A versatile informative diffusion model for single-cell ATAC-seq data generation and analysis

Thank you very much for your detailed and insightful comments on our manuscript. Your comments clearly helped a lot to improve this manuscript! We have summarized your comments and made point-by-point responses and revisions to address your concerns.

1. Thank you for your valuable comment. Our model outperforms the baseline models on both the analysis and generation task. The baseline models are only designed for the scATAC-seq data analysis by leveraging the approximate distribution like ZINB as the posterior. In contrast, our model incorporate both the latent distribution construction and data likelihood learning. These two strategies can coordinate with each other, leading to the improved performance on both sides. The low dimensional cell embeddings could enhance the diffusion model to capture intrinsic high-level factors of variation present in the heterogeneous scATAC-seq data. Furthermore, the mutual information between the cell embeddings and real data points enable ATAC-Diff to avoid ignoring the latent variables as the conditional information when utilizing ELBO as the objectivity compared to the baseline models which are based VAE framework (SCALE, PeakVI).

2. Thank you for your considerable comment. Actually, the Hematopoiesis dataset is the Buenrostro2018 dataset. We utilize Hematopoiesis to indicate the biological process insetad of Buenrostro2018 dataset. You could check the reference [1]. Jason D Buenrostro, M Ryan Corces, Caleb A Lareau, Beijing Wu, Alicia N Schep, Martin J Aryee, Ravindra Majeti, Howard Y Chang, and William J Greenleaf. Integrated single-cell analysis maps the continuous regulatory landscape of human hematopoietic differentiation. Cell, 173(6):1535–1548, 2018.

3. Thank you very much for your valuable comments. We did not compare our model with scBasset since it is only designed for single cell data analysis without any generation ability. Besides, scBasset requires the DNA sequences as the input. However, both Forebrain dataset and PBMC10K dataset do not provide the genome information. In order to further address your concern, we have reported the results of scBasset on the Hematopoiesis (Buenrostro2018) dataset.

While scBasset may outperform our model by leveraging supplementary genomic data for predicting chromatin accessibility, it faces limitations in capturing the distribution of single-cell data due to its reliance on restricted genomic information. Moreover, its deterministic nature restricts its utility in generating scATAC data.

Thank you very much for your detailed and insightful comments on our manuscript. Your comments clearly helped a lot to improve this manuscript!

1. We clarify that in our approach, we utilize fragment counts rather than binary peaks to represent the scATAC-seq data. To alleviate any confusion, we have updated the descriptions related to the data for better clarity.

"Specifically, we use fragment counts to represent the scATAC-seq data.

"We compress the scATAC-seq data (fragment counts) into a lower-dimensional latent space."

2. Our model is constructed solely based on the scATAC-seq data, without incorporating DNA sequence information. Given that our model operates as a conditional generative model, we intend to integrate DNA sequence information as a guiding conditional factor in our future work. In this version, we did not use DNA sequence since some datasets do not include the genome information.

3. Ground truth cell annotations utilized in our study are extracted from the previous publication [1,2,3]. These annotations were extensively characterized through marker peaks corresponding to marker genes, supported by their distinct biological functions enriched with cell-type-specific peaks. We affirm that these cell annotations are not only reliable but also hold significant biological relevance.

The metrics like NMI, ARI, and ASW, which were employed in our evaluation, are commonly utilized to assess the conservation of biological variance in latent features for single-cell dataset benchmarks. This practice is well-documented in [https://www.nature.com/articles/s41592-021-01336-8] and several other prominent single-cell methodologies [4,5,6].

We acknowledge the concern that UMAP visualization can sometimes warp the space, particularly in regions with large distant gaps. Nevertheless, UMAP generally preserves distance relationships within closely-knit local subgroups that share similarities. For instance, subpopulations like EC1, EC2, and EC3, which belong to excitatory cells and are in close proximity, exhibit this preservation, aligning well with their biological characteristics.

Moreover, we have computed the average latent embeddings within each cell type population and calculated the distances across different cell types. We have listed some distances due to the space limitation: EX1-EX2: 0.477, AC-EX1: 0.831. We have added other values in the Appendix.

4.1 We employ unconditional generation to assess the model's ability to capture the entirety of the data distribution. Effective modeling of the data distribution enables us to generate synthetic data for data augmentation, potentially obviating the need for sequencing additional cells and thereby saving valuable time and resources.

To further address your concerns, we have refined Section 4.4.1 to provide a more detailed explanation of the unconditional task

4.2 For the cluster task, we utilize the NMI, ARI, Homo, and ASW to evaluate the performance. The calculation of metrics are formulated as follows.

$NMI(A,B)=\frac{I(A,B)}{\sqrt{H(A)H(B)}}$

$ARI=\frac{RI-\text{Expected}_RI}{\max{(RI)}-\text{Expected}_RI}$

$Homo = 1 - \frac{H(C, Y)}{H(Y)}$

$ASW = \frac{1}{N} \sum_{i=1}^{N} \frac{b_i - a_i}{\max(a_i, b_i)}$

We adopt the average of all scATAC-seq data as the ground truth to calculate the SCC and PCC for the unconditional generation task. For the conditional generation, we average all single cells of the same biological cell type as the ground truth. For the denoising task, it is the same with conditional generation. For the imputation task, we calculate the SCC and PCC of the masked value and the imputed value.

4.3 We cannot calculate the correlation of the generated data and ground truth at the single cell level as the data is not one to one generated. The scATAC-seq technology suffers from many sources of technical noise, leading to dropout events, which means even the ground truth scATAC-seq data is also not complete.

4.4 The SCC and PCC of the conditional generation are lower than the unconditional generation since we average the SCC and PCC of different cell types for conditional generation while we calculate the SCC and PCC of the whole data for conditional generation. There are some cell types which are hard to learn due to limited data. Hence, the PCC and SCC of such cell types are quite low, leading to the lower averaged SCC and PCC.

4.5 We have conducted the ablation study.

Thank you very much for your encouraging comments and the constructive advices on improving the manuscript. Your comments clearly helped a lot to improve the study. We have summarized the suggested comments (Questions) and made point-by-point responses and revisions as follows.

1. Thank you for your valuable feedback. $q_\phi(z|x_0)$ represents the amortized inference distribution, serving as an approximate variational posterior within the generative model. To achieve this parameterization, we employ an auxiliary encoder which is similar to the encoder of VAE model.
2. Thank you for your insightful comments. In our study, $y$ represents the conditional information, encompassing factors like cell type, tissue, and other omics data (e.g., scRNA-seq data). We have chosen to utilize cell type as the guiding conditional information for the evaluation of our model. To address your feedback, we have revised the description of $y as follows:

"We conditioned the diffusion models on the latent 165 variables z and other conditional information y such as cell types, tissue, and other omics data (e.g. scRNA-seq data)."

3. Thanks for your valuable comments. We have incorporated a figure illustrating the dimensions of the tensors within the network. We will put this figure in the global response.
4. We sincerely appreciate your valuable insights. In our methodology, we utilize an exponential distribution where peaks exhibiting lower expression levels are more prone to dropout events compared to those with higher expression levels. This differential dropout mechanism is employed to simulate realistic dropout events in our data. We hypothesize that peaks with higher expression levels are more likely to go undetected during the sequencing process. Consequently, we ensure that a minimum of 80% of values in the simulated dataset are zero, reflecting the sparsity inherent in single-cell sequencing data. This intentional introduction of sparse noise effectively mirrors the prevalent characteristics found in actual single-cell sequencing datasets, which are often characterized by a significant number of 'missing' or 'zero' values."

Dear Authors, You are right. Most likely your rebuttal is sufficient to resolve any unclarities at this point, but if there are any further questions, they'll hopefully be posted asap so that you may still react. thanks for your patience, Area Chair