\"DeepGOZero is a novel method which uses model-theoretic approach for learning ontology embeddings and combine it with neural networks for protein function prediction. DeepGOZero can exploit formal axioms in the GO to make zero-shot predictions, i.e., predict protein functions even if not a single protein in the training phase was associated with that function.\"
\nThis a new approach to the protein function prediction problem. It has proven to really good for this competition. I personally have used it with my present model to achieve better results.
\nLet's discuss what views do you have on this approach and what can be some new approaches emerging from this.
This is the repo link:
\nhttps://github.com/bio-ontology-research-group/deepgozero/tree/main
A big kudos to Maxat Kulmanov, Robert Hoehndorf for their work.
", "messages": [ { "id": "2341041", "postDate": "07/11/2023 20:24:04", "content": "\"DeepGOZero is a novel method which uses model-theoretic approach for learning ontology embeddings and combine it with neural networks for protein function prediction. DeepGOZero can exploit formal axioms in the GO to make zero-shot predictions, i.e., predict protein functions even if not a single protein in the training phase was associated with that function.\"
\nThis a new approach to the protein function prediction problem. It has proven to really good for this competition. I personally have used it with my present model to achieve better results.
\nLet's discuss what views do you have on this approach and what can be some new approaches emerging from this.
This is the repo link:
\nhttps://github.com/bio-ontology-research-group/deepgozero/tree/main
A big kudos to Maxat Kulmanov, Robert Hoehndorf for their work.
", "rawMarkdown": "\"DeepGOZero is a novel method which uses model-theoretic approach for learning ontology embeddings and combine it with neural networks for protein function prediction. DeepGOZero can exploit formal axioms in the GO to make zero-shot predictions, i.e., predict protein functions even if not a single protein in the training phase was associated with that function.\" \n\nThis a new approach to the protein function prediction problem. It has proven to really good for this competition. I personally have used it with my present model to achieve better results. \nLet's discuss what views do you have on this approach and what can be some new approaches emerging from this.\n\nThis is the repo link:\nhttps://github.com/bio-ontology-research-group/deepgozero/tree/main\n\nA big kudos to Maxat Kulmanov, Robert Hoehndorf for their work.", "votes": null }, { "id": "2341767", "postDate": "07/12/2023 09:19:42", "content": "If you don't mind it would be nice if you could share insights how to prep the data in a way to be able to use deepgo zero (if you don't mind, of course without giving away top much Details of your own adjustments…)
", "rawMarkdown": "If you don't mind it would be nice if you could share insights how to prep the data in a way to be able to use deepgo zero (if you don't mind, of course without giving away top much Details of your own adjustments...)", "votes": null }, { "id": "2342295", "postDate": "07/12/2023 17:07:24", "content": "You can use the code in the github repo and train a model for this competition. Just a small hint, using only DeepGOZero wouldn't help, ensembling multiple results always looks promising in such scenarios.
", "rawMarkdown": "You can use the code in the github repo and train a model for this competition. Just a small hint, using only DeepGOZero wouldn't help, ensembling multiple results always looks promising in such scenarios.", "votes": null }, { "id": "2342470", "postDate": "07/12/2023 21:18:45", "content": "Hi! Could you share InterPro Embeddings in case you have prepared them for this model?
\nIf not, could you share the embeddings you used for training and testing?
", "rawMarkdown": "Hi! Could you share InterPro Embeddings in case you have prepared them for this model?\n\nIf not, could you share the embeddings you used for training and testing?", "votes": null }, { "id": "2344568", "postDate": "07/14/2023 15:10:44", "content": "I'm fairly new to this, so you ensemble DeepGOZero with a custom NN architecture, then average the predicted probabilities of each GO term ID for each protein?
", "rawMarkdown": "I'm fairly new to this, so you ensemble DeepGOZero with a custom NN architecture, then average the predicted probabilities of each GO term ID for each protein?", "votes": null }, { "id": "2345871", "postDate": "07/15/2023 19:00:08", "content": "Hi! I haven't used InterPro Embeds till now. But I'll surely like to explore more on this one.
", "rawMarkdown": "Hi! I haven't used InterPro Embeds till now. But I'll surely like to explore more on this one.", "votes": null }, { "id": "2345873", "postDate": "07/15/2023 19:01:14", "content": "Yes that can be one way of doing it.
", "rawMarkdown": "Yes that can be one way of doing it.", "votes": null }, { "id": "2350729", "postDate": "07/19/2023 13:22:32", "content": "@siddhvr , hello
\nPlease explain how to get as great plays in LB as your's ?
\nDo you use only T5 embeding?
", "rawMarkdown": "siddhvr , hello\n\nPlease explain how to get as great plays in LB as your's ?\n\nDo you use only T5 embeding?", "votes": null }, { "id": "2350843", "postDate": "07/19/2023 15:07:32", "content": "Hey!!
\nI can't really open it all up, it would be against the competition rules. But to answer your question, no, I don't use only T5 embeds.
Thank you!
", "rawMarkdown": "Thank you!", "votes": null } ], "comments": [ { "id": 2341767, "author_name": "simonveitner", "author_url": "", "post_date": "07/12/2023 09:19:42", "content": "If you don't mind it would be nice if you could share insights how to prep the data in a way to be able to use deepgo zero (if you don't mind, of course without giving away top much Details of your own adjustments…)
", "votes": null, "replies": [ { "id": 2342295, "author_name": "siddhvr", "author_url": "", "post_date": "07/12/2023 17:07:24", "content": "You can use the code in the github repo and train a model for this competition. Just a small hint, using only DeepGOZero wouldn't help, ensembling multiple results always looks promising in such scenarios.
", "votes": null, "replies": [ { "id": 2344568, "author_name": "bigswipp", "author_url": "", "post_date": "07/14/2023 15:10:44", "content": "I'm fairly new to this, so you ensemble DeepGOZero with a custom NN architecture, then average the predicted probabilities of each GO term ID for each protein?
", "votes": null, "replies": [ { "id": 2345873, "author_name": "siddhvr", "author_url": "", "post_date": "07/15/2023 19:01:14", "content": "Yes that can be one way of doing it.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2342470, "author_name": "aschamp", "author_url": "", "post_date": "07/12/2023 21:18:45", "content": "Hi! Could you share InterPro Embeddings in case you have prepared them for this model?
\nIf not, could you share the embeddings you used for training and testing?
", "votes": null, "replies": [ { "id": 2345871, "author_name": "siddhvr", "author_url": "", "post_date": "07/15/2023 19:00:08", "content": "Hi! I haven't used InterPro Embeds till now. But I'll surely like to explore more on this one.
", "votes": null, "replies": [] } ] }, { "id": 2350729, "author_name": "", "author_url": "", "post_date": "07/19/2023 13:22:32", "content": "@siddhvr , hello
\nPlease explain how to get as great plays in LB as your's ?
\nDo you use only T5 embeding?
", "votes": null, "replies": [ { "id": 2350843, "author_name": "siddhvr", "author_url": "", "post_date": "07/19/2023 15:07:32", "content": "Hey!!
\nI can't really open it all up, it would be against the competition rules. But to answer your question, no, I don't use only T5 embeds.
Thank you!
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2341041": "\"DeepGOZero is a novel method which uses model-theoretic approach for learning ontology embeddings and combine it with neural networks for protein function prediction. DeepGOZero can exploit formal axioms in the GO to make zero-shot predictions, i.e., predict protein functions even if not a single protein in the training phase was associated with that function.\" \n\nThis a new approach to the protein function prediction problem. It has proven to really good for this competition. I personally have used it with my present model to achieve better results. \nLet's discuss what views do you have on this approach and what can be some new approaches emerging from this.\n\nThis is the repo link:\nhttps://github.com/bio-ontology-research-group/deepgozero/tree/main\n\nA big kudos to Maxat Kulmanov, Robert Hoehndorf for their work.", "2341767": "If you don't mind it would be nice if you could share insights how to prep the data in a way to be able to use deepgo zero (if you don't mind, of course without giving away top much Details of your own adjustments...)", "2342295": "You can use the code in the github repo and train a model for this competition. Just a small hint, using only DeepGOZero wouldn't help, ensembling multiple results always looks promising in such scenarios.", "2342470": "Hi! Could you share InterPro Embeddings in case you have prepared them for this model?\n\nIf not, could you share the embeddings you used for training and testing?", "2344568": "I'm fairly new to this, so you ensemble DeepGOZero with a custom NN architecture, then average the predicted probabilities of each GO term ID for each protein?", "2345871": "Hi! I haven't used InterPro Embeds till now. But I'll surely like to explore more on this one.", "2345873": "Yes that can be one way of doing it.", "2350729": "siddhvr , hello\n\nPlease explain how to get as great plays in LB as your's ?\n\nDo you use only T5 embeding?", "2350843": "Hey!!\nI can't really open it all up, it would be against the competition rules. But to answer your question, no, I don't use only T5 embeds.", "2350884": "Thank you!" }, "source": "meta" }