agitter
commented
7 years ago Thanks for the comment @blengerich. Structured data should definitely be discussed somewhere but could fit in a number of places. In my opinion, the decision to incorporate it in specific sections versus the general discussion section might depend on whether we want to cover what has been done in existing papers or propose potential future opportunities.. Some of the examples of structured data that come to mind in papers we've discussed include:
- Sequence (DNA, RNA, amino acid, etc.)
- Temporal data
- Images
- Volumes
- Graphs, specifically molecular graphs
Are there other types of biomedical data that haven't been paired with appropriate network architectures yet, such as the regulatory networks you mentioned? Are there other neural network architectures besides CNNs and RNNs that are well-suited for structured data but haven't been applied in biomedicine yet? More recent and general graph convolutions (e.g. https://tkipf.github.io/graph-convolutional-networks/) could be of interest. Geometric deep learning: going beyond Euclidean data could also also spawn some good discussion.
Structured data arises in a few of the domains discussed throughout the review. For example, CNNs perform tremendously for many tasks based on data from a regular graph like images (subsection here. More recently, CNNs have been developed for computations on arbitrary graphs, specifically for cheminformatics and molecular fingerprinting (#52, #53). Presumably, graph convolutions or a similar technique would be useful in other domains which observe a signal on a graph (gene expression levels on a regulatory network, etc.). So advances in deep learning on structured data may open new domains for convolutional architectures.
Is this topic general enough to be worth mentioning in the discussion section?