Hey guys, I am new to protein function prediction and would like some guidance on how to construct features optimally for models like n-gram or HMM to get the best embeddings for the protein sequences. I am still not comfortable with deep learning so i am sticking to older approaches. Any guidance is highly appreciated.
\nThanks!
", "messages": [ { "id": "2254867", "postDate": "05/11/2023 10:01:26", "content": "Hey guys, I am new to protein function prediction and would like some guidance on how to construct features optimally for models like n-gram or HMM to get the best embeddings for the protein sequences. I am still not comfortable with deep learning so i am sticking to older approaches. Any guidance is highly appreciated.
\nThanks!
", "rawMarkdown": "Hey guys, I am new to protein function prediction and would like some guidance on how to construct features optimally for models like n-gram or HMM to get the best embeddings for the protein sequences. I am still not comfortable with deep learning so i am sticking to older approaches. Any guidance is highly appreciated.\n\nThanks!", "votes": null }, { "id": "2254878", "postDate": "05/11/2023 10:15:02", "content": "You could use transformer protein language model embeddings as features, as I demonstrated in this notebook. In the above example, ESM-2 embeddings are used. Sergei Fironov has also provided ProtT5 embeddings in this dataset.
", "rawMarkdown": "You could use transformer protein language model embeddings as features, as I demonstrated in [this](https://www.kaggle.com/code/viktorfairuschin/esm-2-embeddings-starter) notebook. In the above example, [ESM-2](https://github.com/facebookresearch/esm) embeddings are used. Sergei Fironov has also provided ProtT5 embeddings in [this](https://www.kaggle.com/datasets/sergeifironov/t5embeds) dataset.", "votes": null }, { "id": "2255556", "postDate": "05/11/2023 19:45:42", "content": "Like @viktorfairuschin said, embeddings from language models work well. If you want to read more about it or try an implementation of that idea by people who do this for a living, check out the links below.
\n", "rawMarkdown": "Like @viktorfairuschin said, embeddings from language models work well. If you want to read more about it or try an implementation of that idea by people who do this for a living, check out the links below.\n\n- https://www.nature.com/articles/s41598-020-80786-0\n- https://github.com/Rostlab/goPredSim", "votes": null }, { "id": "2261826", "postDate": "05/16/2023 15:12:28", "content": "Hello 👋
\nI am going to give my take on non-deep learning approaches as you specifically asked for those. An example would be applying some TF-IDF (term frequency, inverse document frequency) techniques to amino-acids sequences.
In short:
\nAfter obtaining these two factors for all amino-acids, you can multiply them and get a sort of embedding for an aminoacid sequence. This reasoning can be applied for sequences of amino-acids potentially and can get you some representation for the whole sequences.
\nIt is a baseline but definitely a start (with its strong limitations).
", "rawMarkdown": "Hello 👋\nI am going to give my take on non-deep learning approaches as you specifically asked for those. An example would be applying some TF-IDF (term frequency, inverse document frequency) techniques to amino-acids sequences. \n\nIn short: \n- TF will enable you to understand which amino-acids are appearing the most in the dataset\n- IDF will enable you to give more importance to the ones that appear the least (the rare ones basically)\n\nAfter obtaining these two factors for all amino-acids, you can multiply them and get a sort of embedding for an aminoacid sequence. This reasoning can be applied for sequences of amino-acids potentially and can get you some representation for the whole sequences. \n\nIt is a baseline but definitely a start (with its strong limitations).", "votes": null } ], "comments": [ { "id": 2254878, "author_name": "viktorfairuschin", "author_url": "", "post_date": "05/11/2023 10:15:02", "content": "You could use transformer protein language model embeddings as features, as I demonstrated in this notebook. In the above example, ESM-2 embeddings are used. Sergei Fironov has also provided ProtT5 embeddings in this dataset.
", "votes": null, "replies": [] }, { "id": 2255556, "author_name": "tilii7", "author_url": "", "post_date": "05/11/2023 19:45:42", "content": "Like @viktorfairuschin said, embeddings from language models work well. If you want to read more about it or try an implementation of that idea by people who do this for a living, check out the links below.
\n", "votes": null, "replies": [] }, { "id": 2261826, "author_name": "wadgek", "author_url": "", "post_date": "05/16/2023 15:12:28", "content": "Hello 👋
\nI am going to give my take on non-deep learning approaches as you specifically asked for those. An example would be applying some TF-IDF (term frequency, inverse document frequency) techniques to amino-acids sequences.
In short:
\nAfter obtaining these two factors for all amino-acids, you can multiply them and get a sort of embedding for an aminoacid sequence. This reasoning can be applied for sequences of amino-acids potentially and can get you some representation for the whole sequences.
\nIt is a baseline but definitely a start (with its strong limitations).
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2254867": "Hey guys, I am new to protein function prediction and would like some guidance on how to construct features optimally for models like n-gram or HMM to get the best embeddings for the protein sequences. I am still not comfortable with deep learning so i am sticking to older approaches. Any guidance is highly appreciated.\n\nThanks!", "2254878": "You could use transformer protein language model embeddings as features, as I demonstrated in [this](https://www.kaggle.com/code/viktorfairuschin/esm-2-embeddings-starter) notebook. In the above example, [ESM-2](https://github.com/facebookresearch/esm) embeddings are used. Sergei Fironov has also provided ProtT5 embeddings in [this](https://www.kaggle.com/datasets/sergeifironov/t5embeds) dataset.", "2255556": "Like @viktorfairuschin said, embeddings from language models work well. If you want to read more about it or try an implementation of that idea by people who do this for a living, check out the links below.\n\n- https://www.nature.com/articles/s41598-020-80786-0\n- https://github.com/Rostlab/goPredSim", "2261826": "Hello 👋\nI am going to give my take on non-deep learning approaches as you specifically asked for those. An example would be applying some TF-IDF (term frequency, inverse document frequency) techniques to amino-acids sequences. \n\nIn short: \n- TF will enable you to understand which amino-acids are appearing the most in the dataset\n- IDF will enable you to give more importance to the ones that appear the least (the rare ones basically)\n\nAfter obtaining these two factors for all amino-acids, you can multiply them and get a sort of embedding for an aminoacid sequence. This reasoning can be applied for sequences of amino-acids potentially and can get you some representation for the whole sequences. \n\nIt is a baseline but definitely a start (with its strong limitations)." }, "source": "meta" }