![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F297623%2F458fb7d08589d7d3d4ca4960db708cc0%2FScreenshot%20from%202019-07-07%2011-19-14.png?generation=1562516371811678&alt=media\")
What is DeepChem?
\n\nDeepChem aims to provide a high quality open-source toolchain that democratizes the use of deep-learning in drug discovery, materials science, quantum chemistry, and biology.
\n\nWhat is ChemNet?\nChemNet is a transferable and generalizable deep neural network for small-molecule property prediction. The idea presented in this: ChemNet paper\nis to train a deep neural network on Tox21, HIV and FreeSolv datasets which are large data sets for chemical properties of the molecules and then use transfer learning to predict the molecule properties of unknown molecules.
\n\nI hope this picture from the above paper gives justice to the idea presented in the paper.
\n\n
The above paper has been incorporated into the DeepChem repository recently here: https://github.com/deepchem/deepchem/blob/master/deepchem/models/tensorgraph/models/chemnet_models.py
\n\nWhat I have done so far?\nMy idea is as follows:
\n\nTake Molecule + scalar coupling constant property (from this notebook -> Convert it into a image -> feed it into ChemCeption Model (as per above code) -> fit -> evaluate (predict scalar coupling constant)
\n\nWhere I am stuck?\nI am stuck at the convert molecule with properties into image stage. I was wondering if some one would be kind enough to help me walk through that code.
\n\nI thought of sharing this idea to help people use another type of models in this competition and honestly, I am impressed by the work done in the above paper and the code and I am itching to try out the ChemNet DNN model on this data set.
\n\nThank you for reading till here.
", "messages": [ { "id": "569983", "postDate": "07/07/2019 16:27:57", "content": "What is DeepChem?
\n\nDeepChem aims to provide a high quality open-source toolchain that democratizes the use of deep-learning in drug discovery, materials science, quantum chemistry, and biology.
\n\nWhat is ChemNet?\nChemNet is a transferable and generalizable deep neural network for small-molecule property prediction. The idea presented in this: ChemNet paper\nis to train a deep neural network on Tox21, HIV and FreeSolv datasets which are large data sets for chemical properties of the molecules and then use transfer learning to predict the molecule properties of unknown molecules.
\n\nI hope this picture from the above paper gives justice to the idea presented in the paper.
\n\n
The above paper has been incorporated into the DeepChem repository recently here: https://github.com/deepchem/deepchem/blob/master/deepchem/models/tensorgraph/models/chemnet_models.py
\n\nWhat I have done so far?\nMy idea is as follows:
\n\nTake Molecule + scalar coupling constant property (from this notebook -> Convert it into a image -> feed it into ChemCeption Model (as per above code) -> fit -> evaluate (predict scalar coupling constant)
\n\nWhere I am stuck?\nI am stuck at the convert molecule with properties into image stage. I was wondering if some one would be kind enough to help me walk through that code.
\n\nI thought of sharing this idea to help people use another type of models in this competition and honestly, I am impressed by the work done in the above paper and the code and I am itching to try out the ChemNet DNN model on this data set.
\n\nThank you for reading till here.
", "rawMarkdown": "What is DeepChem?\n\nDeepChem aims to provide a high quality open-source toolchain that democratizes the use of deep-learning in drug discovery, materials science, quantum chemistry, and biology.\n\nWhat is ChemNet?\nChemNet is a transferable and generalizable deep neural network for small-molecule property prediction. The idea presented in this: [ChemNet paper](https://www.groundai.com/project/chemnet-a-transferable-and-generalizable-deep-neural-network-for-small-molecule-property-prediction/)\nis to train a deep neural network on Tox21, HIV and FreeSolv datasets which are large data sets for chemical properties of the molecules and then use transfer learning to predict the molecule properties of unknown molecules. \n\nI hope this picture from the above paper gives justice to the idea presented in the paper. \n\n\n\nThe above paper has been incorporated into the DeepChem repository recently here: https://github.com/deepchem/deepchem/blob/master/deepchem/models/tensorgraph/models/chemnet_models.py\n\nWhat I have done so far?\nMy idea is as follows:\n\nTake Molecule + scalar coupling constant property (from this [notebook](https://www.kaggle.com/asauve/training-set-molecule-visualization/data) -> Convert it into a image -> feed it into ChemCeption Model (as per above code) -> fit -> evaluate (predict scalar coupling constant)\n\n\nWhere I am stuck?\nI am stuck at the convert molecule with properties into image stage. I was wondering if some one would be kind enough to help me walk through that code. \n\nI thought of sharing this idea to help people use another type of models in this competition and honestly, I am impressed by the work done in the above paper and the code and I am itching to try out the ChemNet DNN model on this data set. \n\nThank you for reading till here.", "votes": null }, { "id": "570115", "postDate": "07/07/2019 19:34:31", "content": "Interesting, although it seems that they mostly use this to predict bulk properties, not specific bonds as is needed in this comp.\n You could try and train one model per coupling type, but I don't see how you would differentiate if you need to predict multiple of the same type of couplings in one \"image\"
\n\nIf you want to go ahead, they say they used RDKit to make the 2D images: \nhttps://www.rdkit.org/docs/GettingStartedInPython.html#working-with-2d-molecules-generating-depictions
", "rawMarkdown": "Interesting, although it seems that they mostly use this to predict bulk properties, not specific bonds as is needed in this comp.\n You could try and train one model per coupling type, but I don't see how you would differentiate if you need to predict multiple of the same type of couplings in one \"image\"\n\nIf you want to go ahead, they say they used RDKit to make the 2D images: \nhttps://www.rdkit.org/docs/GettingStartedInPython.html#working-with-2d-molecules-generating-depictions", "votes": null }, { "id": "570119", "postDate": "07/07/2019 19:51:39", "content": "Thank you for your insight, I'll give it a try and see if it works.
", "rawMarkdown": "Thank you for your insight, I'll give it a try and see if it works.", "votes": null }, { "id": "570126", "postDate": "07/07/2019 20:29:23", "content": "Just to check, but how do you expect to predict separate couplings (i.e. a value that exists between two atoms, and of which there are multiple per molecule) from an image? You would have to somehow highlight that specific bond...
", "rawMarkdown": "Just to check, but how do you expect to predict separate couplings (i.e. a value that exists between two atoms, and of which there are multiple per molecule) from an image? You would have to somehow highlight that specific bond...", "votes": null }, { "id": "570501", "postDate": "07/08/2019 11:40:02", "content": "I just create a kernel showing how to use deepchem to predict additional data for train and test.
\n\nhttps://www.kaggle.com/gtteixeira/deepchem-using-additional-data-for-train-and-test\nAlso I teach how to create 3D molecules.
", "rawMarkdown": "I just create a kernel showing how to use deepchem to predict additional data for train and test.\n\nhttps://www.kaggle.com/gtteixeira/deepchem-using-additional-data-for-train-and-test\nAlso I teach how to create 3D molecules.", "votes": null }, { "id": "854120", "postDate": "05/19/2020 19:20:22", "content": "I tried to install the deepchem library following the notebook introductions. While loading the Tox21 dataset by
\n\nfrom deepchem.molnet import load_tox21
load_tox21()
i get the error message AttributeError: module 'deepchem' has no attribute 'splits'.
\n\nFurther the command import deepchem yield an error.
\n\nDo you know why i will get these errors?
", "rawMarkdown": "I tried to install the deepchem library following the notebook introductions. While loading the Tox21 dataset by\n\n`from deepchem.molnet import load_tox21`\n\n`load_tox21()`\n\ni get the error message AttributeError: module 'deepchem' has no attribute 'splits'.\n\nFurther the command import deepchem yield an error.\n\nDo you know why i will get these errors?", "votes": null } ], "comments": [ { "id": 570115, "author_name": "fnands", "author_url": "", "post_date": "07/07/2019 19:34:31", "content": "Interesting, although it seems that they mostly use this to predict bulk properties, not specific bonds as is needed in this comp.\n You could try and train one model per coupling type, but I don't see how you would differentiate if you need to predict multiple of the same type of couplings in one \"image\"
\n\nIf you want to go ahead, they say they used RDKit to make the 2D images: \nhttps://www.rdkit.org/docs/GettingStartedInPython.html#working-with-2d-molecules-generating-depictions
", "votes": null, "replies": [ { "id": 570119, "author_name": "vikasmalhotra08", "author_url": "", "post_date": "07/07/2019 19:51:39", "content": "Thank you for your insight, I'll give it a try and see if it works.
", "votes": null, "replies": [] }, { "id": 570126, "author_name": "fnands", "author_url": "", "post_date": "07/07/2019 20:29:23", "content": "Just to check, but how do you expect to predict separate couplings (i.e. a value that exists between two atoms, and of which there are multiple per molecule) from an image? You would have to somehow highlight that specific bond...
", "votes": null, "replies": [] } ] }, { "id": 570501, "author_name": "gtteixeira", "author_url": "", "post_date": "07/08/2019 11:40:02", "content": "I just create a kernel showing how to use deepchem to predict additional data for train and test.
\n\nhttps://www.kaggle.com/gtteixeira/deepchem-using-additional-data-for-train-and-test\nAlso I teach how to create 3D molecules.
", "votes": null, "replies": [] }, { "id": 854120, "author_name": "drcapa", "author_url": "", "post_date": "05/19/2020 19:20:22", "content": "I tried to install the deepchem library following the notebook introductions. While loading the Tox21 dataset by
\n\nfrom deepchem.molnet import load_tox21
load_tox21()
i get the error message AttributeError: module 'deepchem' has no attribute 'splits'.
\n\nFurther the command import deepchem yield an error.
\n\nDo you know why i will get these errors?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "569983": "What is DeepChem?\n\nDeepChem aims to provide a high quality open-source toolchain that democratizes the use of deep-learning in drug discovery, materials science, quantum chemistry, and biology.\n\nWhat is ChemNet?\nChemNet is a transferable and generalizable deep neural network for small-molecule property prediction. The idea presented in this: [ChemNet paper](https://www.groundai.com/project/chemnet-a-transferable-and-generalizable-deep-neural-network-for-small-molecule-property-prediction/)\nis to train a deep neural network on Tox21, HIV and FreeSolv datasets which are large data sets for chemical properties of the molecules and then use transfer learning to predict the molecule properties of unknown molecules. \n\nI hope this picture from the above paper gives justice to the idea presented in the paper. \n\n\n\nThe above paper has been incorporated into the DeepChem repository recently here: https://github.com/deepchem/deepchem/blob/master/deepchem/models/tensorgraph/models/chemnet_models.py\n\nWhat I have done so far?\nMy idea is as follows:\n\nTake Molecule + scalar coupling constant property (from this [notebook](https://www.kaggle.com/asauve/training-set-molecule-visualization/data) -> Convert it into a image -> feed it into ChemCeption Model (as per above code) -> fit -> evaluate (predict scalar coupling constant)\n\n\nWhere I am stuck?\nI am stuck at the convert molecule with properties into image stage. I was wondering if some one would be kind enough to help me walk through that code. \n\nI thought of sharing this idea to help people use another type of models in this competition and honestly, I am impressed by the work done in the above paper and the code and I am itching to try out the ChemNet DNN model on this data set. \n\nThank you for reading till here.", "570115": "Interesting, although it seems that they mostly use this to predict bulk properties, not specific bonds as is needed in this comp.\n You could try and train one model per coupling type, but I don't see how you would differentiate if you need to predict multiple of the same type of couplings in one \"image\"\n\nIf you want to go ahead, they say they used RDKit to make the 2D images: \nhttps://www.rdkit.org/docs/GettingStartedInPython.html#working-with-2d-molecules-generating-depictions", "570119": "Thank you for your insight, I'll give it a try and see if it works.", "570126": "Just to check, but how do you expect to predict separate couplings (i.e. a value that exists between two atoms, and of which there are multiple per molecule) from an image? You would have to somehow highlight that specific bond...", "570501": "I just create a kernel showing how to use deepchem to predict additional data for train and test.\n\nhttps://www.kaggle.com/gtteixeira/deepchem-using-additional-data-for-train-and-test\nAlso I teach how to create 3D molecules.", "854120": "I tried to install the deepchem library following the notebook introductions. While loading the Tox21 dataset by\n\n`from deepchem.molnet import load_tox21`\n\n`load_tox21()`\n\ni get the error message AttributeError: module 'deepchem' has no attribute 'splits'.\n\nFurther the command import deepchem yield an error.\n\nDo you know why i will get these errors?" }, "source": "meta" }