Would it make sense to try any pre/post-processing to handle that?", "2796432": "Yes, RDKit considers it as a metal atom. This would pose problems if you use physchem descriptors (molecular weight, lipophilicity, etc). One common approach to handle this is to replace Dy with a methyl group.\n\n\n```python\nfrom rdkit import Chem\nfrom rdkit.Chem import AllChem\ndef transform_Dy(smi):\n\n test_molecule = Chem.MolFromSmiles(smi)\n\n Me = Chem.MolFromSmiles('C')\n\n #create Dy mol object\n Dy = Chem.MolFromSmiles('[Dy]')\n\n new_mol = AllChem.ReplaceSubstructs(test_molecule, Dy, Me)\n\n return Chem.MolToSmiles(new_mol[0])\n```", "2796450": "Thanks a lot!", "2796693": "Like @ruelcedeno said, if you want to use 3D structures or descriptors it definitely matters. I made a notebook about how to handle this: [cheminformatics transformations](https://www.kaggle.com/code/chemdatafarmer/cheminformatics-transformations)\n\nFor 2D descriptors like SMILES or fingerprints I think it's ok to leave it if you assume the attachment point doesn't contribute to binding much, though I haven't experimented much with this yet. I played around with ethylene glycol derived descriptors for 2D and I achieved slightly higher LB scores when I replaced it, but I haven't dug deeply into this yet in a systematic way." }, "source": "meta" }