![\"GeminiMol\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12202244%2Ffbda8e551b5a186a68ce3e878c437d4c%2FExtended_Figure_4.JPG?generation=1713851448085195&alt=media\")
Edit: GeminiMol is Academic Free, due to licensing restrictions, the usage of non-commercially available software may result in ineligibility for prize. Thanks for pointing this out. @hengck23 @shlomoron
\n===============================================
\nUp to this point, I consider the objective of this competition to be a QSAR (quantitative structure-activity relationship) task rather than a DTI (drug-target interaction) task. Therefore, effective small molecular representation, especially those capable of capturing scaffold hopping potential, is of paramount importance. Unsurprisingly, this is also a research area of great interest within the drug discovery community.
\nIndeed, the combination of multiple molecular fingerprints, such as ECFP4+MACCS+TopoTorsion, along with a powerful machine learning method like LightGBM, is sufficient to achieve a good performance in most QSAR tasks. In this competition as well. In additation to molecular fingerprints, you may be exploring alternative molecular representation models, which acquire an understanding (or perhaps just memorization?) of molecular structures through pre-training tasks. These models aim to enhance the predictive performance or generalization ability of QSAR models.
\nOver the past year, we have proposed a model called GeminiMol, which achieves impressive performance on various downstream benchmarks of millions of molecules, despite being pre-trained on only 39,290 molecular structures. These benchmarks include zero-shot molecular similarity comparison tasks, target-level QSAR tasks (similar to this competition), cell-level QSAR tasks, and ADMET property prediction tasks. If any participants have the time to explore novel molecular representation methods, I think GeminiMol may be of assistance to you. You can find the scripts, models and datasets from our GitHub repo: https://github.com/Wang-Lin-boop/GeminiMol?tab=readme-ov-file
\n
GeminiMol employs intermolecular contrastive learning for pre-training. We utilize a graph neural network called WLN to extract molecular features and use these features to predict the 2D maximum common substructure similarity and the 3D conformational space similarity of molecule pairs (represented by shape and pharmacophore similarity). This training strategy allows the model to capture the 3D similarity between molecule pairs with significant 2D structural differences, which may aid in capturing molecular scaffold hopping.
\nRead more here: Conformational Space Profile Enhances Generic Molecular Representation Learning\n
", "messages": [ { "id": "2769155", "postDate": "04/23/2024 07:36:48", "content": "Edit: GeminiMol is Academic Free, due to licensing restrictions, the usage of non-commercially available software may result in ineligibility for prize. Thanks for pointing this out. @hengck23 @shlomoron
\n===============================================
\nUp to this point, I consider the objective of this competition to be a QSAR (quantitative structure-activity relationship) task rather than a DTI (drug-target interaction) task. Therefore, effective small molecular representation, especially those capable of capturing scaffold hopping potential, is of paramount importance. Unsurprisingly, this is also a research area of great interest within the drug discovery community.
\nIndeed, the combination of multiple molecular fingerprints, such as ECFP4+MACCS+TopoTorsion, along with a powerful machine learning method like LightGBM, is sufficient to achieve a good performance in most QSAR tasks. In this competition as well. In additation to molecular fingerprints, you may be exploring alternative molecular representation models, which acquire an understanding (or perhaps just memorization?) of molecular structures through pre-training tasks. These models aim to enhance the predictive performance or generalization ability of QSAR models.
\nOver the past year, we have proposed a model called GeminiMol, which achieves impressive performance on various downstream benchmarks of millions of molecules, despite being pre-trained on only 39,290 molecular structures. These benchmarks include zero-shot molecular similarity comparison tasks, target-level QSAR tasks (similar to this competition), cell-level QSAR tasks, and ADMET property prediction tasks. If any participants have the time to explore novel molecular representation methods, I think GeminiMol may be of assistance to you. You can find the scripts, models and datasets from our GitHub repo: https://github.com/Wang-Lin-boop/GeminiMol?tab=readme-ov-file
\n
GeminiMol employs intermolecular contrastive learning for pre-training. We utilize a graph neural network called WLN to extract molecular features and use these features to predict the 2D maximum common substructure similarity and the 3D conformational space similarity of molecule pairs (represented by shape and pharmacophore similarity). This training strategy allows the model to capture the 3D similarity between molecule pairs with significant 2D structural differences, which may aid in capturing molecular scaffold hopping.
\nRead more here: Conformational Space Profile Enhances Generic Molecular Representation Learning\n
", "rawMarkdown": "Edit: GeminiMol is Academic Free, due to licensing restrictions, the usage of non-commercially available software may result in ineligibility for prize. Thanks for pointing this out. @hengck23 @shlomoron \n\n===============================================\n\n\nUp to this point, I consider the objective of this competition to be a QSAR (quantitative structure-activity relationship) task rather than a DTI (drug-target interaction) task. Therefore, effective small molecular representation, especially those capable of capturing scaffold hopping potential, is of paramount importance. Unsurprisingly, this is also a research area of great interest within the drug discovery community.\n\nIndeed, the combination of multiple molecular fingerprints, such as ECFP4+MACCS+TopoTorsion, along with a powerful machine learning method like LightGBM, is sufficient to achieve a good performance in most QSAR tasks. [In this competition as well](https://www.kaggle.com/competitions/leash-BELKA/discussion/492846). In additation to molecular fingerprints, you may be exploring alternative molecular representation models, which acquire an **understanding (or perhaps just memorization?)** of molecular structures through pre-training tasks. These models aim to enhance the predictive performance or generalization ability of QSAR models. \n\nOver the past year, we have proposed a model called [**GeminiMol**](https://github.com/Wang-Lin-boop/GeminiMol), which achieves impressive performance on various downstream benchmarks of millions of molecules, despite being pre-trained on only 39,290 molecular structures. These benchmarks include zero-shot molecular similarity comparison tasks, target-level QSAR tasks (similar to this competition), cell-level QSAR tasks, and ADMET property prediction tasks. If any participants have the time to explore novel molecular representation methods, I think GeminiMol may be of assistance to you. You can find the scripts, models and datasets from our GitHub repo: https://github.com/Wang-Lin-boop/GeminiMol?tab=readme-ov-file \n\n\n\nGeminiMol employs intermolecular contrastive learning for pre-training. We utilize a graph neural network called WLN to extract molecular features and use these features to predict the 2D maximum common substructure similarity and the **3D conformational space similarity of molecule pairs (represented by shape and pharmacophore similarity)**. This training strategy allows the model to capture the 3D similarity between molecule pairs with significant 2D structural differences, which may aid in capturing molecular scaffold hopping.\n\nRead more here: [Conformational Space Profile Enhances Generic Molecular Representation Learning\n](https://www.biorxiv.org/content/10.1101/2023.12.14.571629)", "votes": null }, { "id": "2769220", "postDate": "04/23/2024 08:09:47", "content": "\"combination of multiple molecular fingerprints, such as ECFP4+MACCS+TopoTorsion,\"
\ni don't think this QSAR alone is not good enough for medal at the end of competition at the end of 3 months.
i don't think QSAR task will out-performed DTI.
\nFurther, molecular fingerprints may work poorly for unseen test blocks.
Nevertheless, QSAR is good as a starting benchmark.
", "rawMarkdown": "\"combination of multiple molecular fingerprints, such as ECFP4+MACCS+TopoTorsion,\"\ni don't think this QSAR alone is not good enough for medal at the end of competition at the end of 3 months.\n\ni don't think QSAR task will out-performed DTI.\nFurther, molecular fingerprints may work poorly for unseen test blocks.\n\nNevertheless, QSAR is good as a starting benchmark.", "votes": null }, { "id": "2769254", "postDate": "04/23/2024 08:43:57", "content": "Undeniably, for the drug discovery community, the significance of DTI models surpasses that of QSAR models (at least in terms of research popularity, haha). This why I said \"up to this point\". Cautiously speaking, if there are a sufficient number of molecules in the test set that exhibit binding sites or conformations that differ from those in the training set, it is indeed possible for DTI models to outperform QSAR models.
\nIn such cases, a sufficiently robust DTI model, capable of surpassing QSAR models on known binding sites, and capable of generalizing to these unknown sites and conformations, would be competitive. However, training such a model would require a larger dataset of protein-ligand pairs, which I believe is quite challenging. In my opinion, it is insufficient to solely utilize the three proteins provided in the training set to train a model that can cover the novel mechanisms of action in the test set.
", "rawMarkdown": "Undeniably, for the drug discovery community, the significance of DTI models surpasses that of QSAR models (at least in terms of research popularity, haha). This why I said \"up to this point\". Cautiously speaking, if there are a sufficient number of molecules in the test set that exhibit binding sites or conformations that differ from those in the training set, it is indeed possible for DTI models to outperform QSAR models. \n\nIn such cases, a sufficiently robust DTI model, capable of surpassing QSAR models on known binding sites, and capable of generalizing to these unknown sites and conformations, would be competitive. However, training such a model would require a larger dataset of protein-ligand pairs, which I believe is quite challenging. In my opinion, it is insufficient to solely utilize the three proteins provided in the training set to train a model that can cover the novel mechanisms of action in the test set.", "votes": null }, { "id": "2771040", "postDate": "04/24/2024 05:30:58", "content": "i find this:
\n\"GeminiMol is released under the Academic Free Licence, which permits academic use, modification and distribution free of charge, but prohibits unauthorised commercial use, including commercial training and as part of a paid computational platform. \"
\nhttps://huggingface.co/AlphaMWang/GeminiMol-MOD
\nHence it think this cannot be used for this kaggle competition
", "rawMarkdown": "i find this:\n\n\"GeminiMol is released under the Academic Free Licence, which permits academic use, modification and distribution free of charge, but prohibits unauthorised commercial use, including commercial training and as part of a paid computational platform. \"\n\nhttps://huggingface.co/AlphaMWang/GeminiMol-MOD\n\nHence it think this cannot be used for this kaggle competition", "votes": null }, { "id": "2771228", "postDate": "04/24/2024 06:45:18", "content": "Excuse me, I don't participate in Kaggle competitions very often. Would Kaggle competitions be considered as a commercial use? I do not consider the use of GeminiMol for Kaggle to be commercial. The prohibition on the use of this platform for commercial training and paid computing platformis solely intended to prevent speculators from exploiting informational asymmetry for profit.
", "rawMarkdown": "Excuse me, I don't participate in Kaggle competitions very often. Would Kaggle competitions be considered as a commercial use? I do not consider the use of GeminiMol for Kaggle to be commercial. The prohibition on the use of this platform for commercial training and paid computing platformis solely intended to prevent speculators from exploiting informational asymmetry for profit.", "votes": null }, { "id": "2771332", "postDate": "04/24/2024 07:40:58", "content": "yes. participation of kaggle is considered as commercial.
\n(e.g. if you use software, dataset etc that is not licensed for commercial use, you will not get the prize money)
but it is better to wait for the organizer and the host to issue an official notice.
\nyou may want to ask the organizer and host for clarification, e.g. by tagging them (with @) in your question.
an famous example is GPL license of yolov5, which causes confusion in the past competition:
\nhttps://www.kaggle.com/c/global-wheat-detection/discussion/163433#919074
it is incredibly confusing. Sometimes the competition itself is deemed as non-commerical, and hence any software can be use. so it is better to ask the organizer and kaggle staffs.
", "rawMarkdown": "yes. participation of kaggle is considered as commercial.\n(e.g. if you use software, dataset etc that is not licensed for commercial use, you will not get the prize money)\n\nbut it is better to wait for the organizer and the host to issue an official notice.\nyou may want to ask the organizer and host for clarification, e.g. by tagging them (with @) in your question.\n\n----\n\nan famous example is GPL license of yolov5, which causes confusion in the past competition:\nhttps://www.kaggle.com/c/global-wheat-detection/discussion/163433#919074\n\n---\n\nit is incredibly confusing. Sometimes the competition itself is deemed as non-commerical, and hence any software can be use. so it is better to ask the organizer and kaggle staffs.", "votes": null }, { "id": "2771409", "postDate": "04/24/2024 08:26:45", "content": "OK, I modified GeminiMol's license statement to allow use in open source competitions. Thanks for your comments !
", "rawMarkdown": "OK, I modified GeminiMol's license statement to allow use in open source competitions. Thanks for your comments !", "votes": null }, { "id": "2771586", "postDate": "04/24/2024 10:15:34", "content": "Good catch @hengck23 . @wanglinboop , changing the license to allow use in a competition is not enough. Since this competition is organized by a for-profit company, the solutions (obviously) have to be such that the company can make use of…this is why section A1 in the rules (winner license) demands our solution to be licensed under MIT up to \"generally commercially available software not owned by you that you used to generate your submission\". Open-source that are not available commercially are no good in most Kaggle competitions.
", "rawMarkdown": "Good catch @hengck23 . @wanglinboop , changing the license to allow use in a *competition* is not enough. Since this competition is organized by a for-profit company, the solutions (obviously) have to be such that the company can make use of...this is why section A1 in the rules (winner license) demands our solution to be licensed under MIT up to \"generally commercially available software not owned by you that you used to generate your submission\". Open-source that are not available commercially are no good in most Kaggle competitions.", "votes": null }, { "id": "2771656", "postDate": "04/24/2024 10:50:51", "content": "https://www.kaggle.com/competitions/leash-BELKA/rules
\nWINNER LICENSE TYPE: MIT Open Source
Under Section 11 (Winners Obligations) of the General Rules below, you hereby grant and will grant Competition Sponsor the following license(s) with respect to your Submission if you are a Competition winner:
\nOpen Source: You hereby license and will license your winning Submission and the source code used to generate the Submission under an MIT Open Source license that in no event limits commercial use of such code or model containing or depending on such code. To the extent your Submission makes use of generally commercially available software not owned by you that you used to generate your submission, but that can be procured by the Competition Sponsor without undue expense, you do not grant the license in the preceding sentence to that software.
\nThis is a bit special.
\nNow @wanglinboop is the creator of the software and own the license of GeminiMol.
\nso if wanglinboop wins, he need to grant the license of his solution (which could be a modified version of GeminiMol)
But if other kaggler uses GeminiMol and win, they cannot change the license.
\nI am not a lawyer, so i cannot confirm if the above this correct. in any case, do obtain confirmation from kaggle staffs and host.
", "rawMarkdown": "https://www.kaggle.com/competitions/leash-BELKA/rules\nWINNER LICENSE TYPE: MIT Open Source\n\n1. WINNER LICENSE.\n\nUnder Section 11 (Winners Obligations) of the General Rules below, you hereby grant and will grant Competition Sponsor the following license(s) with respect to your Submission if you are a Competition winner:\n\nOpen Source: You hereby license and will license your winning Submission and the source code used to generate the Submission under an MIT Open Source license that in no event limits commercial use of such code or model containing or depending on such code. To the extent your Submission makes use of generally commercially available software not owned by you that you used to generate your submission, but that can be procured by the Competition Sponsor without undue expense, you do not grant the license in the preceding sentence to that software.\n\n----\nThis is a bit special.\nNow @wanglinboop is the creator of the software and own the license of GeminiMol.\nso if wanglinboop wins, he need to grant the license of his solution (which could be a modified version of GeminiMol)\n\nBut if other kaggler uses GeminiMol and win, they cannot change the license.\n\nI am not a lawyer, so i cannot confirm if the above this correct. in any case, do obtain confirmation from kaggle staffs and host.", "votes": null }, { "id": "2771677", "postDate": "04/24/2024 11:04:13", "content": "Thanks. I have made revisions to indicate this point. 🤣
", "rawMarkdown": "Thanks. I have made revisions to indicate this point. 🤣", "votes": null } ], "comments": [ { "id": 2769220, "author_name": "hengck23", "author_url": "", "post_date": "04/23/2024 08:09:47", "content": "\"combination of multiple molecular fingerprints, such as ECFP4+MACCS+TopoTorsion,\"
\ni don't think this QSAR alone is not good enough for medal at the end of competition at the end of 3 months.
i don't think QSAR task will out-performed DTI.
\nFurther, molecular fingerprints may work poorly for unseen test blocks.
Nevertheless, QSAR is good as a starting benchmark.
", "votes": null, "replies": [ { "id": 2769254, "author_name": "wanglinboop", "author_url": "", "post_date": "04/23/2024 08:43:57", "content": "Undeniably, for the drug discovery community, the significance of DTI models surpasses that of QSAR models (at least in terms of research popularity, haha). This why I said \"up to this point\". Cautiously speaking, if there are a sufficient number of molecules in the test set that exhibit binding sites or conformations that differ from those in the training set, it is indeed possible for DTI models to outperform QSAR models.
\nIn such cases, a sufficiently robust DTI model, capable of surpassing QSAR models on known binding sites, and capable of generalizing to these unknown sites and conformations, would be competitive. However, training such a model would require a larger dataset of protein-ligand pairs, which I believe is quite challenging. In my opinion, it is insufficient to solely utilize the three proteins provided in the training set to train a model that can cover the novel mechanisms of action in the test set.
", "votes": null, "replies": [] } ] }, { "id": 2771040, "author_name": "hengck23", "author_url": "", "post_date": "04/24/2024 05:30:58", "content": "i find this:
\n\"GeminiMol is released under the Academic Free Licence, which permits academic use, modification and distribution free of charge, but prohibits unauthorised commercial use, including commercial training and as part of a paid computational platform. \"
\nhttps://huggingface.co/AlphaMWang/GeminiMol-MOD
\nHence it think this cannot be used for this kaggle competition
", "votes": null, "replies": [ { "id": 2771228, "author_name": "wanglinboop", "author_url": "", "post_date": "04/24/2024 06:45:18", "content": "Excuse me, I don't participate in Kaggle competitions very often. Would Kaggle competitions be considered as a commercial use? I do not consider the use of GeminiMol for Kaggle to be commercial. The prohibition on the use of this platform for commercial training and paid computing platformis solely intended to prevent speculators from exploiting informational asymmetry for profit.
", "votes": null, "replies": [ { "id": 2771332, "author_name": "hengck23", "author_url": "", "post_date": "04/24/2024 07:40:58", "content": "yes. participation of kaggle is considered as commercial.
\n(e.g. if you use software, dataset etc that is not licensed for commercial use, you will not get the prize money)
but it is better to wait for the organizer and the host to issue an official notice.
\nyou may want to ask the organizer and host for clarification, e.g. by tagging them (with @) in your question.
an famous example is GPL license of yolov5, which causes confusion in the past competition:
\nhttps://www.kaggle.com/c/global-wheat-detection/discussion/163433#919074
it is incredibly confusing. Sometimes the competition itself is deemed as non-commerical, and hence any software can be use. so it is better to ask the organizer and kaggle staffs.
", "votes": null, "replies": [ { "id": 2771409, "author_name": "wanglinboop", "author_url": "", "post_date": "04/24/2024 08:26:45", "content": "OK, I modified GeminiMol's license statement to allow use in open source competitions. Thanks for your comments !
", "votes": null, "replies": [ { "id": 2771586, "author_name": "shlomoron", "author_url": "", "post_date": "04/24/2024 10:15:34", "content": "Good catch @hengck23 . @wanglinboop , changing the license to allow use in a competition is not enough. Since this competition is organized by a for-profit company, the solutions (obviously) have to be such that the company can make use of…this is why section A1 in the rules (winner license) demands our solution to be licensed under MIT up to \"generally commercially available software not owned by you that you used to generate your submission\". Open-source that are not available commercially are no good in most Kaggle competitions.
", "votes": null, "replies": [ { "id": 2771656, "author_name": "hengck23", "author_url": "", "post_date": "04/24/2024 10:50:51", "content": "https://www.kaggle.com/competitions/leash-BELKA/rules
\nWINNER LICENSE TYPE: MIT Open Source
Under Section 11 (Winners Obligations) of the General Rules below, you hereby grant and will grant Competition Sponsor the following license(s) with respect to your Submission if you are a Competition winner:
\nOpen Source: You hereby license and will license your winning Submission and the source code used to generate the Submission under an MIT Open Source license that in no event limits commercial use of such code or model containing or depending on such code. To the extent your Submission makes use of generally commercially available software not owned by you that you used to generate your submission, but that can be procured by the Competition Sponsor without undue expense, you do not grant the license in the preceding sentence to that software.
\nThis is a bit special.
\nNow @wanglinboop is the creator of the software and own the license of GeminiMol.
\nso if wanglinboop wins, he need to grant the license of his solution (which could be a modified version of GeminiMol)
But if other kaggler uses GeminiMol and win, they cannot change the license.
\nI am not a lawyer, so i cannot confirm if the above this correct. in any case, do obtain confirmation from kaggle staffs and host.
", "votes": null, "replies": [ { "id": 2771677, "author_name": "wanglinboop", "author_url": "", "post_date": "04/24/2024 11:04:13", "content": "Thanks. I have made revisions to indicate this point. 🤣
", "votes": null, "replies": [] } ] } ] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "2769155": "Edit: GeminiMol is Academic Free, due to licensing restrictions, the usage of non-commercially available software may result in ineligibility for prize. Thanks for pointing this out. @hengck23 @shlomoron \n\n===============================================\n\n\nUp to this point, I consider the objective of this competition to be a QSAR (quantitative structure-activity relationship) task rather than a DTI (drug-target interaction) task. Therefore, effective small molecular representation, especially those capable of capturing scaffold hopping potential, is of paramount importance. Unsurprisingly, this is also a research area of great interest within the drug discovery community.\n\nIndeed, the combination of multiple molecular fingerprints, such as ECFP4+MACCS+TopoTorsion, along with a powerful machine learning method like LightGBM, is sufficient to achieve a good performance in most QSAR tasks. [In this competition as well](https://www.kaggle.com/competitions/leash-BELKA/discussion/492846). In additation to molecular fingerprints, you may be exploring alternative molecular representation models, which acquire an **understanding (or perhaps just memorization?)** of molecular structures through pre-training tasks. These models aim to enhance the predictive performance or generalization ability of QSAR models. \n\nOver the past year, we have proposed a model called [**GeminiMol**](https://github.com/Wang-Lin-boop/GeminiMol), which achieves impressive performance on various downstream benchmarks of millions of molecules, despite being pre-trained on only 39,290 molecular structures. These benchmarks include zero-shot molecular similarity comparison tasks, target-level QSAR tasks (similar to this competition), cell-level QSAR tasks, and ADMET property prediction tasks. If any participants have the time to explore novel molecular representation methods, I think GeminiMol may be of assistance to you. You can find the scripts, models and datasets from our GitHub repo: https://github.com/Wang-Lin-boop/GeminiMol?tab=readme-ov-file \n\n\n\nGeminiMol employs intermolecular contrastive learning for pre-training. We utilize a graph neural network called WLN to extract molecular features and use these features to predict the 2D maximum common substructure similarity and the **3D conformational space similarity of molecule pairs (represented by shape and pharmacophore similarity)**. This training strategy allows the model to capture the 3D similarity between molecule pairs with significant 2D structural differences, which may aid in capturing molecular scaffold hopping.\n\nRead more here: [Conformational Space Profile Enhances Generic Molecular Representation Learning\n](https://www.biorxiv.org/content/10.1101/2023.12.14.571629)", "2769220": "\"combination of multiple molecular fingerprints, such as ECFP4+MACCS+TopoTorsion,\"\ni don't think this QSAR alone is not good enough for medal at the end of competition at the end of 3 months.\n\ni don't think QSAR task will out-performed DTI.\nFurther, molecular fingerprints may work poorly for unseen test blocks.\n\nNevertheless, QSAR is good as a starting benchmark.", "2769254": "Undeniably, for the drug discovery community, the significance of DTI models surpasses that of QSAR models (at least in terms of research popularity, haha). This why I said \"up to this point\". Cautiously speaking, if there are a sufficient number of molecules in the test set that exhibit binding sites or conformations that differ from those in the training set, it is indeed possible for DTI models to outperform QSAR models. \n\nIn such cases, a sufficiently robust DTI model, capable of surpassing QSAR models on known binding sites, and capable of generalizing to these unknown sites and conformations, would be competitive. However, training such a model would require a larger dataset of protein-ligand pairs, which I believe is quite challenging. In my opinion, it is insufficient to solely utilize the three proteins provided in the training set to train a model that can cover the novel mechanisms of action in the test set.", "2771040": "i find this:\n\n\"GeminiMol is released under the Academic Free Licence, which permits academic use, modification and distribution free of charge, but prohibits unauthorised commercial use, including commercial training and as part of a paid computational platform. \"\n\nhttps://huggingface.co/AlphaMWang/GeminiMol-MOD\n\nHence it think this cannot be used for this kaggle competition", "2771228": "Excuse me, I don't participate in Kaggle competitions very often. Would Kaggle competitions be considered as a commercial use? I do not consider the use of GeminiMol for Kaggle to be commercial. The prohibition on the use of this platform for commercial training and paid computing platformis solely intended to prevent speculators from exploiting informational asymmetry for profit.", "2771332": "yes. participation of kaggle is considered as commercial.\n(e.g. if you use software, dataset etc that is not licensed for commercial use, you will not get the prize money)\n\nbut it is better to wait for the organizer and the host to issue an official notice.\nyou may want to ask the organizer and host for clarification, e.g. by tagging them (with @) in your question.\n\n----\n\nan famous example is GPL license of yolov5, which causes confusion in the past competition:\nhttps://www.kaggle.com/c/global-wheat-detection/discussion/163433#919074\n\n---\n\nit is incredibly confusing. Sometimes the competition itself is deemed as non-commerical, and hence any software can be use. so it is better to ask the organizer and kaggle staffs.", "2771409": "OK, I modified GeminiMol's license statement to allow use in open source competitions. Thanks for your comments !", "2771586": "Good catch @hengck23 . @wanglinboop , changing the license to allow use in a *competition* is not enough. Since this competition is organized by a for-profit company, the solutions (obviously) have to be such that the company can make use of...this is why section A1 in the rules (winner license) demands our solution to be licensed under MIT up to \"generally commercially available software not owned by you that you used to generate your submission\". Open-source that are not available commercially are no good in most Kaggle competitions.", "2771656": "https://www.kaggle.com/competitions/leash-BELKA/rules\nWINNER LICENSE TYPE: MIT Open Source\n\n1. WINNER LICENSE.\n\nUnder Section 11 (Winners Obligations) of the General Rules below, you hereby grant and will grant Competition Sponsor the following license(s) with respect to your Submission if you are a Competition winner:\n\nOpen Source: You hereby license and will license your winning Submission and the source code used to generate the Submission under an MIT Open Source license that in no event limits commercial use of such code or model containing or depending on such code. To the extent your Submission makes use of generally commercially available software not owned by you that you used to generate your submission, but that can be procured by the Competition Sponsor without undue expense, you do not grant the license in the preceding sentence to that software.\n\n----\nThis is a bit special.\nNow @wanglinboop is the creator of the software and own the license of GeminiMol.\nso if wanglinboop wins, he need to grant the license of his solution (which could be a modified version of GeminiMol)\n\nBut if other kaggler uses GeminiMol and win, they cannot change the license.\n\nI am not a lawyer, so i cannot confirm if the above this correct. in any case, do obtain confirmation from kaggle staffs and host.", "2771677": "Thanks. I have made revisions to indicate this point. 🤣" }, "source": "meta" }