I'm working on finding an optimal 3D RNA structure from multiple predicted conformations, either through consensus modeling or REMC (Replica Exchange Monte Carlo).
\nI've also tried training my own model to refine and select the best structure, but so far the results haven’t been great — the TM-scores remain low.
\nHas anyone had success with this kind of approach or have any insights/tips to share?
I'm working on finding an optimal 3D RNA structure from multiple predicted conformations, either through consensus modeling or REMC (Replica Exchange Monte Carlo).
\nI've also tried training my own model to refine and select the best structure, but so far the results haven’t been great — the TM-scores remain low.
\nHas anyone had success with this kind of approach or have any insights/tips to share?
To answer your question, I'll make a parallel with ensembling as it relates to machine learning.
\nAveraging two identical models will give you the same model. Averaging two near-identical models will again be unlikely to produce a performance boost. For strong ML ensembling, one needs diverse models. That means modeling with non-overlapping expertise, or models that have different strengths.
\nBack to RNA models. If you are using an ensemble of structures that were all made from the same MSA and by the same program, chances are that all of those models make the same mistakes. They won't combine well. You will need RNA models produced by different approaches to have any hope they will be diverse enough for a productive ensemble.
", "rawMarkdown": "To answer your question, I'll make a parallel with ensembling as it relates to machine learning.\n\nAveraging two identical models will give you the same model. Averaging two near-identical models will again be unlikely to produce a performance boost. For strong ML ensembling, one needs diverse models. That means modeling with non-overlapping expertise, or models that have different strengths.\n\nBack to RNA models. If you are using an ensemble of structures that were all made from the same MSA and by the same program, chances are that all of those models make the same mistakes. They won't combine well. You will need RNA models produced by different approaches to have any hope they will be diverse enough for a productive ensemble.", "votes": null }, { "id": "3181482", "postDate": "04/18/2025 00:09:43", "content": "Thank you for sharing your great insight!
", "rawMarkdown": "Thank you for sharing your great insight!", "votes": null }, { "id": "3181522", "postDate": "04/18/2025 02:10:51", "content": "typically one use different methods to make say 100 structures. using different model (deep learning, template matching, simulation …). or use different input (rMSA, SS, etc … different MSA subset …)
\nthen, they use energy function to score and rank predicted structures in CASP.
\nthere are other scoring function that use human knowledge or other information … please check CASP solution
I’ve previously experimented with a model called Protenix, and as you know, it provides predicted PDB structures ranked by confidence scores. However, when I calculate the TM-scores on Kaggle while preserving that original ranking, I notice that the order doesn’t always reflect the actual structural quality. Maybe it’s a bit different from what we see with energy-based methods, but anyway what are your thoughts on situations like this?
", "rawMarkdown": "I’ve previously experimented with a model called Protenix, and as you know, it provides predicted PDB structures ranked by confidence scores. However, when I calculate the TM-scores on Kaggle while preserving that original ranking, I notice that the order doesn’t always reflect the actual structural quality. Maybe it’s a bit different from what we see with energy-based methods, but anyway what are your thoughts on situations like this?", "votes": null } ], "comments": [ { "id": 3181268, "author_name": "tilii7", "author_url": "", "post_date": "04/17/2025 16:45:05", "content": "To answer your question, I'll make a parallel with ensembling as it relates to machine learning.
\nAveraging two identical models will give you the same model. Averaging two near-identical models will again be unlikely to produce a performance boost. For strong ML ensembling, one needs diverse models. That means modeling with non-overlapping expertise, or models that have different strengths.
\nBack to RNA models. If you are using an ensemble of structures that were all made from the same MSA and by the same program, chances are that all of those models make the same mistakes. They won't combine well. You will need RNA models produced by different approaches to have any hope they will be diverse enough for a productive ensemble.
", "votes": null, "replies": [ { "id": 3181482, "author_name": "doheon114", "author_url": "", "post_date": "04/18/2025 00:09:43", "content": "Thank you for sharing your great insight!
", "votes": null, "replies": [] } ] }, { "id": 3181522, "author_name": "hengck23", "author_url": "", "post_date": "04/18/2025 02:10:51", "content": "typically one use different methods to make say 100 structures. using different model (deep learning, template matching, simulation …). or use different input (rMSA, SS, etc … different MSA subset …)
\nthen, they use energy function to score and rank predicted structures in CASP.
\nthere are other scoring function that use human knowledge or other information … please check CASP solution
I’ve previously experimented with a model called Protenix, and as you know, it provides predicted PDB structures ranked by confidence scores. However, when I calculate the TM-scores on Kaggle while preserving that original ranking, I notice that the order doesn’t always reflect the actual structural quality. Maybe it’s a bit different from what we see with energy-based methods, but anyway what are your thoughts on situations like this?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3180694": "I'm working on finding an optimal 3D RNA structure from multiple predicted conformations, either through consensus modeling or REMC (Replica Exchange Monte Carlo).\nI've also tried training my own model to refine and select the best structure, but so far the results haven’t been great — the TM-scores remain low.\nHas anyone had success with this kind of approach or have any insights/tips to share?", "3181268": "To answer your question, I'll make a parallel with ensembling as it relates to machine learning.\n\nAveraging two identical models will give you the same model. Averaging two near-identical models will again be unlikely to produce a performance boost. For strong ML ensembling, one needs diverse models. That means modeling with non-overlapping expertise, or models that have different strengths.\n\nBack to RNA models. If you are using an ensemble of structures that were all made from the same MSA and by the same program, chances are that all of those models make the same mistakes. They won't combine well. You will need RNA models produced by different approaches to have any hope they will be diverse enough for a productive ensemble.", "3181482": "Thank you for sharing your great insight!", "3181522": "typically one use different methods to make say 100 structures. using different model (deep learning, template matching, simulation ...). or use different input (rMSA, SS, etc ... different MSA subset ...) \nthen, they use energy function to score and rank predicted structures in CASP.\nthere are other scoring function that use human knowledge or other information ... please check CASP solution", "3181632": "I’ve previously experimented with a model called Protenix, and as you know, it provides predicted PDB structures ranked by confidence scores. However, when I calculate the TM-scores on Kaggle while preserving that original ranking, I notice that the order doesn’t always reflect the actual structural quality. Maybe it’s a bit different from what we see with energy-based methods, but anyway what are your thoughts on situations like this?" }, "source": "meta" }