![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F31669346%2F3bfa0309ca6dd79012dfd5de7f78cec9%2Frna_structure_prediction_pipeline.svg?generation=1775256307987486&alt=media\")
Without extensive access to GPUs, I realized that to create a viable solution, I couldn't change or fine tune the underlying models, I had to change their usage and pipeline. The main problem I found with many of the top notebooks was their gap interpolation. Basically, when the template used in TBM had a gap that did not match with the query sequence, linear interpolation was used almost exclusively to fill in the gaps, essentially creating a straight line of coordinates in between known coordinates. To solve this, I came up with two methods: Filling with Protenix and with parts of existing TBM templates using Biopython's pairwise aligner.
\nFor Protenix, I took gaps of longer than 3 nt, then ran inference with Protenix using 20 flanking residues on each side to provide extra context, then ran rescaling to ensure that the Protenix output was sized reasonably to a distance of around 5.95 Å between each residue. Then to align it to the rest of the structure I used 5 runs of Iterative Weighted Kabsch Alignment with an RMSD gate of 8 Å, threshold of 3, and a minimum of 10 inliers. My Protenix implementation in my notebook also used dynamic chunking for all predictions, using a minimum chunk size of 20 nt and a maximum of the sequence length divided by 4.
\nWhile reading Biopython's aligner documentation, I came across the local mode for the aligner, which I used to take the query gap as an input and find matching subsequences in the training set. After this, I used the same Iterative Weighted Kabsch Alignment as previously mentioned in order to align the gap to the rest of the sequence.
\nMy 2 notebooks for submission featured one with pure Protenix for gap interpolation, and one with that ran Biopython's local aligner first, then had Protenix as a fallback for when quality thresholds weren't met for gap sequences.\n
Being a complete beginner to this field and a college freshman, this competition has been an extremely valuable learning experience for me, teaching me how to think innovatively, turn my ideas into code, and work with large datasets. I enjoyed how this competition rewarded diversity of solutions and innovation, and I strove to create that with my pipeline. Thank you to the competition organizers and Rhiju Das for this amazing competition.
", "messages": [ { "id": 3435053, "postDate": "2026-04-03T16:01:45.790Z", "content": "Without extensive access to GPUs, I realized that to create a viable solution, I couldn't change or fine tune the underlying models, I had to change their usage and pipeline. The main problem I found with many of the top notebooks was their gap interpolation. Basically, when the template used in TBM had a gap that did not match with the query sequence, linear interpolation was used almost exclusively to fill in the gaps, essentially creating a straight line of coordinates in between known coordinates. To solve this, I came up with two methods: Filling with Protenix and with parts of existing TBM templates using Biopython's pairwise aligner.
\nFor Protenix, I took gaps of longer than 3 nt, then ran inference with Protenix using 20 flanking residues on each side to provide extra context, then ran rescaling to ensure that the Protenix output was sized reasonably to a distance of around 5.95 Å between each residue. Then to align it to the rest of the structure I used 5 runs of Iterative Weighted Kabsch Alignment with an RMSD gate of 8 Å, threshold of 3, and a minimum of 10 inliers. My Protenix implementation in my notebook also used dynamic chunking for all predictions, using a minimum chunk size of 20 nt and a maximum of the sequence length divided by 4.
\nWhile reading Biopython's aligner documentation, I came across the local mode for the aligner, which I used to take the query gap as an input and find matching subsequences in the training set. After this, I used the same Iterative Weighted Kabsch Alignment as previously mentioned in order to align the gap to the rest of the sequence.
\nMy 2 notebooks for submission featured one with pure Protenix for gap interpolation, and one with that ran Biopython's local aligner first, then had Protenix as a fallback for when quality thresholds weren't met for gap sequences.\n
Being a complete beginner to this field and a college freshman, this competition has been an extremely valuable learning experience for me, teaching me how to think innovatively, turn my ideas into code, and work with large datasets. I enjoyed how this competition rewarded diversity of solutions and innovation, and I strove to create that with my pipeline. Thank you to the competition organizers and Rhiju Das for this amazing competition.
", "rawMarkdown": "# **1. Approach**\nWithout extensive access to GPUs, I realized that to create a viable solution, I couldn't change or fine tune the underlying models, I had to change their usage and pipeline. The main problem I found with many of the top notebooks was their gap interpolation. Basically, when the template used in TBM had a gap that did not match with the query sequence, linear interpolation was used almost exclusively to fill in the gaps, essentially creating a straight line of coordinates in between known coordinates. To solve this, I came up with two methods: Filling with Protenix and with parts of existing TBM templates using Biopython's pairwise aligner.\n\n# **2. Method**\n## **2.1 Protenix Gap Filling**\nFor Protenix, I took gaps of longer than 3 nt, then ran inference with Protenix using 20 flanking residues on each side to provide extra context, then ran rescaling to ensure that the Protenix output was sized reasonably to a distance of around 5.95 Å between each residue. Then to align it to the rest of the structure I used 5 runs of Iterative Weighted Kabsch Alignment with an RMSD gate of 8 Å, threshold of 3, and a minimum of 10 inliers. My Protenix implementation in my notebook also used dynamic chunking for all predictions, using a minimum chunk size of 20 nt and a maximum of the sequence length divided by 4.\n\n## **2.2 Biopython Local Pairwise Aligner**\nWhile reading Biopython's aligner documentation, I came across the local mode for the aligner, which I used to take the query gap as an input and find matching subsequences in the training set. After this, I used the same Iterative Weighted Kabsch Alignment as previously mentioned in order to align the gap to the rest of the sequence.\n\n# **3. Implementation**\nMy 2 notebooks for submission featured one with pure Protenix for gap interpolation, and one with that ran Biopython's local aligner first, then had Protenix as a fallback for when quality thresholds weren't met for gap sequences.\n\n\n# **4. Other Explorations not Submitted**\n- One of the earlier public implementations of deep learning methods with TBM, where I used Boltz + RNAPro + TBM, shared below\n- Using the Qwen Reasoning model to filter the RNA sequences into groups based on their description and suggest RNA sequences to try as a modeling template, shared below\n- Combining TBM + Protenix + Boltz2\n- Combining TBM + Protenix + Drfold2\n- Averaging multiple mid to high scoring TBM templates (scores between 0.5 and 0.8)\n\n# **5. Conclusion**\nBeing a complete beginner to this field and a college freshman, this competition has been an extremely valuable learning experience for me, teaching me how to think innovatively, turn my ideas into code, and work with large datasets. I enjoyed how this competition rewarded diversity of solutions and innovation, and I strove to create that with my pipeline. Thank you to the competition organizers and Rhiju Das for this amazing competition.", "votes": 3 } ], "comments": [], "raw_markdown_by_id": { "3435053": "# **1. Approach**\nWithout extensive access to GPUs, I realized that to create a viable solution, I couldn't change or fine tune the underlying models, I had to change their usage and pipeline. The main problem I found with many of the top notebooks was their gap interpolation. Basically, when the template used in TBM had a gap that did not match with the query sequence, linear interpolation was used almost exclusively to fill in the gaps, essentially creating a straight line of coordinates in between known coordinates. To solve this, I came up with two methods: Filling with Protenix and with parts of existing TBM templates using Biopython's pairwise aligner.\n\n# **2. Method**\n## **2.1 Protenix Gap Filling**\nFor Protenix, I took gaps of longer than 3 nt, then ran inference with Protenix using 20 flanking residues on each side to provide extra context, then ran rescaling to ensure that the Protenix output was sized reasonably to a distance of around 5.95 Å between each residue. Then to align it to the rest of the structure I used 5 runs of Iterative Weighted Kabsch Alignment with an RMSD gate of 8 Å, threshold of 3, and a minimum of 10 inliers. My Protenix implementation in my notebook also used dynamic chunking for all predictions, using a minimum chunk size of 20 nt and a maximum of the sequence length divided by 4.\n\n## **2.2 Biopython Local Pairwise Aligner**\nWhile reading Biopython's aligner documentation, I came across the local mode for the aligner, which I used to take the query gap as an input and find matching subsequences in the training set. After this, I used the same Iterative Weighted Kabsch Alignment as previously mentioned in order to align the gap to the rest of the sequence.\n\n# **3. Implementation**\nMy 2 notebooks for submission featured one with pure Protenix for gap interpolation, and one with that ran Biopython's local aligner first, then had Protenix as a fallback for when quality thresholds weren't met for gap sequences.\n\n\n# **4. Other Explorations not Submitted**\n- One of the earlier public implementations of deep learning methods with TBM, where I used Boltz + RNAPro + TBM, shared below\n- Using the Qwen Reasoning model to filter the RNA sequences into groups based on their description and suggest RNA sequences to try as a modeling template, shared below\n- Combining TBM + Protenix + Boltz2\n- Combining TBM + Protenix + Drfold2\n- Averaging multiple mid to high scoring TBM templates (scores between 0.5 and 0.8)\n\n# **5. Conclusion**\nBeing a complete beginner to this field and a college freshman, this competition has been an extremely valuable learning experience for me, teaching me how to think innovatively, turn my ideas into code, and work with large datasets. I enjoyed how this competition rewarded diversity of solutions and innovation, and I strove to create that with my pipeline. Thank you to the competition organizers and Rhiju Das for this amazing competition." } }