This was a solo solution built around a hybrid prediction pipeline that combines template-based modeling (TBM), foundation-model inference using Protenix, and a lightweight de novo RNA generator.
\nRather than relying on a single prediction method, the pipeline dynamically combines multiple sources of structural information and generates five diverse candidate structures for each RNA target. This design was motivated by the competition's best-of-five TM-score evaluation metric.
\nThe overall workflow consists of three stages:
\nAdditional geometry refinement and diversity generation are applied throughout the pipeline.
\nThe evaluation metric selects the best prediction among five submitted structures.
\nAs a result, the objective is not only to maximize the quality of a single prediction but also to maximize structural diversity while maintaining biological plausibility.
\nThe solution therefore focuses on:
\nThe first stage attempts to predict structures using known RNA structures from the competition training and validation datasets.
\nThe template pool was constructed by combining:
\nCoordinates were extracted directly from experimental structures.
\nFor each test sequence, candidate templates were identified using a custom global alignment strategy based on BioPython PairwiseAligner.
\nTemplates were ranked according to:
\nOnly templates satisfying minimum similarity thresholds were retained.
\nOnce a suitable template was identified:
\nThis produced a complete initial 3D structure for the target RNA.
\nMultiple template-derived predictions were generated using different perturbation strategies:
\nThese perturbations increased structural diversity while preserving overall fold topology.
\nAll template-derived structures were refined using a custom RNA constraint system.
\nThe refinement process applied:
\nApproximate backbone distances were enforced:
\nLocal backbone geometry was regularized through iterative smoothing.
\nResidues that became unrealistically close were pushed apart through a repulsion mechanism.
\nRefinement strength was adapted according to template confidence.
\nLower-confidence structures received stronger corrections, while high-confidence structures remained largely unchanged.
\nSequences that could not obtain all five predictions through TBM were forwarded to Protenix.
\nRNA multiple sequence alignments were enabled.
\nExternal templates were disabled to encourage generalization to template-free targets.
\nBecause Protenix has sequence length limitations, long RNAs were divided into overlapping chunks.
\nChunk boundaries were processed using overlap windows.
\nPredicted chunks were merged into a full-length structure using:
\nThis allowed the model to process sequences longer than the model's native context window.
\nA simple RNA helix generator was used as a final fallback mechanism.
\nWhenever TBM and Protenix failed to provide sufficient predictions, idealized A-form helical structures were generated and subsequently refined using the same RNA geometry constraints.
\nThis guaranteed that every target received exactly five valid predictions.
\nThe competition evaluates the best prediction among five submitted structures.
\nTo exploit this property, the pipeline intentionally generated diverse conformations using:
\nThe resulting prediction set covered multiple plausible RNA conformations while preserving physically reasonable geometry.
\nSeveral safeguards were implemented:
\nThese mechanisms ensured that valid predictions were produced for all targets.
\nThe final submission was generated by combining predictions from three sources:
\nAll predictions were refined using RNA-specific geometric constraints and assembled into five candidate structures per target.
\nThe solution was designed to balance three objectives:
\nThis balance proved particularly valuable under the competition's best-of-five TM-score evaluation framework.
\nThe most important observation was that prediction diversity can be nearly as important as prediction accuracy.
\nA hybrid pipeline that combines template information, foundation-model predictions, and geometry-aware refinement was more effective than relying on a single prediction source.
\nThe competition also highlighted the importance of robust engineering solutions for long RNA sequences, template-free targets, and incomplete predictions.
", "messages": [ { "id": 3466710, "postDate": "2026-06-04T15:16:06.830Z", "content": "This was a solo solution built around a hybrid prediction pipeline that combines template-based modeling (TBM), foundation-model inference using Protenix, and a lightweight de novo RNA generator.
\nRather than relying on a single prediction method, the pipeline dynamically combines multiple sources of structural information and generates five diverse candidate structures for each RNA target. This design was motivated by the competition's best-of-five TM-score evaluation metric.
\nThe overall workflow consists of three stages:
\nAdditional geometry refinement and diversity generation are applied throughout the pipeline.
\nThe evaluation metric selects the best prediction among five submitted structures.
\nAs a result, the objective is not only to maximize the quality of a single prediction but also to maximize structural diversity while maintaining biological plausibility.
\nThe solution therefore focuses on:
\nThe first stage attempts to predict structures using known RNA structures from the competition training and validation datasets.
\nThe template pool was constructed by combining:
\nCoordinates were extracted directly from experimental structures.
\nFor each test sequence, candidate templates were identified using a custom global alignment strategy based on BioPython PairwiseAligner.
\nTemplates were ranked according to:
\nOnly templates satisfying minimum similarity thresholds were retained.
\nOnce a suitable template was identified:
\nThis produced a complete initial 3D structure for the target RNA.
\nMultiple template-derived predictions were generated using different perturbation strategies:
\nThese perturbations increased structural diversity while preserving overall fold topology.
\nAll template-derived structures were refined using a custom RNA constraint system.
\nThe refinement process applied:
\nApproximate backbone distances were enforced:
\nLocal backbone geometry was regularized through iterative smoothing.
\nResidues that became unrealistically close were pushed apart through a repulsion mechanism.
\nRefinement strength was adapted according to template confidence.
\nLower-confidence structures received stronger corrections, while high-confidence structures remained largely unchanged.
\nSequences that could not obtain all five predictions through TBM were forwarded to Protenix.
\nRNA multiple sequence alignments were enabled.
\nExternal templates were disabled to encourage generalization to template-free targets.
\nBecause Protenix has sequence length limitations, long RNAs were divided into overlapping chunks.
\nChunk boundaries were processed using overlap windows.
\nPredicted chunks were merged into a full-length structure using:
\nThis allowed the model to process sequences longer than the model's native context window.
\nA simple RNA helix generator was used as a final fallback mechanism.
\nWhenever TBM and Protenix failed to provide sufficient predictions, idealized A-form helical structures were generated and subsequently refined using the same RNA geometry constraints.
\nThis guaranteed that every target received exactly five valid predictions.
\nThe competition evaluates the best prediction among five submitted structures.
\nTo exploit this property, the pipeline intentionally generated diverse conformations using:
\nThe resulting prediction set covered multiple plausible RNA conformations while preserving physically reasonable geometry.
\nSeveral safeguards were implemented:
\nThese mechanisms ensured that valid predictions were produced for all targets.
\nThe final submission was generated by combining predictions from three sources:
\nAll predictions were refined using RNA-specific geometric constraints and assembled into five candidate structures per target.
\nThe solution was designed to balance three objectives:
\nThis balance proved particularly valuable under the competition's best-of-five TM-score evaluation framework.
\nThe most important observation was that prediction diversity can be nearly as important as prediction accuracy.
\nA hybrid pipeline that combines template information, foundation-model predictions, and geometry-aware refinement was more effective than relying on a single prediction source.
\nThe competition also highlighted the importance of robust engineering solutions for long RNA sequences, template-free targets, and incomplete predictions.
", "rawMarkdown": "This was a solo solution built around a hybrid prediction pipeline that combines template-based modeling (TBM), foundation-model inference using Protenix, and a lightweight de novo RNA generator.\n\nRather than relying on a single prediction method, the pipeline dynamically combines multiple sources of structural information and generates five diverse candidate structures for each RNA target. This design was motivated by the competition's best-of-five TM-score evaluation metric.\n\nThe overall workflow consists of three stages:\n\n1. Template-Based Modeling (TBM)\n2. Protenix-based structure prediction\n3. De novo fallback generation\n\nAdditional geometry refinement and diversity generation are applied throughout the pipeline.\n\n---\n\n## Competition Strategy\n\nThe evaluation metric selects the best prediction among five submitted structures.\n\nAs a result, the objective is not only to maximize the quality of a single prediction but also to maximize structural diversity while maintaining biological plausibility.\n\nThe solution therefore focuses on:\n\n* exploiting known structures whenever possible,\n* generating multiple alternative conformations,\n* enforcing RNA geometry constraints,\n* maintaining a complete prediction pipeline even for difficult targets.\n\n---\n\n## Phase 1: Template-Based Modeling\n\nThe first stage attempts to predict structures using known RNA structures from the competition training and validation datasets.\n\n### Template Database\n\nThe template pool was constructed by combining:\n\n* train_sequences.csv\n* validation_sequences.csv\n* train_labels.csv\n* validation_labels.csv\n\nCoordinates were extracted directly from experimental structures.\n\n### Sequence Alignment\n\nFor each test sequence, candidate templates were identified using a custom global alignment strategy based on BioPython PairwiseAligner.\n\nTemplates were ranked according to:\n\n* normalized alignment score\n* sequence identity percentage\n\nOnly templates satisfying minimum similarity thresholds were retained.\n\n### Coordinate Transfer\n\nOnce a suitable template was identified:\n\n1. Query and template sequences were aligned.\n2. Matching residues inherited template coordinates.\n3. Missing residues were reconstructed through interpolation.\n\nThis produced a complete initial 3D structure for the target RNA.\n\n### Structural Diversity\n\nMultiple template-derived predictions were generated using different perturbation strategies:\n\n* direct template transfer,\n* coordinate noise injection,\n* hinge motion transformations,\n* chain-level rigid-body perturbations,\n* smooth backbone deformations.\n\nThese perturbations increased structural diversity while preserving overall fold topology.\n\n---\n\n## RNA Geometry Refinement\n\nAll template-derived structures were refined using a custom RNA constraint system.\n\nThe refinement process applied:\n\n### Bond Constraints\n\nApproximate backbone distances were enforced:\n\n* residue i → i+1 ≈ 5.95 Å\n* residue i → i+2 ≈ 10.2 Å\n\n### Laplacian Smoothing\n\nLocal backbone geometry was regularized through iterative smoothing.\n\n### Self-Avoidance\n\nResidues that became unrealistically close were pushed apart through a repulsion mechanism.\n\n### Confidence-Aware Refinement\n\nRefinement strength was adapted according to template confidence.\n\nLower-confidence structures received stronger corrections, while high-confidence structures remained largely unchanged.\n\n---\n\n## Phase 2: Protenix Inference\n\nSequences that could not obtain all five predictions through TBM were forwarded to Protenix.\n\n### RNA MSA\n\nRNA multiple sequence alignments were enabled.\n\n### Template-Free Inference\n\nExternal templates were disabled to encourage generalization to template-free targets.\n\n### Long Sequence Handling\n\nBecause Protenix has sequence length limitations, long RNAs were divided into overlapping chunks.\n\nChunk boundaries were processed using overlap windows.\n\n### Chunk Stitching\n\nPredicted chunks were merged into a full-length structure using:\n\n* Kabsch alignment,\n* overlap matching,\n* weighted coordinate blending.\n\nThis allowed the model to process sequences longer than the model's native context window.\n\n---\n\n## Phase 3: De Novo Fallback\n\nA simple RNA helix generator was used as a final fallback mechanism.\n\nWhenever TBM and Protenix failed to provide sufficient predictions, idealized A-form helical structures were generated and subsequently refined using the same RNA geometry constraints.\n\nThis guaranteed that every target received exactly five valid predictions.\n\n---\n\n## Generation of Five Predictions\n\nThe competition evaluates the best prediction among five submitted structures.\n\nTo exploit this property, the pipeline intentionally generated diverse conformations using:\n\n* alternative templates,\n* coordinate perturbations,\n* hinge motions,\n* chain jittering,\n* smooth backbone deformations.\n\nThe resulting prediction set covered multiple plausible RNA conformations while preserving physically reasonable geometry.\n\n---\n\n## Robustness Measures\n\nSeveral safeguards were implemented:\n\n* automatic C1' atom identification,\n* coordinate collapse detection,\n* sequence chunking for long RNAs,\n* coordinate padding and recovery,\n* fallback prediction generation.\n\nThese mechanisms ensured that valid predictions were produced for all targets.\n\n---\n\n## Final Solution\n\nThe final submission was generated by combining predictions from three sources:\n\n1. Template-Based Modeling\n2. Protenix foundation-model inference\n3. De novo RNA fallback generation\n\nAll predictions were refined using RNA-specific geometric constraints and assembled into five candidate structures per target.\n\nThe solution was designed to balance three objectives:\n\n* structural accuracy,\n* prediction diversity,\n* inference robustness.\n\nThis balance proved particularly valuable under the competition's best-of-five TM-score evaluation framework.\n\n---\n\n## Lessons Learned\n\nThe most important observation was that prediction diversity can be nearly as important as prediction accuracy.\n\nA hybrid pipeline that combines template information, foundation-model predictions, and geometry-aware refinement was more effective than relying on a single prediction source.\n\nThe competition also highlighted the importance of robust engineering solutions for long RNA sequences, template-free targets, and incomplete predictions.\n", "votes": 1 } ], "comments": [], "raw_markdown_by_id": { "3466710": "This was a solo solution built around a hybrid prediction pipeline that combines template-based modeling (TBM), foundation-model inference using Protenix, and a lightweight de novo RNA generator.\n\nRather than relying on a single prediction method, the pipeline dynamically combines multiple sources of structural information and generates five diverse candidate structures for each RNA target. This design was motivated by the competition's best-of-five TM-score evaluation metric.\n\nThe overall workflow consists of three stages:\n\n1. Template-Based Modeling (TBM)\n2. Protenix-based structure prediction\n3. De novo fallback generation\n\nAdditional geometry refinement and diversity generation are applied throughout the pipeline.\n\n---\n\n## Competition Strategy\n\nThe evaluation metric selects the best prediction among five submitted structures.\n\nAs a result, the objective is not only to maximize the quality of a single prediction but also to maximize structural diversity while maintaining biological plausibility.\n\nThe solution therefore focuses on:\n\n* exploiting known structures whenever possible,\n* generating multiple alternative conformations,\n* enforcing RNA geometry constraints,\n* maintaining a complete prediction pipeline even for difficult targets.\n\n---\n\n## Phase 1: Template-Based Modeling\n\nThe first stage attempts to predict structures using known RNA structures from the competition training and validation datasets.\n\n### Template Database\n\nThe template pool was constructed by combining:\n\n* train_sequences.csv\n* validation_sequences.csv\n* train_labels.csv\n* validation_labels.csv\n\nCoordinates were extracted directly from experimental structures.\n\n### Sequence Alignment\n\nFor each test sequence, candidate templates were identified using a custom global alignment strategy based on BioPython PairwiseAligner.\n\nTemplates were ranked according to:\n\n* normalized alignment score\n* sequence identity percentage\n\nOnly templates satisfying minimum similarity thresholds were retained.\n\n### Coordinate Transfer\n\nOnce a suitable template was identified:\n\n1. Query and template sequences were aligned.\n2. Matching residues inherited template coordinates.\n3. Missing residues were reconstructed through interpolation.\n\nThis produced a complete initial 3D structure for the target RNA.\n\n### Structural Diversity\n\nMultiple template-derived predictions were generated using different perturbation strategies:\n\n* direct template transfer,\n* coordinate noise injection,\n* hinge motion transformations,\n* chain-level rigid-body perturbations,\n* smooth backbone deformations.\n\nThese perturbations increased structural diversity while preserving overall fold topology.\n\n---\n\n## RNA Geometry Refinement\n\nAll template-derived structures were refined using a custom RNA constraint system.\n\nThe refinement process applied:\n\n### Bond Constraints\n\nApproximate backbone distances were enforced:\n\n* residue i → i+1 ≈ 5.95 Å\n* residue i → i+2 ≈ 10.2 Å\n\n### Laplacian Smoothing\n\nLocal backbone geometry was regularized through iterative smoothing.\n\n### Self-Avoidance\n\nResidues that became unrealistically close were pushed apart through a repulsion mechanism.\n\n### Confidence-Aware Refinement\n\nRefinement strength was adapted according to template confidence.\n\nLower-confidence structures received stronger corrections, while high-confidence structures remained largely unchanged.\n\n---\n\n## Phase 2: Protenix Inference\n\nSequences that could not obtain all five predictions through TBM were forwarded to Protenix.\n\n### RNA MSA\n\nRNA multiple sequence alignments were enabled.\n\n### Template-Free Inference\n\nExternal templates were disabled to encourage generalization to template-free targets.\n\n### Long Sequence Handling\n\nBecause Protenix has sequence length limitations, long RNAs were divided into overlapping chunks.\n\nChunk boundaries were processed using overlap windows.\n\n### Chunk Stitching\n\nPredicted chunks were merged into a full-length structure using:\n\n* Kabsch alignment,\n* overlap matching,\n* weighted coordinate blending.\n\nThis allowed the model to process sequences longer than the model's native context window.\n\n---\n\n## Phase 3: De Novo Fallback\n\nA simple RNA helix generator was used as a final fallback mechanism.\n\nWhenever TBM and Protenix failed to provide sufficient predictions, idealized A-form helical structures were generated and subsequently refined using the same RNA geometry constraints.\n\nThis guaranteed that every target received exactly five valid predictions.\n\n---\n\n## Generation of Five Predictions\n\nThe competition evaluates the best prediction among five submitted structures.\n\nTo exploit this property, the pipeline intentionally generated diverse conformations using:\n\n* alternative templates,\n* coordinate perturbations,\n* hinge motions,\n* chain jittering,\n* smooth backbone deformations.\n\nThe resulting prediction set covered multiple plausible RNA conformations while preserving physically reasonable geometry.\n\n---\n\n## Robustness Measures\n\nSeveral safeguards were implemented:\n\n* automatic C1' atom identification,\n* coordinate collapse detection,\n* sequence chunking for long RNAs,\n* coordinate padding and recovery,\n* fallback prediction generation.\n\nThese mechanisms ensured that valid predictions were produced for all targets.\n\n---\n\n## Final Solution\n\nThe final submission was generated by combining predictions from three sources:\n\n1. Template-Based Modeling\n2. Protenix foundation-model inference\n3. De novo RNA fallback generation\n\nAll predictions were refined using RNA-specific geometric constraints and assembled into five candidate structures per target.\n\nThe solution was designed to balance three objectives:\n\n* structural accuracy,\n* prediction diversity,\n* inference robustness.\n\nThis balance proved particularly valuable under the competition's best-of-five TM-score evaluation framework.\n\n---\n\n## Lessons Learned\n\nThe most important observation was that prediction diversity can be nearly as important as prediction accuracy.\n\nA hybrid pipeline that combines template information, foundation-model predictions, and geometry-aware refinement was more effective than relying on a single prediction source.\n\nThe competition also highlighted the importance of robust engineering solutions for long RNA sequences, template-free targets, and incomplete predictions.\n" } }