Hi all,\nI've created a dataset containing the ESM-Cambrian embeddings for Train and Test.
\nESM-Cambrian is the successor of ESM2: a foundation model for protein sequences.
\nDataset is available here: https://www.kaggle.com/datasets/dalloliogm/cafa6-esm-cambrian-embeddings/data/data/data
\nNotebook to generate these is here: https://www.kaggle.com/code/dalloliogm/compute-protein-embeddings-with-esm2-esm-c/notebook
\nI thought it would be good to share these, so people don't have to compute them all over again. It takes about 2 hours of compute to convert the sequences into embeddings. I know there are some ESM2 embeddings that have already been shared, but I could not see ESM-C.
\nMore about ESM-C: https://huggingface.co/EvolutionaryScale/esmc-600m-2024-12
", "messages": [ { "id": 3357504, "postDate": "2025-12-01T09:31:22.733Z", "content": "Hi all,\nI've created a dataset containing the ESM-Cambrian embeddings for Train and Test.
\nESM-Cambrian is the successor of ESM2: a foundation model for protein sequences.
\nDataset is available here: https://www.kaggle.com/datasets/dalloliogm/cafa6-esm-cambrian-embeddings/data/data/data
\nNotebook to generate these is here: https://www.kaggle.com/code/dalloliogm/compute-protein-embeddings-with-esm2-esm-c/notebook
\nI thought it would be good to share these, so people don't have to compute them all over again. It takes about 2 hours of compute to convert the sequences into embeddings. I know there are some ESM2 embeddings that have already been shared, but I could not see ESM-C.
\nMore about ESM-C: https://huggingface.co/EvolutionaryScale/esmc-600m-2024-12
", "rawMarkdown": "Hi all,\nI've created a dataset containing the ESM-Cambrian embeddings for Train and Test.\n\nESM-Cambrian is the successor of ESM2: a foundation model for protein sequences. \n\nDataset is available here: https://www.kaggle.com/datasets/dalloliogm/cafa6-esm-cambrian-embeddings/data/data/data\n\nNotebook to generate these is here: https://www.kaggle.com/code/dalloliogm/compute-protein-embeddings-with-esm2-esm-c/notebook\n\n\nI thought it would be good to share these, so people don't have to compute them all over again. It takes about 2 hours of compute to convert the sequences into embeddings. I know there are some ESM2 embeddings that have already been shared, but I could not see ESM-C.\n\nMore about ESM-C: https://huggingface.co/EvolutionaryScale/esmc-600m-2024-12", "votes": 23 }, { "id": 3360652, "postDate": "2025-12-02T12:21:48.860Z", "content": "Thank you! Have you had any impressions of the embedding performance yet? \nFor the moment, my local validation says T5 is better then ESM2.
", "rawMarkdown": "Thank you! Have you had any impressions of the embedding performance yet? \nFor the moment, my local validation says T5 is better then ESM2.", "votes": 1, "replies": [ { "id": 3368958, "postDate": "2025-12-09T16:31:58.003Z", "content": "which size of the esm, the esm is not good in my test
", "rawMarkdown": "which size of the esm, the esm is not good in my test", "replies": [ { "id": 3369768, "postDate": "2025-12-10T10:21:53.217Z", "content": "Thanks for the feedback!\nShapes are Embeddings shapes -> Train: (82404, 320), Test: (224309, 320).
I've just realized that I have used the wrong model for the notebook - esm3_sm_open_v0 which is ESM3, not ESM-Cambrian. I think that for this scope it should not make a big difference, but I'll try to run it with ESM-C. The problem is that ESM-C requires accepting some terms and conditions, and I am not sure I can share the output in a public notebook. You can still copy the notebook and run it with any model you want.
got it, I just said the effect of different embedding, the esm2-15b or 3b is not better than prott5
", "rawMarkdown": "got it, I just said the effect of different embedding, the esm2-15b or 3b is not better than prott5" }, { "id": 3370087, "postDate": "2025-12-10T15:08:31.367Z", "content": "Thanks, that's a useful insight. It would be interesting to do a thorough comparison. What are you trying, out of curiosity?
\nMy approach is the following:
\nyour approach is wonderful. For my supervisor and a gold medal, haha
", "rawMarkdown": "your approach is wonderful. For my supervisor and a gold medal, haha" }, { "id": 3370131, "postDate": "2025-12-10T15:36:02.093Z", "content": "Thanks, but I still need to implement half of it!
", "rawMarkdown": "Thanks, but I still need to implement half of it!" } ] } ] }, { "id": 3362051, "postDate": "2025-12-04T05:52:38.563Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 3360652, "author_name": "MT", "author_url": "", "post_date": "2025-12-02T12:21:48.860000", "content": "Thank you! Have you had any impressions of the embedding performance yet? \nFor the moment, my local validation says T5 is better then ESM2.
", "votes": 1, "replies": [ { "id": 3368958, "author_name": "common-people", "author_url": "", "post_date": "2025-12-09T16:31:58.003000", "content": "which size of the esm, the esm is not good in my test
", "votes": 0, "replies": [ { "id": 3369768, "author_name": "Giovanni Marco Dall'Olio", "author_url": "", "post_date": "2025-12-10T10:21:53.217000", "content": "Thanks for the feedback!\nShapes are Embeddings shapes -> Train: (82404, 320), Test: (224309, 320).
I've just realized that I have used the wrong model for the notebook - esm3_sm_open_v0 which is ESM3, not ESM-Cambrian. I think that for this scope it should not make a big difference, but I'll try to run it with ESM-C. The problem is that ESM-C requires accepting some terms and conditions, and I am not sure I can share the output in a public notebook. You can still copy the notebook and run it with any model you want.
got it, I just said the effect of different embedding, the esm2-15b or 3b is not better than prott5
", "votes": 0, "replies": [] }, { "id": 3370087, "author_name": "Giovanni Marco Dall'Olio", "author_url": "", "post_date": "2025-12-10T15:08:31.367000", "content": "Thanks, that's a useful insight. It would be interesting to do a thorough comparison. What are you trying, out of curiosity?
\nMy approach is the following:
\nyour approach is wonderful. For my supervisor and a gold medal, haha
", "votes": 0, "replies": [] }, { "id": 3370131, "author_name": "Giovanni Marco Dall'Olio", "author_url": "", "post_date": "2025-12-10T15:36:02.093000", "content": "Thanks, but I still need to implement half of it!
", "votes": 0, "replies": [] } ] } ] }, { "id": 3362051, "author_name": "", "author_url": "", "post_date": "2025-12-04T05:52:38.563000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "3357504": "Hi all,\nI've created a dataset containing the ESM-Cambrian embeddings for Train and Test.\n\nESM-Cambrian is the successor of ESM2: a foundation model for protein sequences. \n\nDataset is available here: https://www.kaggle.com/datasets/dalloliogm/cafa6-esm-cambrian-embeddings/data/data/data\n\nNotebook to generate these is here: https://www.kaggle.com/code/dalloliogm/compute-protein-embeddings-with-esm2-esm-c/notebook\n\n\nI thought it would be good to share these, so people don't have to compute them all over again. It takes about 2 hours of compute to convert the sequences into embeddings. I know there are some ESM2 embeddings that have already been shared, but I could not see ESM-C.\n\nMore about ESM-C: https://huggingface.co/EvolutionaryScale/esmc-600m-2024-12", "3360652": "Thank you! Have you had any impressions of the embedding performance yet? \nFor the moment, my local validation says T5 is better then ESM2.", "3362051": "" } }