Potential Reorganization

[manolik]

Hi all, This is such an awesome review, and i'd love to use it as a basis for teaching, but i've had to jump back-and-forth through sections, as the current structure doesn't flow as naturally for a class. I had some thoughts about potentially restructuring and wanted to get folks' feedback on this. What do you think? https://www.dropbox.com/s/guv6eovze8e1s6g/Deep%20Learning%20Review%20Reorganization.xlsx?dl=0 Thanks! Manolis

[agitter]

@manolik this is great, thanks for the suggestions. I see many benefits to a major reorganization:

• We can create a more logical organization around the content rather than the existing sequence of sections, which was a product of our initial categorize/study/treat framework and the order in which sections were contributed
• It will make version 2.0 more of a standalone product and give us a better reason to go through the effort of updating the review (book?)
• It can engage new contributors

The only real downside is that it will take more effort to write and especially to edit. I'm okay with that. If we put out a call for help with specific pieces, we can attract domain experts to contribute.

If others generally think this is a good idea, we can create add a new outline to the repository starting from the spreadsheet above.

[akundaje]

I support the idea of a re-org as well

[cgreene]

I also strongly support the idea of reorganization. I wonder if it would make sense to have a Imaging, Text, and EHR (perhaps in that order?) in a module. EHRs are hot as an area, but it might help to lead off with how folks are handling many of the components of them before jumping into them.

I made @manolik's into a markdown version of the table.

Introduction	-1		DL's role outside of biology
			How this document was composed (online, open, living) & context for those who read prior version
Module0	0	Foundations	ML and Deep Learning
			CNNs, RNNs, Model Architectures
			Training, Regularization
			Dimensionality reduction
			Autoencoders, Generative Models, GANs
Module1	1	Pre-Tx	DNA accessibility
			Promoters and enhancers
			Transcription factors
			DNA methylation
	2	Expression	Gene expression
			Single-cell data
	3	Post-Tx	Splicing
			Micro-RNA binding
Module2	4	Variation	Sequencing and variant calling
			Comparative genomics
			Alignment, Evolution
	5	Genetics	Population genetics, phasing
			eQTLs
			Genetics, GWAS, Rare variants
	6	EHRs	Patient categorization
			Electronic health records
	7	Specialized	Neuroscience
			Metagenomics
			Cancer
Module3	8	PPIs	Protein-protein interactions
			PPI networks and graph analysis
	9	Protein Structure	Protein secondary and tertiary structure
			Structure determination and cryoEM
	10	3D structure	MHC-peptide binding
	11	Therapeutics	Drug development
			Drug repositioning
Module4	12	Imaging	Morphological phenotypes
			Imaging applications in healthcare
			Video processing and analysis
	13	Text	Text applications in healthcare
			Clinical decision making
			Interactive question/answer systems
Module5	14	Models	Customizing: bias-vs-variance
			Uncertainty quantification
			Multimodal, multi-task, and transfer learning
			Interpretation
	15	Practical issues	Optimization
			Data limitations
			Hardware limitations and scaling
			Data, code, and model sharing
Module6	16	Conclusions	Summarize how field has changed since prior review (mid 2017)
	17	Methods	Summarize number of contributors, number of commits, use of manubot, etc. Primarily for others undertaking similar efforts.

I added a brief intro / conclusions section. While I could see the adjustments I proposed, I think they are small when compared to the scale of the reorganization. I like where this is going and strongly support this concept! 👏

[manolik]

Yes, greatly agree.

1. about intro on "DL beyond bio". To avoid distracting from the main message, I would perhaps put this as a section at the end of the "Module0.Foundations" module, after introducing the core methods, which i think are more relevant.

2. About how the article was written, i would add that to the beginning of Module6-section 17, as folks reading this article/book will care more about the content than the (very awesomely cool) way it was written

[evancofer]

I agree. Reorganizing things in this manner looks like a good idea.

[SystemsResearch]

This structure looks good. I am just wondering if sections 6 and 7 would be fit better later on. The progression could be:

• module 0 as above
• module 1 (Tx/expression): sections 1, 2, 3
• module 2 (genetics/genomics): sections 4, 5
• module 3 (proteins/drugs): sections 8, 9, 10, 11
• module 4 (phenotypes/patients): sections 12, 13, 6, 7
• module 5 as above
• module 6 as above

[cgreene]

I do think imaging, text, and EHRs make a lot of sense together. Perhaps what's currently "specialized" could have it's own module just before Module 5. These sections would likely be shorter, but that's ok. Sleep, if I understand correctly, might go into neuroscience there for instance.