\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4172517%2F2c1f400a7a6b42ef0e3fb8864ab7b978%2Fphoto1688681456.jpeg?generation=1688851313762452&alt=media\")
đź“ą Video record of the webinar: https://youtu.be/hpYfLLqyo9Y
\nThe webinar starts with short overview by Liza Geraseva of the methods used on previous CAFA challenges.
\n(See also presentation attached to the announcement: https://www.kaggle.com/competitions/cafa-5-protein-function-prediction/discussion/418295 )
In a discussion part of the webinar - one can find comments by Robert Hoehndorf (current top1),
\nwho suggests to use their recent version DeepGOZero - https://academic.oup.com/bioinformatics/article/38/Supplement_1/i238/6617515
\nwhich is based on a beautiful idea: GO-hierarchy is embedded by itself preserving the hierarchy structure,
\nand then embeddings of the proteins are projected on the space of GO embedding.
\nThat is the idea.
\n(See discussion around timestamp: https://youtu.be/hpYfLLqyo9Y?t=1437 and figure 2 in the paper above).
Predrag Radivojac (from the organizers team) shared his vision on the difficulties and broad perspective of the competition. A part of the difficulties come from the fact even small change in the sequence might change function a lot. It has also been mentioned that Molecular Function terms are expected to be better captured by sequence data, while Biological Process data might be better captured from the other external sources - like protein-protein interactions datasets, pathways etc. At the same time Cellular Component is something in the middel and widely research open question what should capture it better. It was also remarked for CCO difference between profound methods and baseline (blast-like methods) - is the smallest one - so it might be the point for the improvment for profound methods.
\nHere is also zoom \"chat\" comments:
\n18:09:55 From Quancheng Liu : blast-knn can do better, as far as I know, like around 0.49
\n18:13:26 From Robert Hoehndorf : There is also a neuro-symbolic version of DeepGO for 0-shot learning, but also for exploiting the Gene Ontology axioms much better than DeepGO or DeepGOPlus do: https://academic.oup.com/bioinformatics/article/38/Supplement_1/i238/6617515
\nIt's a much improved version of the hierarchical classifier used in the original DeepGO.
18:24:32 From Robert Hoehndorf : I believe the results are not official for CAFA4 yet :)
\n18:24:47 From Predrag Radivojac : True
18:37:39 From Robert Hoehndorf : This is the very big question!
\n18:38:31 From Predrag Radivojac : We do not know how different the test set would be from the proteins used for the public leaderboard. We do know it will be larger.
\n18:39:57 From Predrag Radivojac : I can give a few thoughts on the metrics
\n18:46:52 From Robert Hoehndorf : https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-9-S5-S2 for curation methods!
\n18:47:07 From Robert Hoehndorf : classic paper :)
\n18:47:37 From Denis Moiseenko : Thank you!
\n18:47:56 From Predrag Radivojac : https://link.springer.com/book/10.1007/978-1-4939-3743-1
\n18:50:51 From DANA MARY : In case of multifunctional proteins how do we ho about with the function prediction
\n18:51:15 From DANA MARY : How do we label them
\n18:51:37 From Simon : They can have multiple labels =1 at same time.
\n18:51:41 From Zc : Some proteins can participate in many different pathways! They would be associated with many different GO labels as a result
\n18:52:20 From Zc : Proteins also take on different roles with different PTMs, localization, microenvironemnt etc.
\n18:52:55 From Zc : Go terms are a very simple way to label these functions
\n18:52:56 From Robert Hoehndorf : also: for these \"hard\" cases Pedrag mentions, a little known fact: there are /real/ negative annotations available (not many, but quite a few still) at https://www.ebi.ac.uk/GOA/downloads (look for the \"NOT\" annotations)
\n18:53:04 From DANA MARY : Do we prefer molecular function ontology over other ontology classes for function prediction
\n18:56:51 From Denis Moiseenko : Thank you again!
\n18:57:16 From DANA MARY : Thank you for the clarification
\n18:57:28 From Simon : Thank you to experienced people for explaining for biology noobs :)
\n18:57:51 From Denis Moiseenko : Sorry
\n18:57:51 From Denis Moiseenko : I asked about uncharted proteins and isoforms
\n18:57:51 From Denis Moiseenko : uncurated
\n18:57:55 From Denis Moiseenko : (Uniparc)
\n18:58:08 From Max Buckley : What level of data representation would you suggest for CCO? If MFO is sequence and BPO is systems level (protein protein interaction etc.)
\n19:00:55 From Zc : Historically BPO has been harder to predict than CCO and MFO, is this still true today for our models?
\n19:01:19 From Robert Hoehndorf : Yes.
\n19:01:28 From Robert Hoehndorf : (if you use Fmax)
\n19:02:57 From Predrag Radivojac : Yes
\n19:03:06 From Robert Hoehndorf : I will have to leave. Thank you very much for organizing this meeting.
\n19:09:22 From DANA MARY : No no more questions. Thank you
\n19:12:22 From Predrag Radivojac : Lucy Colwell’s group has new deep learning methods. I would check her group’s web site.
\n19:13:00 From Predrag Radivojac : https://research.google/people/106569/
\n19:14:26 From Zc : Proteinfer by Cowell:
https://elifesciences.org/articles/80942
\n19:14:33 From Zc : Colwell*
\n19:14:43 From Denis Moiseenko : Thank you for organizing this seminar, thanks Liza for presentation and Predrag for answering questions right from the heart of the challenge!
\n19:15:17 From Simon : Thank you for the many thought provoking ideas and for organizing!
\n19:15:29 From Predrag Radivojac : Thank you all for organizing this webinar and participating.
\n19:16:04 From Quancheng Liu : Thank you guys!
\n19:18:25 From Michał Szafarczyk : Thank you, it was a very instructive meeting!
đź“ą Video record of the webinar: https://youtu.be/hpYfLLqyo9Y
\nThe webinar starts with short overview by Liza Geraseva of the methods used on previous CAFA challenges.
\n(See also presentation attached to the announcement: https://www.kaggle.com/competitions/cafa-5-protein-function-prediction/discussion/418295 )
In a discussion part of the webinar - one can find comments by Robert Hoehndorf (current top1),
\nwho suggests to use their recent version DeepGOZero - https://academic.oup.com/bioinformatics/article/38/Supplement_1/i238/6617515
\nwhich is based on a beautiful idea: GO-hierarchy is embedded by itself preserving the hierarchy structure,
\nand then embeddings of the proteins are projected on the space of GO embedding.
\nThat is the idea.
\n(See discussion around timestamp: https://youtu.be/hpYfLLqyo9Y?t=1437 and figure 2 in the paper above).
Predrag Radivojac (from the organizers team) shared his vision on the difficulties and broad perspective of the competition. A part of the difficulties come from the fact even small change in the sequence might change function a lot. It has also been mentioned that Molecular Function terms are expected to be better captured by sequence data, while Biological Process data might be better captured from the other external sources - like protein-protein interactions datasets, pathways etc. At the same time Cellular Component is something in the middel and widely research open question what should capture it better. It was also remarked for CCO difference between profound methods and baseline (blast-like methods) - is the smallest one - so it might be the point for the improvment for profound methods.
\nHere is also zoom \"chat\" comments:
\n18:09:55 From Quancheng Liu : blast-knn can do better, as far as I know, like around 0.49
\n18:13:26 From Robert Hoehndorf : There is also a neuro-symbolic version of DeepGO for 0-shot learning, but also for exploiting the Gene Ontology axioms much better than DeepGO or DeepGOPlus do: https://academic.oup.com/bioinformatics/article/38/Supplement_1/i238/6617515
\nIt's a much improved version of the hierarchical classifier used in the original DeepGO.
18:24:32 From Robert Hoehndorf : I believe the results are not official for CAFA4 yet :)
\n18:24:47 From Predrag Radivojac : True
18:37:39 From Robert Hoehndorf : This is the very big question!
\n18:38:31 From Predrag Radivojac : We do not know how different the test set would be from the proteins used for the public leaderboard. We do know it will be larger.
\n18:39:57 From Predrag Radivojac : I can give a few thoughts on the metrics
\n18:46:52 From Robert Hoehndorf : https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-9-S5-S2 for curation methods!
\n18:47:07 From Robert Hoehndorf : classic paper :)
\n18:47:37 From Denis Moiseenko : Thank you!
\n18:47:56 From Predrag Radivojac : https://link.springer.com/book/10.1007/978-1-4939-3743-1
\n18:50:51 From DANA MARY : In case of multifunctional proteins how do we ho about with the function prediction
\n18:51:15 From DANA MARY : How do we label them
\n18:51:37 From Simon : They can have multiple labels =1 at same time.
\n18:51:41 From Zc : Some proteins can participate in many different pathways! They would be associated with many different GO labels as a result
\n18:52:20 From Zc : Proteins also take on different roles with different PTMs, localization, microenvironemnt etc.
\n18:52:55 From Zc : Go terms are a very simple way to label these functions
\n18:52:56 From Robert Hoehndorf : also: for these \"hard\" cases Pedrag mentions, a little known fact: there are /real/ negative annotations available (not many, but quite a few still) at https://www.ebi.ac.uk/GOA/downloads (look for the \"NOT\" annotations)
\n18:53:04 From DANA MARY : Do we prefer molecular function ontology over other ontology classes for function prediction
\n18:56:51 From Denis Moiseenko : Thank you again!
\n18:57:16 From DANA MARY : Thank you for the clarification
\n18:57:28 From Simon : Thank you to experienced people for explaining for biology noobs :)
\n18:57:51 From Denis Moiseenko : Sorry
\n18:57:51 From Denis Moiseenko : I asked about uncharted proteins and isoforms
\n18:57:51 From Denis Moiseenko : uncurated
\n18:57:55 From Denis Moiseenko : (Uniparc)
\n18:58:08 From Max Buckley : What level of data representation would you suggest for CCO? If MFO is sequence and BPO is systems level (protein protein interaction etc.)
\n19:00:55 From Zc : Historically BPO has been harder to predict than CCO and MFO, is this still true today for our models?
\n19:01:19 From Robert Hoehndorf : Yes.
\n19:01:28 From Robert Hoehndorf : (if you use Fmax)
\n19:02:57 From Predrag Radivojac : Yes
\n19:03:06 From Robert Hoehndorf : I will have to leave. Thank you very much for organizing this meeting.
\n19:09:22 From DANA MARY : No no more questions. Thank you
\n19:12:22 From Predrag Radivojac : Lucy Colwell’s group has new deep learning methods. I would check her group’s web site.
\n19:13:00 From Predrag Radivojac : https://research.google/people/106569/
\n19:14:26 From Zc : Proteinfer by Cowell:
https://elifesciences.org/articles/80942
\n19:14:33 From Zc : Colwell*
\n19:14:43 From Denis Moiseenko : Thank you for organizing this seminar, thanks Liza for presentation and Predrag for answering questions right from the heart of the challenge!
\n19:15:17 From Simon : Thank you for the many thought provoking ideas and for organizing!
\n19:15:29 From Predrag Radivojac : Thank you all for organizing this webinar and participating.
\n19:16:04 From Quancheng Liu : Thank you guys!
\n19:18:25 From Michał Szafarczyk : Thank you, it was a very instructive meeting!
Excellent summary, thanks for sharing
", "rawMarkdown": "Excellent summary, thanks for sharing", "votes": null }, { "id": "2335852", "postDate": "07/08/2023 21:22:24", "content": "“ for these \"hard\" cases Pedrag mentions, a little known fact: there are /real/ negative annotations available (not many, but quite a few still) at https://www.ebi.ac.uk/GOA/downloads (look for the \"NOT\" annotations)”
\n
Excellent summary, thanks for sharing
", "votes": null, "replies": [] }, { "id": 2335852, "author_name": "", "author_url": "", "post_date": "07/08/2023 21:22:24", "content": "“ for these \"hard\" cases Pedrag mentions, a little known fact: there are /real/ negative annotations available (not many, but quite a few still) at https://www.ebi.ac.uk/GOA/downloads (look for the \"NOT\" annotations)”
\n