I'm excited to share my inference and submission pipeline for the ongoing Kaggle RNA 3D Folding Prediction Challenge! For this competition, I've been using Boltz-1, a powerful and fully open-source biomolecular structure prediction model released by MIT.
\nBoltz-1 is a end-to-end deep learning model for predicting the 3D structures of biomolecular complexes. It includes well-supported training, inference, and evaluation pipelines, making it an excellent open-source choice for both research and competition settings. The code is available on GitHub here: https://github.com/jwohlwend/boltz
\nAccording to the official Boltz-1 paper, the model achieves highly competitive performance compared to AlphaFold3 (AF3) variants. https://www.biorxiv.org/content/10.1101/2024.11.19.624167v3
\nAdditionally, the recent benchmark study, \"Limits of deep-learning-based RNA prediction methods\", evaluated various co-folding structure prediction models — and Boltz came out on top, showing the best overall performance.
\n(It’s worth noting that Proteinix was not tested in that study.)
I've built a full inference and submission pipeline based on Boltz-1.
\nIt covers:
\nInput preprocessing
\nRunning Boltz-1 for 3D RNA prediction
\nFormatting outputs for submission
\nCode is available here: https://www.kaggle.com/code/youhanlee/boltz-1-inference-submission
\nI'm incredibly honored to be a co-host of this competition. RNA structure prediction has been a central interest in my ML journey, and helping RNA-folding competition is a milestone I’m proud of.
\nThis moment is also personally meaningful: I won gold in Kaggle’s first RNA competition, and it made me become a Kaggle Grandmaster. Now, being on the organizing side is exciting.
\nLooking forward to seeing what everyone builds — let’s push the limits of RNA folding together! 🧬
", "messages": [ { "id": 3194705, "postDate": "2025-05-06T07:53:19.597Z", "content": "I'm excited to share my inference and submission pipeline for the ongoing Kaggle RNA 3D Folding Prediction Challenge! For this competition, I've been using Boltz-1, a powerful and fully open-source biomolecular structure prediction model released by MIT.
\nBoltz-1 is a end-to-end deep learning model for predicting the 3D structures of biomolecular complexes. It includes well-supported training, inference, and evaluation pipelines, making it an excellent open-source choice for both research and competition settings. The code is available on GitHub here: https://github.com/jwohlwend/boltz
\nAccording to the official Boltz-1 paper, the model achieves highly competitive performance compared to AlphaFold3 (AF3) variants. https://www.biorxiv.org/content/10.1101/2024.11.19.624167v3
\nAdditionally, the recent benchmark study, \"Limits of deep-learning-based RNA prediction methods\", evaluated various co-folding structure prediction models — and Boltz came out on top, showing the best overall performance.
\n(It’s worth noting that Proteinix was not tested in that study.)
I've built a full inference and submission pipeline based on Boltz-1.
\nIt covers:
\nInput preprocessing
\nRunning Boltz-1 for 3D RNA prediction
\nFormatting outputs for submission
\nCode is available here: https://www.kaggle.com/code/youhanlee/boltz-1-inference-submission
\nI'm incredibly honored to be a co-host of this competition. RNA structure prediction has been a central interest in my ML journey, and helping RNA-folding competition is a milestone I’m proud of.
\nThis moment is also personally meaningful: I won gold in Kaggle’s first RNA competition, and it made me become a Kaggle Grandmaster. Now, being on the organizing side is exciting.
\nLooking forward to seeing what everyone builds — let’s push the limits of RNA folding together! 🧬
", "rawMarkdown": "I'm excited to share my inference and submission pipeline for the ongoing Kaggle RNA 3D Folding Prediction Challenge! For this competition, I've been using Boltz-1, a powerful and fully open-source biomolecular structure prediction model released by MIT.\n\n# Why Boltz-1?\nBoltz-1 is a end-to-end deep learning model for predicting the 3D structures of biomolecular complexes. It includes well-supported training, inference, and evaluation pipelines, making it an excellent open-source choice for both research and competition settings. The code is available on GitHub here: https://github.com/jwohlwend/boltz\n\nAccording to the official Boltz-1 paper, the model achieves highly competitive performance compared to AlphaFold3 (AF3) variants. https://www.biorxiv.org/content/10.1101/2024.11.19.624167v3\n\nAdditionally, the recent benchmark study, \"Limits of deep-learning-based RNA prediction methods\", evaluated various co-folding structure prediction models — and Boltz came out on top, showing the best overall performance.\n(It’s worth noting that Proteinix was not tested in that study.)\n\n# Inference & Submission Code\nI've built a full inference and submission pipeline based on Boltz-1.\n\nIt covers:\n\nInput preprocessing\n\nRunning Boltz-1 for 3D RNA prediction\n\nFormatting outputs for submission\n\nCode is available here: https://www.kaggle.com/code/youhanlee/boltz-1-inference-submission\n\n# A Personal Note\nI'm incredibly honored to be a co-host of this competition. RNA structure prediction has been a central interest in my ML journey, and helping RNA-folding competition is a milestone I’m proud of.\n\nThis moment is also personally meaningful: I won gold in Kaggle’s first RNA competition, and it made me become a Kaggle Grandmaster. Now, being on the organizing side is exciting.\n\nLooking forward to seeing what everyone builds — let’s push the limits of RNA folding together! 🧬", "votes": 21 }, { "id": 3194780, "postDate": "2025-05-06T09:27:58.480Z", "content": "Thank you for sharing your work! It might be a great help to me.
", "rawMarkdown": "Thank you for sharing your work! It might be a great help to me.", "votes": 1, "replies": [ { "id": 3194854, "postDate": "2025-05-06T11:13:04.397Z", "content": "Wow, an impressively high score on LB! Did you fine-tune this model or is it based on your own model architecture?
", "rawMarkdown": "Wow, an impressively high score on LB! Did you fine-tune this model or is it based on your own model architecture?", "votes": 1, "replies": [ { "id": 3194862, "postDate": "2025-05-06T11:18:18.550Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3194864, "postDate": "2025-05-06T11:19:11.363Z", "content": "I haven’t used Boltz yet, but it might be a great method to try. Once I manage to beat the vfold_human_expert score as quickly as I can, I’ll share it. Hopefully, I’ll be able to do it soon.
", "rawMarkdown": "I haven’t used Boltz yet, but it might be a great method to try. Once I manage to beat the vfold_human_expert score as quickly as I can, I’ll share it. Hopefully, I’ll be able to do it soon.", "votes": 2 }, { "id": 3194887, "postDate": "2025-05-06T11:55:12.623Z", "content": "Good Luck!
", "rawMarkdown": "Good Luck!", "votes": 1 } ] }, { "id": 3195226, "postDate": "2025-05-06T19:30:57.787Z", "content": "Thanks! I hope it would be helpful.
", "rawMarkdown": "Thanks! I hope it would be helpful." } ] }, { "id": 3196630, "postDate": "2025-05-07T08:41:03.423Z", "content": "Thank you very much for your sharing; it has been extremely beneficial for a beginner like me.
", "rawMarkdown": "Thank you very much for your sharing; it has been extremely beneficial for a beginner like me." }, { "id": 3195350, "postDate": "2025-05-06T23:29:07.060Z", "content": "it seems the predictions are not stable.
", "rawMarkdown": "it seems the predictions are not stable." }, { "id": 3194945, "postDate": "2025-05-06T13:07:43.033Z", "content": "Thank you for this
", "rawMarkdown": "Thank you for this", "votes": 1 } ], "comments": [ { "id": 3194780, "author_name": "doheon114", "author_url": "", "post_date": "2025-05-06T09:27:58.480000", "content": "Thank you for sharing your work! It might be a great help to me.
", "votes": 1, "replies": [ { "id": 3194854, "author_name": "Nigmat Rahim", "author_url": "", "post_date": "2025-05-06T11:13:04.397000", "content": "Wow, an impressively high score on LB! Did you fine-tune this model or is it based on your own model architecture?
", "votes": 1, "replies": [ { "id": 3194862, "author_name": "", "author_url": "", "post_date": "2025-05-06T11:18:18.550000", "content": "", "votes": 0, "replies": [] }, { "id": 3194864, "author_name": "doheon114", "author_url": "", "post_date": "2025-05-06T11:19:11.363000", "content": "I haven’t used Boltz yet, but it might be a great method to try. Once I manage to beat the vfold_human_expert score as quickly as I can, I’ll share it. Hopefully, I’ll be able to do it soon.
", "votes": 2, "replies": [] }, { "id": 3194887, "author_name": "Nigmat Rahim", "author_url": "", "post_date": "2025-05-06T11:55:12.623000", "content": "Good Luck!
", "votes": 1, "replies": [] } ] }, { "id": 3195226, "author_name": "Youhan Lee", "author_url": "", "post_date": "2025-05-06T19:30:57.787000", "content": "Thanks! I hope it would be helpful.
", "votes": 0, "replies": [] } ] }, { "id": 3196630, "author_name": "Feng", "author_url": "", "post_date": "2025-05-07T08:41:03.423000", "content": "Thank you very much for your sharing; it has been extremely beneficial for a beginner like me.
", "votes": 0, "replies": [] }, { "id": 3195350, "author_name": "MengYe", "author_url": "", "post_date": "2025-05-06T23:29:07.060000", "content": "it seems the predictions are not stable.
", "votes": 0, "replies": [] }, { "id": 3194945, "author_name": "hellodont", "author_url": "", "post_date": "2025-05-06T13:07:43.033000", "content": "Thank you for this
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "3194705": "I'm excited to share my inference and submission pipeline for the ongoing Kaggle RNA 3D Folding Prediction Challenge! For this competition, I've been using Boltz-1, a powerful and fully open-source biomolecular structure prediction model released by MIT.\n\n# Why Boltz-1?\nBoltz-1 is a end-to-end deep learning model for predicting the 3D structures of biomolecular complexes. It includes well-supported training, inference, and evaluation pipelines, making it an excellent open-source choice for both research and competition settings. The code is available on GitHub here: https://github.com/jwohlwend/boltz\n\nAccording to the official Boltz-1 paper, the model achieves highly competitive performance compared to AlphaFold3 (AF3) variants. https://www.biorxiv.org/content/10.1101/2024.11.19.624167v3\n\nAdditionally, the recent benchmark study, \"Limits of deep-learning-based RNA prediction methods\", evaluated various co-folding structure prediction models — and Boltz came out on top, showing the best overall performance.\n(It’s worth noting that Proteinix was not tested in that study.)\n\n# Inference & Submission Code\nI've built a full inference and submission pipeline based on Boltz-1.\n\nIt covers:\n\nInput preprocessing\n\nRunning Boltz-1 for 3D RNA prediction\n\nFormatting outputs for submission\n\nCode is available here: https://www.kaggle.com/code/youhanlee/boltz-1-inference-submission\n\n# A Personal Note\nI'm incredibly honored to be a co-host of this competition. RNA structure prediction has been a central interest in my ML journey, and helping RNA-folding competition is a milestone I’m proud of.\n\nThis moment is also personally meaningful: I won gold in Kaggle’s first RNA competition, and it made me become a Kaggle Grandmaster. Now, being on the organizing side is exciting.\n\nLooking forward to seeing what everyone builds — let’s push the limits of RNA folding together! 🧬", "3194780": "Thank you for sharing your work! It might be a great help to me.", "3196630": "Thank you very much for your sharing; it has been extremely beneficial for a beginner like me.", "3195350": "it seems the predictions are not stable.", "3194945": "Thank you for this" } }