We sincerely appreciate the reviewer’s thoughtful comments and the ideas provided for additional experiments.

We understand that the synthetic task serves as a controlled environment to test the model’s performance, and we agree that the simplifications involved may not fully reflect the complexity of real-world scenarios. However, our intention was to provide a baseline evaluation to better understand the model’s behavior in a controlled setting before applying it to more complex, real-world tasks. We will clarify this in the revised manuscript to ensure that the purpose of the synthetic task is better understood.

The length of 50 was selected, as our main application goal is the design of aptamers, which are typically short (20-100 bases). While experiments with longer sequences will likely degrade the performance of other methods, this will not affect BAnG. ROPE (or an alternative positional encoding) is required to incorporate nucleotide order and relative distance in chain information. Furthermore, while variable sequence length may negatively impact the performance of other methods, it should have no influence on BAnG. These points are supported by the additional experiments we have conducted, the results will be provided in the revised manuscript and can be found here.

We are grateful for the reviewer’s suggestions, which have helped guide us in exploring these additional experiments. We hope the clarifications and new results will address the concerns raised and the reviewer will consider increasing their score. If there are any additional specific changes or analyses you would like us to address, we are more than happy to make further revisions.

Thank you for your feedback and for bringing attention to key aspects of our study. Below, we address your concerns and clarify key points regarding our methodology and evaluation.

RNA structural information from PDB was used only during data preprocessing to identify interacting nucleotides; however, it was not incorporated into the training process. Our model does not require RNA structure information for inference, nor does it internally predict or rely on it in any way. In contrast, some other methods depend on structural data for training, inference, or both, making our approach more broadly applicable, especially in cases where such information is unavailable. We acknowledge that certain explanations may have been challenging to interpret and will take this into account in our text revisions.

For training, we filtered proteins to include only those shorter than 500 residues, while RNA and DNA sequences were not subject to length restrictions (Appendix C.1). Consequently, proteins in our dataset tend to be shorter. Nucleotide sequences were cropped during training, leading to an effective training length of 205±114 residues, primarily to reduce computational memory requirements. Long RNA sequences are typically derived from the ribosome. Given that our ultimate goal is to generate aptamers—short RNA sequences—the cropping is appropriate, as our model doesn't need to learn long-range dependencies. We acknowledge that a sequence-based split is not ideal; however, it was applied solely during training to enhance data diversity and facilitate optimal weight selection, rather than to support any specific claims (Appendix C.1). Our test set contains proteins with less than 25% sequence similarity to the training data (Appendix F). Given that proteins with such low similarity are estimated to have less than a 10% chance of sharing a similar fold (doi.org/10.1093/protein/12.2.85), this suggests that our model demonstrates a degree of generalization across both structural and sequential spaces.

Evaluating generative models is inherently challenging, as direct comparison with ground truth is not possible. Additionally, to the best of our knowledge, there are no established computational protocols for assessing RNA affinity to proteins. Given these challenges, we opted to use a deep learning method for scoring, similar to practices in other domains like computer vision. Among available methods, we selected DeepCLIP due to several key advantages: it has undergone experimental validation, is retrainable for individual proteins, has demonstrated strong performance in benchmark studies (doi:10.1093/bib/bbad307), and does not rely on RNA structural information. We took extensive measures to verify its suitability (detailed in Appendix E.1), including rigorous training and testing of DeepCLIP models for each evaluation sample and selecting only those with exceptionally high performance metrics (AUROC > 0.95). While the reviewer suggests that such models may have a tendency for false positives and negatives, this would only imply that RNA-BAnG performs even better than our current estimates indicate. Nonetheless, we ensured that our DeepCLIP training and testing datasets contained an equal number of positive and negative samples to mitigate any potential biases (Appendix E.1).

The claim that aptamer binding motifs are the primary determinants of protein-RNA interactions, while the rest of the RNA sequence plays a lesser role, is supported by the findings of a prior study (doi.org/10.1038/nature12311). This paper, which provided the experimental data used in our evaluation, also states that RNA structure is not a critical factor in protein interactions for most of the analyzed samples.

Targeting specific protein binding sites with our model is currently possible through protein truncation. We did not incorporate a more direct method, since the primary protein-aptamer interaction experiment, SELEX, is not binding site specific as well. It has not been a priority for the current study, but it could be explored in our future work.

We appreciate the reviewer’s insights, which have helped us refine our explanations. We will also ensure that our model's limitations are properly discussed in the conclusion. Given the additional clarifications provided, we hope the reviewer will consider reassessing their score. If there are additional aspects you would like us to address or specific analyses you would like to see, we are open to further suggestions.

Thank you for your helpful feedback. We appreciate your suggestions and have made adjustments accordingly. Below, we address the points you raised regarding our methods and analysis. Additionally, we are exploring ways to provide a more detailed analysis of the binding motifs, as suggested, to further improve the clarity and depth of our evaluation.

In response to your suggestion, we have conducted ablation studies to evaluate the impact of non-traditional architecture choices, such as positional encodings in Cross Attention and Geometrical Attention. These studies have demonstrated the significant contribution of these components to the model’s performance, and the results will be included in the Appendix of the revised manuscript for further clarity.

We hope that these changes address the concerns raised and improve the overall quality of our manuscript. Thank you again for your insightful comments and suggestions. If there are any additional specific changes or analyses you would like us to address, we are more than happy to make further revisions.

We appreciate the reviewer’s insightful comments. Let us now address the specific questions and concerns raised in the review regarding our evaluation choices and methodological decisions.

Expanding the comparison to additional baseline models is challenging because most existing models (cited in the paper) rely on large training sets of RNA sequences known to interact with a specific protein to generate new interacting sequences. However, such datasets are often difficult to obtain or may not exist at all, making our model—which operates without this requirement—more versatile and practical. This same constraint also prevents us from expanding the GenerRNA comparison set. Regarding RNAFLOW, its assessment on proteins with no sequence similarity to our training data was intended to make the comparison more rigorous for our model and should not have negatively impacted RNAFLOW’s performance. The decision to restrict the comparison set in this way was due to RNAFLOW’s long generation time and its clear limitations in the given settings, particularly its inability to operate on non-truncated proteins. Since truncation is not technically possible—binding sites are unknown, just as they typically are in real-world applications—our evaluation setup naturally reflects this constraint.

Both comparisons with baseline models are quantified using statistical measures. Each sample consists of 1,000 generated RNA sequences, and we evaluate performance by calculating the proportion of sequences that achieve high affinity scores (section 4.4). This ensures a robust statistical basis for our comparisons and allows for a meaningful assessment of model performance. To provide a more quantitative comparison, we computed the proportion of sequences above a threshold, similarly to Figure 6, and calculated areas under threshold-dependent performance curves for both RNAFLOW and GenerRNA. The results further strengthen our claim and can be found here.

Evaluating generative models is inherently challenging, as direct comparison with ground truth is not possible. Additionally, to the best of our knowledge, there are no established computational protocols for assessing RNA affinity to proteins. DeepCLIP is experimentally validated, does not rely on RNA structure (which is unavailable in our case), and has demonstrated strong performance in benchmark studies (doi:10.1093/bib/bbad307). To address evaluation challenges, we have added a more traditional approach by calculating sequence similarity using BLASTN. Results for this evaluation can be found here.

Regarding BAnG’s broader applicability, its extension to alternative biological tasks, such as protein-peptide interactions, is planned for future work as it requires substantial additional effort, particularly in data mining and preparation, making it beyond the scope of this paper.

While the effectiveness of BAnG for multiple mutually exclusive motifs was verified with the DoubleBind task (Table 2), its performance when multiple motifs appear simultaneously in a sequence depends on the uncertainty of their relative positioning. If motif placement is completely random, BAnG will face similar challenges to autoregressive models. However, if relative positions are fixed, BAnG’s performance remains stable. These points are supported by the additional experiments we have conducted, their results can be found here. Ultimately, BAnG is designed to reduce uncertainty in the target probability distribution, and increased noise (uncertainty) in the data makes learning more difficult for any model.

Anchor point selection plays a key role in minimizing uncertainty in the resulting factorization distribution. In synthetic tasks, any residue with a fixed distance from the synthetic motif could serve as an anchor point. However, in real-world scenarios, binding motifs vary in size and content, making random selection among interacting residues a practical approach to reduce uncertainty and diversify training data. In general, anchor point selection is driven by the goal of uncertainty reduction, but the optimal choice remains case-dependent.

Finally, regarding decoding strategies, ProteinMPNN’s sampling method is equivalent to the Iterative Max Logit strategy tested in our manuscript. However, extending BAnG to allow decoding in any order contradicts its core probability factorization and is therefore not feasible.

We are grateful for the reviewer’s thoughtful feedback, which have helped clarify key aspects of our approach and provided valuable ideas for additional experiments and evaluation metrics. We hope our responses address the reviewer's concerns and the reviewer will consider increasing their score. If there are any additional specific changes or analyses you would like us to address, we are more than happy to make further revisions.

We appreciate the reviewer’s thoughtful feedback and suggestions. We will add the necessary citations and work on further clarifying our methods and biological explanations to ensure the content is more accessible.

In response to the reviewer’s request, we have conducted ablation studies to assess the impact of non-traditional architecture choices, such as positional encodings in Cross Attention and Geometrical Attention. These studies have demonstrated the valuable contribution of these components to the model’s performance, and the results will be included in the Appendix for further clarity.

We agree that an explanation of RNA-binding motifs would strengthen the biological motivation for our approach. Aptamer RNA binding motifs are short, conserved sequence patterns that serve as recognition sites for RNA-binding proteins (RBPs). Many RBPs specifically recognize distinct nucleotide sequences, such as AU-rich or GU-rich elements, while others rely on secondary structures like hairpins and stem-loops for recognition, though these cases are relatively rare. We will incorporate a brief discussion of these motifs in the revised manuscript, referencing doi.org/10.1038/nature12311, which also provides the evaluation data used in our study.

We have additionally explored the effects of protein data preprocessing on model performance. Our observations indicate that segmenting the protein structure into domains or removing intrinsically disordered regions (IDRs) can improve generation results in certain cases. However, such preprocessing is only feasible if meta-information about protein-RNA interactions is available, as RNAs may interact with multiple domains or IDRs simultaneously. We will add this information into the revision.

Lastly, regarding BAnG’s broader applicability, we plan to extend its use to other biological tasks, such as protein-peptide interactions. However, this would require substantial additional effort and is beyond the scope of the current paper.

We greatly appreciate the reviewer’s valuable comments, which have helped us improve the clarity of our explanations. We hope the added details and planned additions will address the concerns raised by the reviewer and enhance the overall quality of our paper. As the requested revisions were focused on improving clarity, we hope the reviewer will consider increasing their score. If there are any additional specific changes or analyses you would like us to address, we are more than happy to make further revisions.