scKGOT: Intercellular Signaling Inference with Knowledge Graph Optimal Transport for Single-cell Transcriptomics

1. Limited Clarity on Dynamic Adaptability: While the paper claims that Optimal Transport allows scKGOT to dynamically adapt to different biological conditions, it lacks clear explanations and quantitative examples demonstrating how this dynamic adaptability outperforms static models in specific biological scenarios.
2. Dependence on Complete Pathway Information: scKGOT relies on a comprehensive Ligand-Receptor-Pathway Knowledge Graph (LRP-KG) as a prior, which may not be fully available or accurate for all biological contexts. This dependency raises questions about the model’s robustness and effectiveness in datasets with incomplete or biased pathway annotations.
3. Complexity and Accessibility: The use of Optimal Transport, combined with pathway knowledge integration, makes scKGOT computationally complex and potentially challenging for researchers without advanced computational expertise. This complexity could limit its accessibility and reproducibility.

It is very difficult to understand the rationale and evaluation methods used in this paper. On the one hand, the authors make a good point that pathways in sender and receiver cells are related via causal links (i.e. the pathway in the sender cell creates a ligand that get secreted and physically interacts with a receptor on a receiver cell, which causes a pathway to activate within the receiver cell). On the other hand, the main problem I have is that there is no ground truth for this type of relationship presented, just a general comparison of learned networks and prior networks. Further much of the presented biology makes very little sense. For example, in figure 2d, there is a relationship where "pathways in cancer" generates a large number of ligands, which bind to a large number of receptors in another "pathways in cancer" pathway. This really makes not sense biologically and I cannot understand how it makes sense to predict it as relevant.

Another example "By focusing on the largest connected components, we aimed to gain insights into how different cell types communicate within various biological contexts, contributing to a deeper under- standing of processes such as immune modulation, cell signaling, and tissue remodeling." - why is the largest connected component interesting here? The networks in Figure 3 almost look like a random assortment of pathways and don't provide any cell context information, which seriously undermines their use as a tool to interpret the results of scKGOT.

Multiple statements in the paper demonstrate poor understanding of the biological application area e.g. "2,223,641 and 1,651,421 records of ligand-receptor interaction facts respectively, including binding, dephosphorylation and activation, etc." - these are mostly not direct ligand-receptor relationships, but rather are relationships within pathways.

1- Literature review in the introduction section needs to be improved: For example, the introduction section is very limited in reviewing previous works that utilized prior biological pathway knowledge to enhance model interpretability in the field of omics data analysis. The authors may need to refer to some previous works such as Spatalk [1], EXPORT [2], VEGA[3] that utilize knowledge graphs to enhance model interpretability when analyzing omics data.

2- Mathematical description of paper needs to be significantly improved: Some mathematical notations are not described at all when describing the proposed model. For example, what are s1 and s2 in equation (2)? or a and b in equation (3)? also what's the intuition behind inequalities in (3)? Major revision is needed in mathematical description of the model to help readers better understand the model

3- The results need to be better described:

--For example whats the difference between left and right figures in Figure 1 (top row)? What does different colors mean? Also what does (CTDB) version of benchmarking tools mean in Figure 1 (bottom row) mean? Can authors provide more descriptive caption for figure 1?

--In Figure 2C, why authors have focused on "Eph-Ephrin Signaling" or "Signaling By Notch4" pathways? Do they have any specific characteristics that authors have picked them for visualization? What does each subfigure indicate?

-- In Figure 2D, can authors focus on one specific LRP pair that is previously well studied and describe the model results on that specific LRP pair? Its hard to see what result is novel from scKGOT in Figure 2D.

References: [1] Shao, X., Li, C., Yang, H. et al. Knowledge-graph-based cell-cell communication inference for spatially resolved transcriptomic data with SpaTalk. Nat Commun 13, 4429 (2022). https://doi.org/10.1038/s41467-022-32111-8 [2] Biologically Interpretable VAE with Supervision for Transcriptomics Data Under Ordinal Perturbations Seyednami Niyakan, Byung-Jun Yoon, Xiaoning Qian, Xihaier Luo bioRxiv doi:https://doi.org/10.1101/2024.03.28.587231 [3] Seninge, L., Anastopoulos, I., Ding, H. et al. VEGA is an interpretable generative model for inferring biological network activity in single-cell transcriptomics. Nat Commun 12, 5684 (2021). https://doi.org/10.1038/s41467-021-26017-0

While scKGOT presents an innovative and mathematically sound approach for analyzing cell-cell communication, the manuscript would significantly benefit from more detailed methodological descriptions. The current presentation leaves several critical implementation questions unanswered. Specifically, the paper should elaborate on the preprocessing pipeline for single-cell transcriptomics data, including how it handles common technical challenges like dropout effects, batch variations, and cell type annotation reliability. A crucial aspect that requires more clarity is the construction and quality control of the knowledge graph from multiple pathway databases. While the authors utilize both KEGG and Reactome databases (with over 2 million interaction records), the methodology for resolving redundancies and potentially conflicting information between these databases remains unclear. The integration strategy for different database formats, confidence scores, and annotation systems would be valuable information for readers looking to implement or build upon this approach and can be added as supplementary information. Minor Comments: Writing: There is redundant information in the introduction section (paragraphs 3 and 4 are repetitive) Implementation: Additional details about computational requirements and scalability would be valuable Method availability: Information about code availability and documentation would benefit the community