![\"![\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8111559%2Fff9dfe3c79a024a4877087946c20f4ff%2F1234.png?generation=1759458540965027&alt=media\")
First of all, thanks to Kaggle and the organizers for putting together such a challenging and interesting competition.
\nWe finished 3rd on the public leaderboard before the unseen data release, and also 3rd on the private leaderboard.
\nEarlier we shared a short overview of our solution. This post goes into more detail.
\nrMSA
\nWe generated our own rMSA data using the official rMSA code provided by the hosts.
Training dataset
\nWe tested two versions:
Both versions of the training data and labels are available here: Google Drive.
\nWe built our ensemble using models from these official repositories:
\n\nWe initially explored new architectures, but given the time and resource limits, it was more effective to focus on fine-tuning and combining existing models.
\nmax_steps; a full dataset run took about a day. We are sharing our original submission notebooks. Some commented-out code remains, which shows parts of our experiment history.
\nMany thanks to the community members who shared resources and insights:
\nFinally, thanks to Eigen Company for providing resources for this competition.
", "messages": [ { "id": "3295548", "postDate": "09/29/2025 04:48:51", "content": "First of all, thanks to Kaggle and the organizers for putting together such a challenging and interesting competition.
\nWe finished 3rd on the public leaderboard before the unseen data release, and also 3rd on the private leaderboard.
\nEarlier we shared a short overview of our solution. This post goes into more detail.
\nrMSA
\nWe generated our own rMSA data using the official rMSA code provided by the hosts.
Training dataset
\nWe tested two versions:
Both versions of the training data and labels are available here: Google Drive.
\nWe built our ensemble using models from these official repositories:
\n\nWe initially explored new architectures, but given the time and resource limits, it was more effective to focus on fine-tuning and combining existing models.
\nmax_steps; a full dataset run took about a day. We are sharing our original submission notebooks. Some commented-out code remains, which shows parts of our experiment history.
\nMany thanks to the community members who shared resources and insights:
\nFinally, thanks to Eigen Company for providing resources for this competition.
", "rawMarkdown": "# Our 3rd Place Solution – Stanford RNA 3D Folding\n\nFirst of all, thanks to Kaggle and the organizers for putting together such a challenging and interesting competition. \n\nWe finished **3rd on the public leaderboard** before the unseen data release, and also **3rd on the private leaderboard**. \n\nEarlier we shared a short overview of our solution. This post goes into more detail. \n\n---\n\n## Summary of Our Solution\n- Ensemble of **DRfold2**, **Protenix**, and **Boltz-1** \n- Fine-tuned **Protenix** with newly released RNA datasets \n\n---\n\n## Detailed Solution\n\n### Dataset\n\n- **rMSA** \n We generated our own rMSA data using the [official rMSA code](https://github.com/pylelab/rMSA) provided by the hosts. \n - The v2 rMSA released by the organizers did not cover the full set of recently published data, so we had to build our own. \n - This process took about **14 days**, even with multiprocessing and several servers. \n - Our rMSA data is available here: [Google Drive](https://drive.google.com/drive/folders/15bhXWsR6QuDQo6U4j8Ii8-OE4B4SmC7i?usp=drive_link). \n\n- **Training dataset** \n We tested two versions: \n 1. **Full RNA dataset** – included all recently available data, such as [CASP16](https://www.kaggle.com/datasets/tant64/casp16) uploaded by @tant64. To check quality, we compared some of the labels with the [CASP16 GitLab repository](https://gitlab.com/arneelof/CASP16-predictions/-/tree/main/rna_results?ref_type=heads). \n 2. **RNA-only dataset** – excluded complexes (RNA bound to proteins/DNA), following the clarification from the hosts [here](https://www.kaggle.com/competitions/stanford-rna-3d-folding/discussion/575745). \n\nBoth versions of the training data and labels are available here: [Google Drive](https://drive.google.com/drive/folders/1XKYzk2oCcHPt6DB_wLNYL-7w3s1R793n). \n\n---\n\n### Models\n\nWe built our ensemble using models from these official repositories: \n- [DRfold2](https://github.com/leeyang/DRfold2) \n- [Protenix](https://github.com/bytedance/Protenix) \n- [Boltz](https://github.com/jwohlwend/boltz) \n\nWe initially explored new architectures, but given the time and resource limits, it was more effective to focus on fine-tuning and combining existing models. \n\n---\n\n#### DRfold2\n- Performed best on sequences **<400 nt** with v1 data. \n- Running the full pipeline with optimization, clustering, and all 80 released checkpoints was too slow. \n- We found that **energy selection + Arena** gave most of the gain, while the other steps added little to the TM-score but took significant runtime. \n- So our DRfold2 setup used **energy selection + Arena** only for the post-processing. \n\n---\n\n#### Protenix\n- Trained on an **NVIDIA GH200 (96GB VRAM)**. \n- Limited to sequences **<800 nt** due to memory issues. \n- Used fine-tuning code from @lihaoweicvch ([discussion link](https://www.kaggle.com/competitions/stanford-rna-3d-folding/discussion/573495)). \n- Fine-tuning **without rMSA** did not help, so we trained **with rMSA**. \n- Only modified `max_steps`; a full dataset run took about a day. \n- Early checkpoints (2–3 cycles) performed worse; performance improved with longer fine-tuning. \n- We also tested **multiple Protenix outputs → DRfold energy selection → Arena**, but it did not improve results. \n- For inference, we modified [@geraseva’s code](https://www.kaggle.com/code/geraseva/protenix) to include rMSA. \n\n---\n\n#### Boltz-1\n- While not the strongest model on its own, including **one Boltz prediction** improved ensemble diversity. \n- This helped because the scoring metric favored picking the best out of multiple diverse outputs. \n- We used [@youhanlee’s inference notebook](https://www.kaggle.com/code/youhanlee/boltz-1-inference-submission). \n\n---\n\n### Final Submissions & Code\n\nWe are sharing our original submission notebooks. Some commented-out code remains, which shows parts of our experiment history. \n\n1. **[DRfold2 + Protenix](https://www.kaggle.com/code/yekim102/drfold4-pro-msa-pro-msa-base-rna)** \n - Public LB: **0.60338** | Private LB: **0.52787** \n - <400 nt: 3 × DRfold2 + 2 × Protenix (RNA-only, with MSA) \n - >400 nt: 2 × Protenix (RNA-only) + 2 × Protenix (All) + Protenix baseline \n\n2. **[Protenix + Boltz](https://www.kaggle.com/code/yekim102/fin-pro-all-1-rna-2-base1-boltz1)** \n - Public LB: **0.61253** | Private LB: **0.54312** \n - 2 × Protenix (RNA-only) + Protenix (All) + Protenix baseline + Boltz baseline \n\n---\n\n## Acknowledgements\n\nMany thanks to the community members who shared resources and insights: \n- **@hengck23** – for extensive discussions and guidance. \n- **@geraseva** – for early Protenix inference code. \n- **@lihaoweicvch** – for Protenix fine-tuning code. \n- **@youhanlee** – for updated Boltz inference code. \n\nFinally, thanks to **Eigen Company** for providing resources for this competition.", "votes": null }, { "id": "3295653", "postDate": "09/29/2025 09:47:41", "content": "congradualation.
\nHow do you ensemble the outputs of different models?
\nBecause each model outputs coordinates.
@garybios The competition's submission.csv expects competitors to predict 5 different \"candidates\" for each input sequence. Most competitors ensembled models in 3 ways:
Hope this answers your question
", "rawMarkdown": "garybios The competition's `submission.csv` expects competitors to predict 5 different \"candidates\" for each input sequence. Most competitors ensembled models in 3 ways:\n1. **Length based:** check the length of the input sequence, if L > threshold use model A if L <= threshold use model B. Example: if the sequence is shorter than 400 nucleotides use DRFold2 and if its longer use Protenix.\n2. **Vertical ensemble:** if you have 5 possible candidates use X/5 for model A and Y/5 for model B. Example: use 2 candidates for Protenix and 3 for DRFold2.\n3. **Align and aggregate:** align all sequences to a reference coordinate frame and then aggregate (using the mean usually). I have not seen many succesfull attempts on this approach though.\n\nHope this answers your question", "votes": null }, { "id": "3297305", "postDate": "10/02/2025 20:16:00", "content": "@yekim102 @ryankim99 @ouiqdmw thanks for a beautiful writeup and sharing your insights as early as June! Congrats on having the top AF3-like solution.
\nA question for you. In earlier discussions, it seemed like use of CASP16 blind predictions to train Protenix led to a big increase in accuracy on public LB:
\n\n\nWe included Casp 16 data so that bumps up the public score.
\n
\n If we don't use the casp 16 data, we get about 0.42.
\n (from https://www.kaggle.com/competitions/stanford-rna-3d-folding/discussion/584487)
But actually… our Private LB does not include any CASP16 targets, and your model still did super well!
\nSo I wonder if CASP16 predictions were so necessary.
\nDo you have a checkpoint somewhere where you trained without CASP16 targets (but perhaps with a very up to date PDB)? I wonder how it scores on Private LB.
", "rawMarkdown": "yekim102 @ryankim99 @ouiqdmw thanks for a beautiful writeup and sharing your insights as early as June! Congrats on having the top AF3-like solution.\n\nA question for you. In earlier discussions, it seemed like use of CASP16 blind predictions to train Protenix led to a big increase in accuracy on public LB:\n\n>We included Casp 16 data so that bumps up the public score.\n> If we don't use the casp 16 data, we get about 0.42.\n(from https://www.kaggle.com/competitions/stanford-rna-3d-folding/discussion/584487) \n\nBut actually... our Private LB does not include any CASP16 targets, and your model still did super well! \n\nSo I wonder if CASP16 predictions were so necessary.\n\nDo you have a checkpoint somewhere where you trained *without* CASP16 targets (but perhaps with a very up to date PDB)? I wonder how it scores on Private LB.", "votes": null }, { "id": "3297374", "postDate": "10/03/2025 02:30:17", "content": "Thank you Rhiju!
\nUnfortunately, we don’t have experiments without the CASP16 targets, as we were limited by computational resources.
\nThe RNA-only trained model performed well on the private LB, so I believe the CASP16 targets were not what boosted the score. Instead, the improvement likely came from fitting to the RNA-only targets (excluding those labeled as “complex”).
congradualation.
\nHow do you ensemble the outputs of different models?
\nBecause each model outputs coordinates.
@garybios The competition's submission.csv expects competitors to predict 5 different \"candidates\" for each input sequence. Most competitors ensembled models in 3 ways:
Hope this answers your question
", "votes": null, "replies": [] } ] }, { "id": 3297305, "author_name": "rhijudas", "author_url": "", "post_date": "10/02/2025 20:16:00", "content": "@yekim102 @ryankim99 @ouiqdmw thanks for a beautiful writeup and sharing your insights as early as June! Congrats on having the top AF3-like solution.
\nA question for you. In earlier discussions, it seemed like use of CASP16 blind predictions to train Protenix led to a big increase in accuracy on public LB:
\n\n\nWe included Casp 16 data so that bumps up the public score.
\n
\n If we don't use the casp 16 data, we get about 0.42.
\n (from https://www.kaggle.com/competitions/stanford-rna-3d-folding/discussion/584487)
But actually… our Private LB does not include any CASP16 targets, and your model still did super well!
\nSo I wonder if CASP16 predictions were so necessary.
\nDo you have a checkpoint somewhere where you trained without CASP16 targets (but perhaps with a very up to date PDB)? I wonder how it scores on Private LB.
", "votes": null, "replies": [ { "id": 3297374, "author_name": "yekim102", "author_url": "", "post_date": "10/03/2025 02:30:17", "content": "Thank you Rhiju!
\nUnfortunately, we don’t have experiments without the CASP16 targets, as we were limited by computational resources.
\nThe RNA-only trained model performed well on the private LB, so I believe the CASP16 targets were not what boosted the score. Instead, the improvement likely came from fitting to the RNA-only targets (excluding those labeled as “complex”).