![\"RNA](\"https://raw.githubusercontent.com/ammomahdavikia/asset-holding/main/rna_data_structures.png\")
Let me go through our journey from the start!
\nThe Stanford RNA 3D Folding Challenge was our first deep dive into applying AI to a new domain: RNA structure prediction. While we had some background in molecular biology and machine learning, combining them in such an open-ended problem was both challenging and exciting.
\nWe joined the competition not only to test our skills, but also to explore how modern deep learning methods could be applied to one of the most fundamental questions in computational biology: How does RNA fold into its functional 3D structure?
\nAs we explored the competition, we quickly realized that RNA 3D folding is not an isolated problem. It is part of a broader scientific effort to predict biomolecular structures — an effort that has been benchmarked for decades in the CASP competition (Critical Assessment of Structure Prediction).
\nFor those unfamiliar, CASP is held every two years and brings together top research groups worldwide to test their methods on blind structure prediction tasks. Reviewing CASP16, the most recent edition that included RNA, made it clear: the path to state-of-the-art RNA folding goes through CASP solutions.
\nHere is a summary of the top-5 performers in CASP16 for RNA folding:
\n| Team | \nCore Model/Framework | \nMSA | \nRefinement Method | \n
|---|---|---|---|
| Vfold | \nVfold-Pipeline, IsRNA, RNAJP | \n✘ | \nSimRNA + Filtering + Templates + Literature information for ranking | \n
| Guangzhou | \nGuijunlab-complex, AlphaFold2, AlphaFold3 | \n✔ | \nCustom model quality assessment (GuijunLab-QA:QA4) + DeepAssembly | \n
| Kiharalab | \nNuFold, DeepFoldRNA, DRfold, RosettaFold2NA, RhoFold, trRosettaRNA, AlphaFold3 | \n✔ | \nRanked by combining Rosetta score and ARES | \n
| Yang-server | \ntrRosettaRNA | \n✔ | \nRestraint-based GD (PyRosetta-based) | \n
| GeneSilico | \nMain: (ModeRNA + SimRNA) / Additional: (RNAComposer, RNA-BRiQ, RNAJP, trRosettaRNA, AlphaFold3) | \n✔ | \nRMSD scoring + Template refinement (QRNAS) | \n
Wow, they used bunch of native and open models. But what are these models? How do they work? Let's dive deeper on the basics.
\nCurrent methods for RNA 3D folding, can be divided in following categories:
\nTop-performing teams combined both established structural biology tools and modern deep learning frameworks, often in creative ensembles.
\nThis huge list raises an important question for us: What exactly do these frameworks do under the hood, and how do they differ?
\nTo answer that, we should get familiar with types of data in RNA folding era. There are four different types of RNA data structures commonly used for RNA 3D folding:
\n
Now we are ready to dive deeper. Here is the quick summary of the inputs for some of the chosen frameworks and their potential novelty:
\n![\"State-of-the-RNArt\"](\"https://raw.githubusercontent.com/ammomahdavikia/asset-holding/main/State-of-the-RNArt.png\")
As you can see, they are very diverse and complementary. For those interested in experimenting, here is a curated list of implementations, along with Kaggle notebooks and GitHub repositories. Feel free to check them out.
\n| Method | \nKaggle | \nGitHub | \n
|---|---|---|
| DRFold2 | \nView on Kaggle | \nRepo | \n
| NuFold | \nView on Kaggle | \nRepo | \n
| RhoFold | \nView on Kaggle #1 View on Kaggle #2 | \nRepo | \n
| RibonanzaNet | \nView on Kaggle | \nRepo | \n
| Protenix (AF3) | \nView on Kaggle | \nRepo | \n
| Boltz-1 (AF3) | \nView on Kaggle | \nRepo | \n
| trRosettaRNA | \nView on Kaggle | \nDownload | \n
Beside main modules, there are some methods for structure refinement. At first, I underestimated refinement, but soon realized it was a bottleneck…
\nThey are used for refining the predicted structure with different approaches in order to stabilize molecules thermodynamically. There are some available methods out there and it is one of the key factors for smooth prediction and filling the gap between GPUs and nature.
\nMy take until now? Disregarding a potential breakthrough solution, top performing solutions would come out of smart ensembling of models + efficient refinement. Also handling long RNAs is a challenge and key factor.
\nWhich models did I choose? I focused primarily on deep learning–based methods, as I faced compute and storage constraints. Thermodynamic methods were too CPU-heavy, and template-based approaches required specialized data handling far beyond my current setup.
\nFrom the surveyed frameworks, I successfully ran inference with: Protenix, Boltz-1, DRFold2, NuFold, RhoFold+, DeepFoldRNA, and trRosettaRNA.
\nI also attempted Vfold, but could not get it running in the Kaggle environment.
By the way, here is the final ensemble I used:
\n![\"State-of-the-RNArt\"](\"https://raw.githubusercontent.com/ammomahdavikia/asset-holding/main/stanford_rna_3d_folding_ensemble.png\")
I explored multiple variations of Protenix, DRFold2, Boltz-1 and trRosettaRNA since they showed far better standalone results.
\nMost available checkpoints were trained several years ago. It is obvious that fine-tuning on recent published RNAs on PDB would improve the performance on newly discovered structures. We tried to fine-tune Protenix on newly published structures, but limited GPU access prevented full experiments. We also tried to reimplement AF3 with some tweaks and vectorized optimization but couldn't make it on time.
\nAdditionally, literature review showed decent performance of trRosettaRNA2 but the model was not publicly released until after the competition ended. I believe it could help the overall performance (specially on small sequences).
\nWe also tried some new approaches handling long sequences but they didn't work either. First, we tried blind chunking of the long sequences plus adding some overlaps between chunks for minimum data loss. We also tried to chunk into sequences based on their position in secondary structure but it didn't lead to performance boost either.
\nI am deeply grateful to Kaggle, Stanford University, and all the organizers and hosts who made this competition possible. Bringing together such a challenging and inspiring problem is no small feat, and it provided a unique opportunity to learn and grow.
\nSpecial thanks to @rhijudas for outstanding hosting and guidance throughout the competition, and to @hengck23 for their remarkable contributions to the Kaggle community — I learned a great deal from your shared work and insights.
\nFinally, a big thanks to the broader Kaggle community, whose discussions, notebooks, and shared experiments created an environment of collaboration and discovery.
\nLet me go through our journey from the start!
\nThe Stanford RNA 3D Folding Challenge was our first deep dive into applying AI to a new domain: RNA structure prediction. While we had some background in molecular biology and machine learning, combining them in such an open-ended problem was both challenging and exciting.
\nWe joined the competition not only to test our skills, but also to explore how modern deep learning methods could be applied to one of the most fundamental questions in computational biology: How does RNA fold into its functional 3D structure?
\nAs we explored the competition, we quickly realized that RNA 3D folding is not an isolated problem. It is part of a broader scientific effort to predict biomolecular structures — an effort that has been benchmarked for decades in the CASP competition (Critical Assessment of Structure Prediction).
\nFor those unfamiliar, CASP is held every two years and brings together top research groups worldwide to test their methods on blind structure prediction tasks. Reviewing CASP16, the most recent edition that included RNA, made it clear: the path to state-of-the-art RNA folding goes through CASP solutions.
\nHere is a summary of the top-5 performers in CASP16 for RNA folding:
\n| Team | \nCore Model/Framework | \nMSA | \nRefinement Method | \n
|---|---|---|---|
| Vfold | \nVfold-Pipeline, IsRNA, RNAJP | \n✘ | \nSimRNA + Filtering + Templates + Literature information for ranking | \n
| Guangzhou | \nGuijunlab-complex, AlphaFold2, AlphaFold3 | \n✔ | \nCustom model quality assessment (GuijunLab-QA:QA4) + DeepAssembly | \n
| Kiharalab | \nNuFold, DeepFoldRNA, DRfold, RosettaFold2NA, RhoFold, trRosettaRNA, AlphaFold3 | \n✔ | \nRanked by combining Rosetta score and ARES | \n
| Yang-server | \ntrRosettaRNA | \n✔ | \nRestraint-based GD (PyRosetta-based) | \n
| GeneSilico | \nMain: (ModeRNA + SimRNA) / Additional: (RNAComposer, RNA-BRiQ, RNAJP, trRosettaRNA, AlphaFold3) | \n✔ | \nRMSD scoring + Template refinement (QRNAS) | \n
Wow, they used bunch of native and open models. But what are these models? How do they work? Let's dive deeper on the basics.
\nCurrent methods for RNA 3D folding, can be divided in following categories:
\nTop-performing teams combined both established structural biology tools and modern deep learning frameworks, often in creative ensembles.
\nThis huge list raises an important question for us: What exactly do these frameworks do under the hood, and how do they differ?
\nTo answer that, we should get familiar with types of data in RNA folding era. There are four different types of RNA data structures commonly used for RNA 3D folding:
\n
Now we are ready to dive deeper. Here is the quick summary of the inputs for some of the chosen frameworks and their potential novelty:
\n
As you can see, they are very diverse and complementary. For those interested in experimenting, here is a curated list of implementations, along with Kaggle notebooks and GitHub repositories. Feel free to check them out.
\n| Method | \nKaggle | \nGitHub | \n
|---|---|---|
| DRFold2 | \nView on Kaggle | \nRepo | \n
| NuFold | \nView on Kaggle | \nRepo | \n
| RhoFold | \nView on Kaggle #1 View on Kaggle #2 | \nRepo | \n
| RibonanzaNet | \nView on Kaggle | \nRepo | \n
| Protenix (AF3) | \nView on Kaggle | \nRepo | \n
| Boltz-1 (AF3) | \nView on Kaggle | \nRepo | \n
| trRosettaRNA | \nView on Kaggle | \nDownload | \n
Beside main modules, there are some methods for structure refinement. At first, I underestimated refinement, but soon realized it was a bottleneck…
\nThey are used for refining the predicted structure with different approaches in order to stabilize molecules thermodynamically. There are some available methods out there and it is one of the key factors for smooth prediction and filling the gap between GPUs and nature.
\nMy take until now? Disregarding a potential breakthrough solution, top performing solutions would come out of smart ensembling of models + efficient refinement. Also handling long RNAs is a challenge and key factor.
\nWhich models did I choose? I focused primarily on deep learning–based methods, as I faced compute and storage constraints. Thermodynamic methods were too CPU-heavy, and template-based approaches required specialized data handling far beyond my current setup.
\nFrom the surveyed frameworks, I successfully ran inference with: Protenix, Boltz-1, DRFold2, NuFold, RhoFold+, DeepFoldRNA, and trRosettaRNA.
\nI also attempted Vfold, but could not get it running in the Kaggle environment.
By the way, here is the final ensemble I used:
\n
I explored multiple variations of Protenix, DRFold2, Boltz-1 and trRosettaRNA since they showed far better standalone results.
\nMost available checkpoints were trained several years ago. It is obvious that fine-tuning on recent published RNAs on PDB would improve the performance on newly discovered structures. We tried to fine-tune Protenix on newly published structures, but limited GPU access prevented full experiments. We also tried to reimplement AF3 with some tweaks and vectorized optimization but couldn't make it on time.
\nAdditionally, literature review showed decent performance of trRosettaRNA2 but the model was not publicly released until after the competition ended. I believe it could help the overall performance (specially on small sequences).
\nWe also tried some new approaches handling long sequences but they didn't work either. First, we tried blind chunking of the long sequences plus adding some overlaps between chunks for minimum data loss. We also tried to chunk into sequences based on their position in secondary structure but it didn't lead to performance boost either.
\nI am deeply grateful to Kaggle, Stanford University, and all the organizers and hosts who made this competition possible. Bringing together such a challenging and inspiring problem is no small feat, and it provided a unique opportunity to learn and grow.
\nSpecial thanks to @rhijudas for outstanding hosting and guidance throughout the competition, and to @hengck23 for their remarkable contributions to the Kaggle community — I learned a great deal from your shared work and insights.
\nFinally, a big thanks to the broader Kaggle community, whose discussions, notebooks, and shared experiments created an environment of collaboration and discovery.
\nCongrats 🎉
", "rawMarkdown": "Congrats 🎉", "votes": null } ], "comments": [ { "id": 3297080, "author_name": "mokar2001", "author_url": "", "post_date": "10/02/2025 10:14:26", "content": "Congrats 🎉
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3296861": "Let me go through our journey from the start!\n## **Introduction & Motivation**\n\nThe Stanford RNA 3D Folding Challenge was our first deep dive into applying AI to a new domain: RNA structure prediction. While we had some background in molecular biology and machine learning, combining them in such an open-ended problem was both challenging and exciting. \n\nWe joined the competition not only to test our skills, but also to explore how modern deep learning methods could be applied to one of the most fundamental questions in computational biology: **How does RNA fold into its functional 3D structure?**\n## **Background: RNA 3D Folding Landscape**\n\nAs we explored the competition, we quickly realized that RNA 3D folding is not an isolated problem. It is part of a broader scientific effort to predict biomolecular structures — an effort that has been benchmarked for decades in the [**CASP competition** (Critical Assessment of Structure Prediction)](https://predictioncenter.org/). \n\nFor those unfamiliar, CASP is held every two years and brings together top research groups worldwide to test their methods on blind structure prediction tasks. Reviewing CASP16, the most recent edition that included RNA, made it clear: **the path to state-of-the-art RNA folding goes through CASP solutions.** \n\nHere is a summary of the top-5 performers in CASP16 for RNA folding:\n\n| Team | Core Model/Framework | MSA | Refinement Method |\n| --------------- | ----------------------------------------------------------------------------------------------- | --- | ------------------------------------------------------------------- |\n| **Vfold** | Vfold-Pipeline, IsRNA, RNAJP | ✘ | SimRNA + Filtering + Templates + Literature information for ranking |\n| **Guangzhou** | Guijunlab-complex, AlphaFold2, AlphaFold3 | ✔ | Custom model quality assessment (GuijunLab-QA:QA4) + DeepAssembly |\n| **Kiharalab** | NuFold, DeepFoldRNA, DRfold, RosettaFold2NA, RhoFold, trRosettaRNA, AlphaFold3 | ✔ | Ranked by combining Rosetta score and ARES |\n| **Yang-server** | trRosettaRNA | ✔ | Restraint-based GD (PyRosetta-based) |\n| **GeneSilico** | Main: (ModeRNA + SimRNA) / Additional: (RNAComposer, RNA-BRiQ, RNAJP, trRosettaRNA, AlphaFold3) | ✔ | RMSD scoring + Template refinement (QRNAS) |\n\n## **Lessons from CASP16**\n\nWow, they used bunch of native and open models. But what are these models? How do they work? Let's dive deeper on the basics.\n\nCurrent methods for RNA 3D folding, can be divided in following categories:\n+ **Ab initio** (or prediction-based) methods tend to simulate the physics of the system.\n+ **Template-based** (or fragment-assembly) approaches rely on the fact that molecules that have evolution similitude adopt similar structures. A database of known RNA structures is used as a reference.\n+ **Deep learning** approaches use available data to create a neural network architecture that predicts RNA 3D structures from different data sources.\n\nTop-performing teams combined both **established structural biology tools** and **modern deep learning frameworks**, often in creative ensembles.\n\nThis huge list raises an important question for us: **What exactly do these frameworks do under the hood, and how do they differ?**\n\n## **Data Foundations**\n\nTo answer that, we should get familiar with types of data in RNA folding era. There are four different types of RNA data structures commonly used for RNA 3D folding:\n\n1. **Raw Sequence**: The linear string of nucleotides (A, U, C, G) that make up the RNA molecule. This primary sequence serves as the foundation for predicting higher-order structures and functions.\n2. **Secondary Structure**: Represents the base-pairing interactions within an RNA sequence, typically visualized as dot-bracket notation or base-pair probability matrices. It captures elements like stems, loops, and bulges, which are crucial for understanding RNA folding and function.\n3. **Tertiary Structure**: The 3D spatial conformation of the RNA molecule, describing how it folds in real space. This includes complex interactions like pseudoknots and long-range contacts, often derived from experimental techniques or computational modeling.\n4. **Multiple Sequence Alignment (MSA)**: An alignment of homologous RNA sequences from different organisms. MSAs highlight conserved regions and covariation patterns, offering evolutionary insights that can improve structural and functional predictions.\n \n\n\n## **Survey of Core Frameworks**\n\nNow we are ready to dive deeper. Here is the quick summary of the inputs for some of the chosen frameworks and their potential novelty:\n\n\n\nAs you can see, they are very diverse and complementary. For those interested in experimenting, here is a curated list of implementations, along with Kaggle notebooks and GitHub repositories. Feel free to check them out.\n\n\n| Method | Kaggle | GitHub |\n|--------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------|\n| **DRFold2** | [View on Kaggle](https://www.kaggle.com/code/hengck23/lb0-321-simple-drfold-no-msa) | [Repo](https://github.com/leeyang/DRfold2) |\n| **NuFold** | [View on Kaggle](https://www.kaggle.com/code/amirmmahdavikia/stanford-rna-3d-folding-nufold-inference) | [Repo](https://github.com/kiharalab/NuFold) |\n| **RhoFold** | [View on Kaggle #1](https://www.kaggle.com/code/ogurtsov/rhofold-ribonanzanet-msas-lb-0-215)