Weakness 1

We believe that RiboFlow, building upon the foundation of SOTA generative models, introduces the following innovations:

• By incorporating a distance-aware, ligand-guided module, RiboFlow achieves ligand-conditioned RNA backbone design for the first time. This significantly fills a void in the field of RNA design and introduces a novel design task to the AI4Science community.
• RiboFlow curates the largest currently available dataset of experimentally resolved and validated RNA-ligand complexes. This addresses the problem of data scarcity in the field, which is a significant contribution for data-driven deep learning.

Weakness 2

Thank you for raising this excellent question. We believe that directly replacing the RoseTTAFold2NA module in RNAFlow with AlphaFold 3, while theoretically feasible, would lead to an unacceptably low sampling efficiency in practice, making it difficult to apply to real-world engineering tasks like drug discovery. The core bottleneck lies in RNAFlow's generative paradigm: in each iteration of the sampling process, a large-scale folding network (RoseTTAFold2NA or AlphaFold 3) must be called to predict the structure, causing the total sampling time to accumulate dramatically.

To illustrate this more concretely, we conducted a test:

• On an H100 GPU, folding a 100-nucleotide (nt) RNA with AlphaFold 3 (without computing MSA features) requires approximately 30 seconds for a single inference. Following the five sampling steps recommended in the RNAFlow paper, designing just one 100-nt RNA would take 150 seconds (30s × 5). This is precisely one of the main motivations behind our development of RiboFlow (our model).
• RiboFlow adopts a different framework that generates structures directly in geometric space, thereby bypassing the bottleneck of repeatedly calling a large folding model in a loop. As shown in our experiments (see Figure 10), RiboFlow can generate an RNA of the same length in just 6 to 10 seconds, representing an efficiency improvement of at least 15 to 25 times compared to the aforementioned approach.

Therefore, an efficient generative framework like RiboFlow is crucial for engineering tasks such as drug discovery and sensor design, which require exploring vast chemical spaces and rapidly generating thousands of candidate molecules.

Weakness 3

We will address your concerns point by point:

• We are honored to note that many reviewers have shown great interest in our data processing methods. We apologize that the presentation of the dataset's collection and processing in the current version of the main text and appendix is not sufficiently comprehensive. In a future revision, we plan to highlight the contribution of the RiboBind dataset and add a more detailed description of this workflow to better help the community advance the design of ligand-bound RNA molecules.
• Our model does not use a separate module to enforce this sequence-structure consistency. In fact, this consistency is achieved implicitly through our joint denoising framework. At each denoising step $t$, our model makes predictions conditioned on the joint state of both the noised sequence and structure. This means that when predicting the denoising direction for the sequence, the model must incorporate the current structural information, and vice versa. This continuous, interdependent conditioning process forces the sequence and structure to co-evolve throughout the generation trajectory, ultimately converging to a self-consistent solution.
• Regarding Section 4.5, the ligand information is directly incorporated into the RNA generation during the sampling process. We apologize for the ambiguous description here. We will provide a detailed description of this process in the appendix in a future version.

Question 1

Equation (1) forms the foundation of our generative model. Under the Flow Matching framework, it defines a probabilistic path by which a complete RNA structure, including both its geometry and sequence, evolves from a completely random state ($t = 0$) to its true target state ($t=1$).

• The left-hand side, $p_t(\mathcal{T}_t|\mathcal{T}_1)$, represents the overall conditional probability: the probability of observing a certain “noised” state $\mathcal{T}_t$ at intermediate time $t$, given the final ground-truth RNA structure $\mathcal{T}_1$.
• The right-hand product form is the key to this formulation. It originates from a core modeling simplification: we assume that, conditioned on the true structure $\mathcal{T}_1$, the generative processes of each nucleotide are independent. Therefore, the complex joint probability of the entire RNA can be factorized into the product of three component probabilities for each nucleotide $i$: translation ($p_t(x_t^i|x_1^i)$), rotation ($p_t(r_t^i|r_1^i)$), and category ($p_t(c_t^i|c_1^i)$).

This formulation decomposes the task of designing a complete RNA into $N_t$ independent and more tractable sub-problems—each focusing on the generation of an individual nucleotide’s properties. The model only needs to learn how each nucleotide’s three basic attributes evolve from noise to their true states, without having to directly handle the complex interdependencies among all atoms. This dramatically reduces the complexity of the generative modeling task.

Question 2

We believe that RiboBind is currently the most comprehensive dataset of experimentally resolved and validated data in the field of RNA-ligand interaction research. In our response to reviewer N2Zg, we also compared it with two authoritative datasets, PDBbind and RNAmigos2. We have ensured that all data from PDBbind and RNAmigos2, after sequence deduplication, are included in RiboBind. Meanwhile, we noted that RNAmigos2 also provides some synthetic datasets for RNA-ligand interactions, but this portion of the data is evaluated solely through docking software, without wet-lab experimental validation.

Question 3

The dynamic cropping strategy is inspired by the spatial interface cropping strategy used in AlphaFold 3. By cropping the original structure, we create a smaller structure that contains the core interaction information between RNA and ligand. This allows the model to focus more on the pocket region that interacts with the ligand molecule, thereby enabling the model to learn the structure around the pocket more accurately and efficiently. This is very important for designing RNAs with affinity for a given ligand. Therefore, as you can see, using the cropping method often yields good results on affinity metrics. However, the cropped structure is only a part of the original structure, which impacts the plausibility of the structures generated by the model to some degree. How to better balance the ratio of original data to cropped data is also one of our current research directions.

Limitations

Thank you for your kind words. We are indeed actively exploring protein-conditioned RNA design. However, unlike small molecules, proteins are structurally more complex, which makes learning the protein-RNA interactions a more challenging task. At the same time, inspired by works such as RhoDesign, we are also attempting to use RiboFlow to design and optimize existing fluorescent aptamers. We hope that RiboFlow can serve as a useful tool to advance the development of the RNA engineering field.

Question 1

Effectively capturing the conformational flexibility of RNA is critical for accurate design. The works you mentioned, such as RNAFlow and gRNAde, represent an important direction for modeling flexibility through ensembles. Our model, RiboFlow, adopts another validated strategy. Our approach aims to precisely define a single, physically plausible conformation by explicitly modeling the 8 torsion angles of RNA. We believe this method directly addresses the fundamental physical cause of RNA flexibility, the rotational degrees of freedom of the backbone. Rather than representing multiple discrete final states (an ensemble), we choose to directly model the continuous degrees of freedom that determine these states. This idea of precisely defining biomolecular structures through local frames and torsion angles has been proven to be extremely effective in several landmark works, such as AlphaFold 2 and RNA-FrameFlow.

The ablation study you suggested, removing torsion angle information, is difficult to perform directly in our ligand-conditioned RNA design task, as it would fundamentally undermine the premise of the task itself. To evaluate the interaction between RNA and a ligand, the model must generate a precise conformation that includes key atomic coordinates. The positions of these key atoms are precisely determined by the torsion angles we predict. If the torsion angle module were removed, our model would degenerate into one that can only generate a coarse-grained backbone (e.g., only C4' atoms), making it completely unable to form a chemically complete pocket for ligand binding. Consequently, meaningful ligand-guided design or evaluation would be impossible.

Nevertheless, to address your core question about the importance of torsion angles, we cite the ablation analysis performed by RNA-FrameFlow on the unconditional backbone generation task. (Given the limited time, we directly cite their published results here). Their study systematically compared model performance with varying levels of atomic detail: a. Frame atoms (C3', C4', O4'); b. Frame and 3 non-frame atoms (C1', P, O3'); and c. Frame and 7 non-frame atoms (C1', P, O3', C5', OP1, OP2). Observing the experimental results reveals that capturing flexibility by precisely defining the local geometry of a single conformation is also an effective strategy.

Question 2

Thank you for raising this critical question about the model’s level of structural detail. In fact, our model generates a full-atom structure that includes the majority of heavy atoms (excluding only the base-specific atoms beyond N1/N9), with an output format similar to AlphaFold2. At each denoising step, once the geometry of the backbone is predicted, we reconstruct the full-atom coordinates of each residue based on known nucleotide chemical geometry, namely, ideal bond lengths and bond angles. This is akin to the predefined nucleotide geometry information used in AlphaFold2 and is implemented following the open-source project OpenComplex on GitHub.

This reconstruction process ensures that the nucleobases not only possess correct chemical identities but, more importantly, have physically accurate and chemically reasonable 3D orientations. This allows for precise modeling of the physicochemical environment required for interactions with ligands. As for generating fully atomic RNA models, we are also actively exploring the use of full-atom foundation models such as AlphaFold3 to design RNAs with improved ligand-binding capabilities.

Question 3

I will address your questions point by point:

1. Why is the rotation diffusion schedule linear during training but exponential during sampling?

This design choice is thoroughly explained in the MultiFlow paper (see Appendix I.3 “Diffusion schedule and reduced noise sampling” and I.4 “Hyperparameters”). The paper clearly states that translation and rotation vectors follow different sampling strategies. This is because rotations occur in a curved, non-Euclidean space, the SO(3) manifold, where physically plausible rotational conformations form a narrow target region. During training, the goal is to teach the model to recover the original structure from any level of noise $t$. A linear schedule ensures that the model is uniformly exposed to various noise levels across the training process. However, during sampling, an exponential schedule allocates more discrete steps in the low-noise region, allowing finer refinement of the output and thus significantly improving structural detail and fidelity in the final predictions.

2. Whether the RNA sequence sampling schedule follows a similar or different trajectory?

For sequence sampling, we adopt a linear schedule, which has been explicitly stated in our anonymous codebase.

3. How are the four variable types—translations, rotations, torsions, and nucleotide types—balanced during sampling, given that for $t \in [0, 0.5]$ (even up to $t \approx 0.8$), the RNA structure is far from a valid fold, yet around 80% of the sequence has already been sampled?

Specifically, translations and nucleotide types (as categorical variables) follow a linear schedule, which provides a stable and computationally efficient generation trajectory. In contrast, rotations are governed by an exponential schedule, a strategy validated by models like MultiFlow and FrameFlow, allowing for more computational steps in the later phase of sampling to finely adjust critical structural details such as bond angles. As for torsion angles, inspired by AlphaFold2, we do not model them via flow matching. Instead, at each step, we directly predict them based on the current backbone geometry, as torsion angles are highly dependent on the local structural environment. These three distinct generation processes are coordinated through a shared denoising network. At each step, the network receives the current states of all variables and simultaneously predicts their updates, ensuring that all components co-evolve coherently and converge toward a physically plausible final conformation within a unified framework.

Limitations

Thank you for your comment, which indeed addresses the core positioning of our work. The core innovation of RiboFlow lies in addressing the challenge of generating a higher-quality RNA backbone that is more compatible with specific ligands. The goal of our work is not to create a brand-new sequence design (inverse folding) model intended to surpass existing state-of-the-art tools.

To this end, we proposed a sequence-structure co-design strategy. Here, the role of sequence information is to act as a guide for structure generation, rather than being the final design product itself. We leverage the joint constraints of sequence and structure to help the model explore a more plausible conformational space, thereby generating a superior backbone. The results in Table 1 validate this core idea, showing that the accuracy of the generated backbone is significantly improved by introducing sequence information. Based on this positioning of our model, we believe that a direct benchmark comparison against SOTA inverse folding tools on a sequence recovery task would likely not accurately evaluate the true contribution of our work. This is because it would overlook the value of our model in the critical upstream stage of optimizing and improving backbone quality.

To eliminate this potential misunderstanding and in response to your suggestion, we will add a clear statement to the discussion section of the paper. This will clarify that RiboFlow is positioned as an upstream, high-quality backbone generator, designed to work synergistically with, rather than replace, downstream dedicated inverse folding tools. We believe that such a clarification will more accurately reflect the contribution of our work and place it in the proper context within the field of RNA design.

Weakness 1

Q: Flow on labels seemed off, campbell 2024 used a rate matrix with CTMC formulation, where RiboFlow seemed using a linear interpolation. Did not find explicit theory in the orig paper that said such a linear interpolation is fine.

A: For our sequence design, we primarily referred to FlexSBDD[1] (see Equation 8) and PPFlow[2] (see Section 3.4). In their work, linear interpolation was adopted for categorical types, and their experiments demonstrated that this sampling method is reasonable, a topic which is discussed in detail in both papers. We adopted their approach in RiboFlow, and our experimental results have also confirmed that linear interpolation for nucleotide types is effective.

Weakness 2

Q: Many RNA-small molecule complexes are shaped by tight interactions with proteins which could bias the conformation space. Do the authors filter these complexes out of the dataset?

A: In constructing RiboBind, we primarily followed the workflow outlined in HariBoss (Section 2.1 Construction of the database). Specifically, we first queried all structures containing at least one RNA molecule and one ligand using the RCSB GraphQL API. We then selected those structures where the ligand forms non-covalent interactions with RNA.

It is worth noting that HariBoss did not explicitly exclude complexes containing proteins. If we strictly constrain the dataset to structures without any proteins, we retrieve 1,264 RNA-ligand complex structures (based on an RCSB advanced search with the following filters: Number of Distinct RNA Entities ≥ 1, Number of Distinct Protein Entities = 0, and Total Number of Non-polymer Instances ≥ 1). This set shares about 80% overlap with the broader set of 1,591 structures we obtained using a more relaxed search criterion.

Weakness 3

Q: Many ligands share high chemical/physical similarity, only removing ligands by identity will still lead to leakage. Authors should report ligand similarity across splits.

A: To better understand potential ligand-level leakage, we computed the maximum Tanimoto similarity (based on Morgan fingerprints) between each validation ligand and all training ligands. Under RNA sequence-based splitting, about 65% of validation ligands have similarity ≥0.7 with training ligands; under RNA structure-based splitting, about 60% do. This is expected since our dataset is split based on RNA sequence or structure, and certain ligands co-occur with highly similar or identical RNAs.

We note that the core challenge in our task is generalizing across RNA sequence and structure diversity. The model's performance depends on capturing RNA-ligand interactions rather than memorizing specific ligand structures. Nevertheless, we acknowledge this limitation and plan to incorporate ligand scaffold-based splitting strategies in future work to better disentangle ligand and RNA generalization.

Weaknes 4

Q: My fear is that due to the low data setting, models will simply tend to predict the most likely RNA independent of the ligand.

A: To assess the ligand specificity of our model, we selected two chemically distinct ligands, A2F and GTP (A2F smiles: c1[nH]c2c(nc(nc2n1)F)N, GTP: c1nc2c(n1C3C(C(C(O3)COP(=O)(O)OP(=O)(O)OP(=O)(O)O)O)O)N=C(NC2=O)N), with a Tanimoto similarity of only 0.1096. For each ligand, we used RiboFlow to generate 50 RNA sequences of length 100 nt, followed by gRNAde selection to retain 8 ligand-compatible candidates, and finally used RhoFold to predict their 3D structures. We then selected those with scTM > 0.45 for downstream analysis.

To quantify the structural variation induced by different ligands, we computed the RMSD between RNA backbones generated from A2F and those from GTP. Among all cross-ligand pairs, the minimum RMSD was 3.7 Å, and the maximum reached 7.9 Å. These values suggest that RiboFlow does not simply memorize the most likely RNA structure, but generates ligand-specific folds in response to different chemical inputs.

Weakness 5

Q: There have been other datasets curating RNA-ligand interactions, authors should discuss these and position theirs against them.

A: Thank you for your question. PDBBind is a comprehensive database primarily focused on experimentally measured binding affinity data for protein–ligand complexes, though it also includes a small subset of RNA–ligand complexes. In its latest version (v.2024), it contains 234 RNA–ligand complexes. According to our inspection, all of these entries are already included in the RiboBind dataset, making PDBBind essentially a high-quality subset of RiboBind in the context of RNA–ligand interactions.

The RNAmigos2 dataset, on the other hand, was introduced alongside the recently proposed RNA structure-based virtual screening tool RNAmigos2. We downloaded the dataset released with the paper, which includes 1,740 experimentally validated RNA–ligand binding site structures. After deduplication, we verified that all of these experimental structures are also present in the RiboBind dataset. In addition to the experimental data, the RNAmigos2 dataset also contains 1.3 million synthetic affinity data points, generated via in silico molecular docking. These synthetic interactions were computed by docking approximately 800 ligands (sourced from the ChEMBL database) against the 1,740 experimentally solved RNA structures in all pairwise combinations.

Since RiboBind currently only includes experimentally determined structural data, it remains fully consistent with the experimental portion of the RNAmigos2 dataset. However, we plan to incorporate the synthetic data in future versions of RiboBind to increase the chemical diversity and overall scale of the dataset, thereby enhancing the generative performance of our models.

Question 1

We sincerely apologize for the confusion caused here, as we omitted an experimental table in the appendix that would have supported this conclusion. What we intended to convey is that increasing the number of sampling steps can improve generation quality to some extent, but this improvement saturates beyond a certain point. In our experiments, we observed that around 50 sampling steps, the model’s performance reaches this upper limit, and further increasing the number of steps does not lead to significant gains in generation quality. To clarify this, we now provide a more detailed table supporting our conclusion, using RiboFlow as an example:

---

Ref:

[1] Zhang, Z., Wang, M., & Liu, Q. (2024). Flexsbdd: Structure-based drug design with flexible protein modeling. Advances in Neural Information Processing Systems. [2] Lin, H., Zhang, O., et al. (2024). PPflow: Target-aware peptide design with torsional flow matching. arXiv preprint arXiv:2405.06642.