Robust estimation of cancer and immune cell-type proportions from bulk tumor ATAC-Seq data

Reviewer #1 (Public Review):

Summary:

Building upon their famous tool for the deconvolution of human transcriptomics data (EPIC), Gabriel et al. implemented a new methodology for the quantification of the cellular composition of samples profiled with Assay for Transposase-Accessible Chromatin sequencing (ATAC-Seq). To build a signature for ATAC-seq deconvolution, they first created a compendium of ATAC-seq data and derived chromatin accessibility marker peaks and reference profiles for 21 cell types, encompassing immune cells, endothelial cells, and fibroblasts. They then coupled this novel signature with the EPIC deconvolution framework based on constrained least-square regression to derive a dedicated tool called EPIC-ATAC. The method was then assessed using real and pseudo-bulk RNA-seq data from human peripheral blood mononuclear cells (PBMC) and, finally, applied to ATAC-seq data from breast cancer tumors to show it accurately quantifies their immune contexture.

Strengths:

Overall, the work is of very high quality. The proposed tool is timely; its implementation, characterization, and validation are based on rigorous methodologies and resulted in robust results. The newly-generated, validation data and the code are publicly available and well-documented. Therefore, I believe this work and the associated resources will greatly benefit the scientific community.

Weaknesses:

A few aspects can be improved to clarify the value and applicability of the EPIC-ATAC and the transparency of the benchmarking analysis.

Most of the validation results in the main text assess the methods on all cell types together, by showing the correlation, RMSE, and scatterplots of the estimated vs. true cell fractions. This approach is valuable for showing the overall method performance and for detecting systematic biases and noisy estimates. However, it provides very limited insights regarding the capability of the methods to estimate the individual cell types, which is the ultimate aim of deconvolution analysis. This limitation is exacerbated for rare cell types, which could even have a negative correlation with the ground truth fractions, but not weigh much on the overall RMSE and correlation. I would suggest integrating into the main text and figures an in-depth assessment of the individual cell types. In particular, it should be shown and discussed which cell types can be accurately quantified and which ones are less reliable.

In the benchmarking analysis, EPIC-ATAC is compared to several deconvolution methods, most of which were originally developed for transcriptomics data. This comparison is not completely fair unless their peculiarities and the limitations of tweaking them to work with ATAC-seq data are discussed. For instance, some methods (including the original EPIC) correct for cell-type-specific mRNA bias, which is not present in ATAC-seq data and might, thus, result in systematic errors.

On a similar note, it could be made more explicit which adaptations were introduced in EPIC, besides the ad-hoc ATAC-seq signature, to make it applicable to this type of data.

Given that the final applicability of EPIC-ATAC is on real bulk RNA-seq data, whose characteristics might not be completely recapitulated by pseudo-bulk samples, it would be interesting to see EPIC and EPIC-ATAC compared on a dataset with matched, real bulk RNA-seq and ATAC-seq, respectively. It would nicely complement the analysis of Figure 7 and could be used to dissect the commonalities and peculiarities of these two approaches.

Reviewer #2 (Public Review):

Summary:

The manuscript expands the current bulk sequencing data deconvolution toolkit to include ATAC-seq. The EPIC-ATAC tool successfully predicts accurate proportions of immune cells in bulk tumour samples and EPIC-ATAC seems to perform well in benchmarking analyses. The authors achieve their aim of developing a new bulk ATAC-seq deconvolution tool.

Strengths:

The manuscript describes simple and understandable experiments to demonstrate the accuracy of EPIC-ATAC. They have also been incredibly thorough with their reference dataset collections. The authors have been robust in their benchmarking endeavours and measured EPIC-ATAC against multiple datasets and tools.

Weaknesses:

Currently, the tool has a narrow applicability in that it estimates the percentage of immune cells in a bulk ATAC-seq experiment.

Comments:

(1) Has any benchmarking been done on the runtime of the tool? Although EPIC-ATAC seems to "win" in benchmarking metrics, sometimes the differences are quite small. If EPIC-ATAC takes forever to run, compared to another tool that is a lot quicker, might some people prefer to sacrifice 0.01 in correlation for a quicker running tool?

(2) In Figure 3B the data points look a bit squashed in the bottom-left corner. Could the plot be replotted with the data point spread out? There also seems to be some inter-patient variability. Could the authors comment on that?

(3) Could the authors comment on the possibility of expanding EPIC-ATAC into more than a percentage prediction tool? Perhaps EPIC-ATAC could remove the immune cell signal from the bulk ATAC-seq data to "purify" the uncharacterised cells in silico, or generate pseudo-ATAC-seq tracks of the identified cell types.