Thanks to the organizers for a great competition!
\n3-phase ensemble pipeline combining Template-Based Modeling (TBM), Protenix (template-free) prediction, and RNAPro (NVIDIA's Part 1 winner).
\ntest_sequences.csv\n |\n v\n+-----------+ pairwise alignment, top_n=50\n| Phase 1 |--> TBM Predictions (up to 5 slots)\n| TBM |\n+-----------+\n | targets with remaining slots\n v\n+-----------+ 2-5 samples (adaptive)\n| Phase 2 |--> Protenix (template-free)\n| Protenix |\n+-----------+\n | combine TBM + Protenix into 5 slots\n v\nconvert_templates_to_pt_files.py --> templates.pt\n |\n v\n+-----------+\n| Phase 3 | RNAPro (NVIDIA, Part 1 winner)\n| RNAPro | N_sample=1, N_step=200, max_len=1000\n+-----------+\n | merge (slots 1-4: RNAPro, slot 5: Phase 1-2 best)\n v\nsubmission.csv (final)\nPublic notebooks built on TBM + Protenix were scoring 0.43–0.44 on the public LB. When I forked and resubmitted the same notebooks, scores varied by ~0.02–0.03 across reruns due to Protenix's stochastic diffusion. This got me thinking — how much of these scores is real accuracy vs. lucky sampling?\nSince most participants seemed to be on TBM + Protenix, I figured I needed a different model to stand out. That led me to RNAPro.
\nThe core TBM (alignment parameters, constraints, slot perturbations) is mostly the same as public notebooks. The one addition is dynamic slot allocation — adjusting the TBM vs Protenix ratio based on template quality.\nFor each test target, find similar sequences from the template pool (train + validation) using pairwise alignment, ranked by normalized alignment score.
\nInstead of filling all 5 slots with TBM, allocate slots based on how good the best template is:\nAllocate 5 prediction slots based on the best template's quality. The table below shows the maximum TBM slots — the actual count may be lower if the template pool lacks enough sequences passing the similarity/identity thresholds. Remaining slots are filled by Protenix.
\nTemplate Quality Assessment\n------------------------------------------------------------\npct_identity ≥ 80% --> HIGH --> up to 5 TBM\nAND match_rate ≥ 90%\npct_identity ≥ 50% --> MEDIUM --> up to 3 TBM\notherwise --> LOW --> up to 1 TBM\n------------------------------------------------------------\nIn practice, most targets classified as HIGH still had fewer than 5 qualifying templates — e.g., only 1 out of 28 targets achieved a full 5 TBM slots in this run.\nEach TBM slot applies a different perturbation to diversify predictions:
\nSlot 1: Best template (direct adaptation)\nSlot 2: + Gaussian noise (scaled by dissimilarity)\nSlot 3: + Hinge motion (longest chain segment)\nSlot 4: + Per-chain jitter\nSlot 5: + Smooth wiggle deformation\n |\n v\nadaptive_rna_constraints() <-- stereochemical regularization\nKey parameters: top_n=50, MIN_PERCENT_IDENTITY=50%
For targets with remaining empty slots (medium/low quality templates), run Protenix (qiweiyin's adjusted weights) for template-free structure prediction, with MSA and template disabled. The number of Protenix samples is adaptive per target — 2 for medium-quality templates, 4 for low-quality (or 5 if no qualifying TBM template exists). TBM and Protenix predictions are combined into 5 slots per target:
\nPer target:\n+-----------------------------------------------------+\n| Slot 1 | Slot 2 | Slot 3 | Slot 4 | Slot 5 |\n|---------|---------|---------|---------|-----------|\n| TBM | TBM | TBM | Protenix| Protenix | <- medium quality\n| TBM | TBM | TBM | TBM | TBM | <- high quality\n| TBM | Protenix| Protenix| Protenix| Protenix | <- low quality\n+-----------------------------------------------------+\nRun RNAPro (NVIDIA, Part 1 winner, score 0.640) using Phase 1-2 predictions as template input. The RNAPro inference pipeline is based on theoviel's public notebook, adapted to use our Phase 1-2 output as templates. For targets where RNAPro succeeds (≤1000nt), slots 1-4 are replaced with RNAPro predictions while slot 5 retains the Phase 1-2 best prediction. Targets exceeding 1000nt keep Phase 1-2 predictions entirely.
\nPhase 1-2 submission.csv\n |\n v\nconvert_templates_to_pt_files.py --> templates.pt\n |\n v\n +-------------------------+\n | RNAPro |\n | checkpoint: 500M |\n | + RibonanzaNet2 (MSA) |\n | |\n | N_sample=1 |\n | N_step=200, N_cycle=10 |\n | max_len=1000 |\n +------------+------------+\n |\n +------------v------------+\n | Per-target merge: |\n | |\n | RNAPro OK (≤1000nt): |\n | Slots 1-4 <- RNAPro |\n | Slot 5 <- Phase1-2|\n | best |\n | |\n | RNAPro skip (>1000nt) |\n | or fail/timeout: |\n | -> KEEP Phase 1-2 |\n +------------+------------+\n |\n v\n submission.csv (final)\nRNAPro runs sequentially on a single GPU in the submitted notebook.
\nBefore pursuing multi-model ensembles, I tried many variations on the Protenix + TBM pipeline. None of the changes showed clear improvement:
\nAfter running out of ideas on Protenix tuning, I tried other models:
\nAfter integrating RNAPro into the pipeline, I was unable to conduct thorough ablation studies or hyperparameter tuning. The Kaggle GPU weekly quota was nearly exhausted from the preceding experiments, and the competition deadline left little room for systematic exploration of ensemble strategies (e.g., weighted merging, per-target model selection, or increasing RNAPro's N_sample).
+-----------------------------------------------------+\n| Hardware: Kaggle P100 x1 Time limit: 8 hours |\n+-------------+--------------+------------------------+\n| Phase | Time | Notes |\n+-------------+--------------+------------------------+\n| Phase 1 TBM | ~11 min | CPU-only |\n| Phase 2 Ptx | ~79 min | GPU (diffusion model) |\n| Phase 3 RNAPro | ~282 min | GPU (sequential) |\n+-------------+--------------+------------------------+\n| Total | ~6.2 hours | within 8h limit |\n+-------------+--------------+------------------------+\nThank you to the Stanford and Kaggle organizers for hosting this competition, and to all participants who shared their code publicly.
\nThanks to the organizers for a great competition!
\n3-phase ensemble pipeline combining Template-Based Modeling (TBM), Protenix (template-free) prediction, and RNAPro (NVIDIA's Part 1 winner).
\ntest_sequences.csv\n |\n v\n+-----------+ pairwise alignment, top_n=50\n| Phase 1 |--> TBM Predictions (up to 5 slots)\n| TBM |\n+-----------+\n | targets with remaining slots\n v\n+-----------+ 2-5 samples (adaptive)\n| Phase 2 |--> Protenix (template-free)\n| Protenix |\n+-----------+\n | combine TBM + Protenix into 5 slots\n v\nconvert_templates_to_pt_files.py --> templates.pt\n |\n v\n+-----------+\n| Phase 3 | RNAPro (NVIDIA, Part 1 winner)\n| RNAPro | N_sample=1, N_step=200, max_len=1000\n+-----------+\n | merge (slots 1-4: RNAPro, slot 5: Phase 1-2 best)\n v\nsubmission.csv (final)\nPublic notebooks built on TBM + Protenix were scoring 0.43–0.44 on the public LB. When I forked and resubmitted the same notebooks, scores varied by ~0.02–0.03 across reruns due to Protenix's stochastic diffusion. This got me thinking — how much of these scores is real accuracy vs. lucky sampling?\nSince most participants seemed to be on TBM + Protenix, I figured I needed a different model to stand out. That led me to RNAPro.
\nThe core TBM (alignment parameters, constraints, slot perturbations) is mostly the same as public notebooks. The one addition is dynamic slot allocation — adjusting the TBM vs Protenix ratio based on template quality.\nFor each test target, find similar sequences from the template pool (train + validation) using pairwise alignment, ranked by normalized alignment score.
\nInstead of filling all 5 slots with TBM, allocate slots based on how good the best template is:\nAllocate 5 prediction slots based on the best template's quality. The table below shows the maximum TBM slots — the actual count may be lower if the template pool lacks enough sequences passing the similarity/identity thresholds. Remaining slots are filled by Protenix.
\nTemplate Quality Assessment\n------------------------------------------------------------\npct_identity ≥ 80% --> HIGH --> up to 5 TBM\nAND match_rate ≥ 90%\npct_identity ≥ 50% --> MEDIUM --> up to 3 TBM\notherwise --> LOW --> up to 1 TBM\n------------------------------------------------------------\nIn practice, most targets classified as HIGH still had fewer than 5 qualifying templates — e.g., only 1 out of 28 targets achieved a full 5 TBM slots in this run.\nEach TBM slot applies a different perturbation to diversify predictions:
\nSlot 1: Best template (direct adaptation)\nSlot 2: + Gaussian noise (scaled by dissimilarity)\nSlot 3: + Hinge motion (longest chain segment)\nSlot 4: + Per-chain jitter\nSlot 5: + Smooth wiggle deformation\n |\n v\nadaptive_rna_constraints() <-- stereochemical regularization\nKey parameters: top_n=50, MIN_PERCENT_IDENTITY=50%
For targets with remaining empty slots (medium/low quality templates), run Protenix (qiweiyin's adjusted weights) for template-free structure prediction, with MSA and template disabled. The number of Protenix samples is adaptive per target — 2 for medium-quality templates, 4 for low-quality (or 5 if no qualifying TBM template exists). TBM and Protenix predictions are combined into 5 slots per target:
\nPer target:\n+-----------------------------------------------------+\n| Slot 1 | Slot 2 | Slot 3 | Slot 4 | Slot 5 |\n|---------|---------|---------|---------|-----------|\n| TBM | TBM | TBM | Protenix| Protenix | <- medium quality\n| TBM | TBM | TBM | TBM | TBM | <- high quality\n| TBM | Protenix| Protenix| Protenix| Protenix | <- low quality\n+-----------------------------------------------------+\nRun RNAPro (NVIDIA, Part 1 winner, score 0.640) using Phase 1-2 predictions as template input. The RNAPro inference pipeline is based on theoviel's public notebook, adapted to use our Phase 1-2 output as templates. For targets where RNAPro succeeds (≤1000nt), slots 1-4 are replaced with RNAPro predictions while slot 5 retains the Phase 1-2 best prediction. Targets exceeding 1000nt keep Phase 1-2 predictions entirely.
\nPhase 1-2 submission.csv\n |\n v\nconvert_templates_to_pt_files.py --> templates.pt\n |\n v\n +-------------------------+\n | RNAPro |\n | checkpoint: 500M |\n | + RibonanzaNet2 (MSA) |\n | |\n | N_sample=1 |\n | N_step=200, N_cycle=10 |\n | max_len=1000 |\n +------------+------------+\n |\n +------------v------------+\n | Per-target merge: |\n | |\n | RNAPro OK (≤1000nt): |\n | Slots 1-4 <- RNAPro |\n | Slot 5 <- Phase1-2|\n | best |\n | |\n | RNAPro skip (>1000nt) |\n | or fail/timeout: |\n | -> KEEP Phase 1-2 |\n +------------+------------+\n |\n v\n submission.csv (final)\nRNAPro runs sequentially on a single GPU in the submitted notebook.
\nBefore pursuing multi-model ensembles, I tried many variations on the Protenix + TBM pipeline. None of the changes showed clear improvement:
\nAfter running out of ideas on Protenix tuning, I tried other models:
\nAfter integrating RNAPro into the pipeline, I was unable to conduct thorough ablation studies or hyperparameter tuning. The Kaggle GPU weekly quota was nearly exhausted from the preceding experiments, and the competition deadline left little room for systematic exploration of ensemble strategies (e.g., weighted merging, per-target model selection, or increasing RNAPro's N_sample).
+-----------------------------------------------------+\n| Hardware: Kaggle P100 x1 Time limit: 8 hours |\n+-------------+--------------+------------------------+\n| Phase | Time | Notes |\n+-------------+--------------+------------------------+\n| Phase 1 TBM | ~11 min | CPU-only |\n| Phase 2 Ptx | ~79 min | GPU (diffusion model) |\n| Phase 3 RNAPro | ~282 min | GPU (sequential) |\n+-------------+--------------+------------------------+\n| Total | ~6.2 hours | within 8h limit |\n+-------------+--------------+------------------------+\nThank you to the Stanford and Kaggle organizers for hosting this competition, and to all participants who shared their code publicly.
\n