download alphafold2 supplimentary paper. it has algorithm for all modules like this
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F9063a612d627bd6fa56a33dba089b83a%2FSelection_167.png?generation=1743239753827033&alt=media\")
download drfold2. it implements the above and follows very close to the step and index
drfold2 uses RAN LM to replace MSA
I noticed that no one has mentioned this library yet, so I wanted to share it here.
\nThere’s a library called OpenComplex, which extends OpenFold to support structure prediction for RNA and RNA-protein complexes.
\nhttps://github.com/baaihealth/OpenComplex/tree/master
\nRight now, I'm exploring the idea of adapting ESMFold —which predicts protein structures from protein language model embeddings— for RNA. I'm experimenting with combining an appropriate RNA language model with ESMFold and OpenComplex.
\nhttps://github.com/facebookresearch/esm
\nThat said, these models are a bit outdated, and setting up the environment and resolving dependencies has been a bit tricky… I might share more if I make interesting progress.
", "messages": [ { "id": "3162419", "postDate": "03/29/2025 07:51:07", "content": "I noticed that no one has mentioned this library yet, so I wanted to share it here.
\nThere’s a library called OpenComplex, which extends OpenFold to support structure prediction for RNA and RNA-protein complexes.
\nhttps://github.com/baaihealth/OpenComplex/tree/master
\nRight now, I'm exploring the idea of adapting ESMFold —which predicts protein structures from protein language model embeddings— for RNA. I'm experimenting with combining an appropriate RNA language model with ESMFold and OpenComplex.
\nhttps://github.com/facebookresearch/esm
\nThat said, these models are a bit outdated, and setting up the environment and resolving dependencies has been a bit tricky… I might share more if I make interesting progress.
", "rawMarkdown": "I noticed that no one has mentioned this library yet, so I wanted to share it here.\n\nThere’s a library called **OpenComplex**, which extends OpenFold to support structure prediction for RNA and RNA-protein complexes.\n\nhttps://github.com/baaihealth/OpenComplex/tree/master\n\nRight now, I'm exploring the idea of adapting ESMFold —which predicts protein structures from protein language model embeddings— for RNA. I'm experimenting with combining an appropriate RNA language model with ESMFold and OpenComplex.\n\nhttps://github.com/facebookresearch/esm\n\nThat said, these models are a bit outdated, and setting up the environment and resolving dependencies has been a bit tricky... I might share more if I make interesting progress.", "votes": null }, { "id": "3162462", "postDate": "03/29/2025 08:38:52", "content": "https://www.biorxiv.org/content/10.1101/2025.03.25.643589v1
\nthe paper and some results
fyi,
\nseveral af3 clones (proteinx, chai-1) already included esm in their to replace MSA for RNA prediction.
\ncheck their repo and technical report
Ah, you're right—I'd forgotten that Chai-1 has a no-MSA option. I'll take a look!
\nIn any case, I'm working on building a simple structure prediction model from scratch as a way to deepen my understanding of this competition :)
", "rawMarkdown": "Ah, you're right—I'd forgotten that Chai-1 has a no-MSA option. I'll take a look!\n\nIn any case, I'm working on building a simple structure prediction model from scratch as a way to deepen my understanding of this competition :)", "votes": null }, { "id": "3162478", "postDate": "03/29/2025 09:08:32", "content": "yet another paper:
\nNeuralPLexer3: Accurate Biomolecular Complex Structure Prediction with Flow Models
\nhttps://arxiv.org/pdf/2412.10743v2
qoute:
\n\"• Nucleic acids: On CASP15 RNA targets [32, 33], NP3 demonstrated comparable performance to AF3 (46.5%
\nversus 47.3%, respectively) and slightly below that of Alchemy_RNA2 34, which relies on explicitly
\ncurated human inputs. While AF3 conditions its predictions on RNA MSAs, NP3 uses RNA language
\nmodels (LMs) and achieves similar performance—a noteworthy result, as MSAs are typically considered
\nmore effective than LMs for conditioning structure predictions. This outcome also highlights the potential of
\nutilizing LM-based approaches in scenarios where RNA MSAs are unavailable or impractical to generate\"
however, i believe msa and esm should both in the total solution. you can can make 5 prediction, why don't use multiple methods?
", "rawMarkdown": "yet another paper:\nNeuralPLexer3: Accurate Biomolecular Complex Structure Prediction with Flow Models\nhttps://arxiv.org/pdf/2412.10743v2\n\nqoute:\n\"• Nucleic acids: On CASP15 RNA targets [32, 33], NP3 demonstrated comparable performance to AF3 (46.5%\nversus 47.3%, respectively) and slightly below that of Alchemy_RNA2 [34] (54.8%), which relies on explicitly\ncurated human inputs. While AF3 conditions its predictions on RNA MSAs, NP3 uses RNA language\nmodels (LMs) and achieves similar performance—a noteworthy result, as MSAs are typically considered\nmore effective than LMs for conditioning structure predictions. This outcome also highlights the potential of\nutilizing LM-based approaches in scenarios where RNA MSAs are unavailable or impractical to generate\"\n\n----\n\nhowever, i believe msa and esm should both in the total solution. you can can make 5 prediction, why don't use multiple methods?", "votes": null }, { "id": "3162485", "postDate": "03/29/2025 09:14:55", "content": "\":In any case, I'm working on building a simple structure prediction model from scratch as a way to deepen my understanding of this competition :)\"
\nmy suggestion:
\ndownload alphafold2 supplimentary paper. it has algorithm for all modules like this
\n
download drfold2. it implements the above and follows very close to the step and index
drfold2 uses RAN LM to replace MSA
you can do likewise for alphafold3
", "rawMarkdown": "\":In any case, I'm working on building a simple structure prediction model from scratch as a way to deepen my understanding of this competition :)\"\n\nmy suggestion:\n- download alphafold2 supplimentary paper. it has algorithm for all modules like this\n\n\n\n- download drfold2. it implements the above and follows very close to the step and index\n- drfold2 uses RAN LM to replace MSA\n\n---\n\nyou can do likewise for alphafold3", "votes": null } ], "comments": [ { "id": 3162462, "author_name": "hengck23", "author_url": "", "post_date": "03/29/2025 08:38:52", "content": "https://www.biorxiv.org/content/10.1101/2025.03.25.643589v1
\nthe paper and some results
fyi,
\nseveral af3 clones (proteinx, chai-1) already included esm in their to replace MSA for RNA prediction.
\ncheck their repo and technical report
Ah, you're right—I'd forgotten that Chai-1 has a no-MSA option. I'll take a look!
\nIn any case, I'm working on building a simple structure prediction model from scratch as a way to deepen my understanding of this competition :)
", "votes": null, "replies": [ { "id": 3162485, "author_name": "hengck23", "author_url": "", "post_date": "03/29/2025 09:14:55", "content": "\":In any case, I'm working on building a simple structure prediction model from scratch as a way to deepen my understanding of this competition :)\"
\nmy suggestion:
\ndownload alphafold2 supplimentary paper. it has algorithm for all modules like this
\n
download drfold2. it implements the above and follows very close to the step and index
drfold2 uses RAN LM to replace MSA
you can do likewise for alphafold3
", "votes": null, "replies": [] } ] } ] }, { "id": 3162478, "author_name": "hengck23", "author_url": "", "post_date": "03/29/2025 09:08:32", "content": "yet another paper:
\nNeuralPLexer3: Accurate Biomolecular Complex Structure Prediction with Flow Models
\nhttps://arxiv.org/pdf/2412.10743v2
qoute:
\n\"• Nucleic acids: On CASP15 RNA targets [32, 33], NP3 demonstrated comparable performance to AF3 (46.5%
\nversus 47.3%, respectively) and slightly below that of Alchemy_RNA2 34, which relies on explicitly
\ncurated human inputs. While AF3 conditions its predictions on RNA MSAs, NP3 uses RNA language
\nmodels (LMs) and achieves similar performance—a noteworthy result, as MSAs are typically considered
\nmore effective than LMs for conditioning structure predictions. This outcome also highlights the potential of
\nutilizing LM-based approaches in scenarios where RNA MSAs are unavailable or impractical to generate\"
however, i believe msa and esm should both in the total solution. you can can make 5 prediction, why don't use multiple methods?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3162419": "I noticed that no one has mentioned this library yet, so I wanted to share it here.\n\nThere’s a library called **OpenComplex**, which extends OpenFold to support structure prediction for RNA and RNA-protein complexes.\n\nhttps://github.com/baaihealth/OpenComplex/tree/master\n\nRight now, I'm exploring the idea of adapting ESMFold —which predicts protein structures from protein language model embeddings— for RNA. I'm experimenting with combining an appropriate RNA language model with ESMFold and OpenComplex.\n\nhttps://github.com/facebookresearch/esm\n\nThat said, these models are a bit outdated, and setting up the environment and resolving dependencies has been a bit tricky... I might share more if I make interesting progress.", "3162462": "https://www.biorxiv.org/content/10.1101/2025.03.25.643589v1\nthe paper and some results", "3162465": "fyi,\n\nseveral af3 clones (proteinx, chai-1) already included esm in their to replace MSA for RNA prediction.\ncheck their repo and technical report", "3162475": "Ah, you're right—I'd forgotten that Chai-1 has a no-MSA option. I'll take a look!\n\nIn any case, I'm working on building a simple structure prediction model from scratch as a way to deepen my understanding of this competition :)", "3162478": "yet another paper:\nNeuralPLexer3: Accurate Biomolecular Complex Structure Prediction with Flow Models\nhttps://arxiv.org/pdf/2412.10743v2\n\nqoute:\n\"• Nucleic acids: On CASP15 RNA targets [32, 33], NP3 demonstrated comparable performance to AF3 (46.5%\nversus 47.3%, respectively) and slightly below that of Alchemy_RNA2 [34] (54.8%), which relies on explicitly\ncurated human inputs. While AF3 conditions its predictions on RNA MSAs, NP3 uses RNA language\nmodels (LMs) and achieves similar performance—a noteworthy result, as MSAs are typically considered\nmore effective than LMs for conditioning structure predictions. This outcome also highlights the potential of\nutilizing LM-based approaches in scenarios where RNA MSAs are unavailable or impractical to generate\"\n\n----\n\nhowever, i believe msa and esm should both in the total solution. you can can make 5 prediction, why don't use multiple methods?", "3162485": "\":In any case, I'm working on building a simple structure prediction model from scratch as a way to deepen my understanding of this competition :)\"\n\nmy suggestion:\n- download alphafold2 supplimentary paper. it has algorithm for all modules like this\n\n\n\n- download drfold2. it implements the above and follows very close to the step and index\n- drfold2 uses RAN LM to replace MSA\n\n---\n\nyou can do likewise for alphafold3" }, "source": "meta" }