Given the well-recognized variability in data reliability, ranging from experimentally validated interactions to in silico predictions, could the authors clarify how data quality, uncertainty, or source confidence is quantitatively factored into the network construction and subsequent analyses? For example, in top-down approaches where comprehensive interactomes are filtered, is there a risk of amplifying noise or bias due to inconsistent annotation standards across databases (as mentioned with “YES1” gene identifiers)? How do the proposed tools or frameworks account for this, especially when drawing clinical or pharmacological conclusions? Additionally, since machine learning models, particularly GNNs, are increasingly used to augment or infer network connections, how do the authors address the explainability and generalizability of predictions made on incomplete or noisy biological data? Are there benchmark validation strategies that could strengthen the trust in such models beyond retrospective validations?
I also have a question that’s been on my mind after reading the paper.
The authors do a great job discussing link prediction using GNNs and the integration of multi-source biological data. But I was wondering: how do the authors handle the challenge of conflicting or context-specific interactions in large-scale network integration, especially when merging datasets from different cell types, tissues, or disease states?
For example, a protein–protein interaction might be valid in a hepatic context but not in neural tissue — yet it could still be retained in a global interactome. In such cases, how are tissue-specificity or context-awareness factored into network synthesis and downstream predictions like drug-target inference or essential gene identification? And would incorporating spatiotemporal expression data or cell-type–specific interaction layers improve the accuracy of network-driven conclusions?