Overall Manuscript Structure

[cgreene]

The overall aims of the Headline Review articles are outlined in the README. Here's a document structure that I am playing around with to target the review at this question: What would need to be true for deep learning to transform how we categorize, study, and treat individuals to maintain or restore health?

• Relevant areas where methods inspired by deep learning are already having an impact
  • Sequence -> Function
  • Transcriptional regulation
  • Patient information
  • Imaging + Bio
• The structures of problem statements which use deep learning towards these ends
  • Supervised approaches
    • Convolutional NNs on genome
    • Principles in which multiple synergistic patterns are learned simultaneously
    • more examples of shared properties across approaches
  • Unsupervised approaches
    • some denoising autoencoder work is common across systems
    • more shared properties
• Perspectives towards the future & overall question.
  • Which challenges do we think will be resolved first?
  • Are there any approaches/data types that have taken off in other fields but that are under-utilized here?
  • What initiatives or data do we think are particularly interesting for/amenable to deep learning analyses and why?
• Overall summary on state of the field & reflection towards overall question.

There are some wonderful github-based reading groups/lists by @pimentel @hussius @gokceneraslan. If any of you have feedback as we structure this review, please provide it. If you'd like to participate - dive in!

[cgreene]

Also wanted to tag @YosephBarash and @davek44 who have been active in this area for their thoughts.

[michaelmhoffman]

What is the format of the text itself? GitHub markdown?

[michaelmhoffman]

Does this need to actually have anything to do with "systems pharmacology"?

[cgreene]

@michaelmhoffman It should probably roughly touch on topics that could be construed as systems pharmacology. My read is that the precision medicine perspective + deep learning on genomic/transcriptomic/proteomic/etc data gets us close enough.

Format of the text itself will be markdown [eventually I'll convert it to LaTeX and reformat]. I think we will use something like [@doi:doi_link] for citations. @dhimmel has code to automatically pull down doi metadata and covert to bibtex.

[agitter]

Are there any approaches/data types that have taken off in other fields but that are under-utilized here?

Reinforcement learning perhaps? http://karpathy.github.io/2016/05/31/rl/ gives a brief intro. I'm not aware of examples in biology or medicine.

[michaelmhoffman]

Reinforcement learning hooked up to some experimental system would be fun.

[dhimmel]

@dhimmel has code to automatically pull down doi metadata and covert to bibtex.

@cgreene, the code is here. Let me know when formatting time comes and I can help with the auto-conversion of citations.

[agitter]

@dhimmel that looks great. May want to pair it with Arxiv2bib at formatting time.

[dhimmel]

@agitter thanks! How about the following citation conventions:

1. Always use a DOI for the version of record if available. Cite DOIs using [@doi:10.15363/thinklab.4]
2. If no DOI for the version of record, use a PubMed ID. For example, [@pmid:26158728].
3. If the article is an arxiv preprint, use [@arxiv:1508.06576].
4. If the article has none of the above, big problem. File an issue.

You can do multiple citations using: [@doi:10.15363/thinklab.4 @pmid:26158728 @arxiv:1508.06576]

[cgreene]

@dhimmel proposed citation conventions look good to me. Do you want to file a PR to add it to the contribution instructions?

[agitter]

@cgreene there are a lot of Imaging + Bio deep learning papers out there. Should we take a more targeted approach for logging them as issues, such as focusing on those that pertain to human disease and medicine, instead of trying to catalog everything? What might be the main points of this subsection?

[cgreene]

@agitter : I guess I'd say, if one could make an argument that it's relevant to our current guiding question [which I think still needs a bit of refinement - but probably an increase in specificity, not a decrease] then those are the ones for which we should file an issue.

[cgreene]

Going to close this now that the discussion has been captured in subsequent issues.

[agitter]

@dhimmel This issue was closed, but I wanted to ask a follow up question now that we're starting to write. The citation conventions above will be great for making the citations machine-readable for automated bibliography construction. Do you have any ideas for how to make them human-readable as well? For example, in latex/bibtex I might use \cite{Zhou2015_deep_sea} so that anyone reviewing my text knows which paper I'm discussing. That will be harder to do when reviewing DOIs and PMIDs.

[dhimmel]

Do you have any ideas for how to make them human-readable as well?

DOIs have some semantic meaning (often contain a journal abbreviation). But we could define another category such as [tag:Zhou2015_deep_sea] and then you'd also have to update a mapping file where Zhou2015_deep_sea would point to a valid machine-readable citation. Do you think that is a good solution?

[agitter]

@dhimmel yes, I think that would be the best solution for machine- and human-readable citations. It adds overhead for the authors so we'll have to weigh those tradeoffs.

@cgreene do you have an opinion about creating a citation mapping file?

[cgreene]

I defer to you and Daniel.

On Thu, Oct 27, 2016, 8:14 AM Anthony Gitter [email protected]
wrote:

@dhimmel https://github.com/dhimmel yes, I think that would be the best
solution for machine- and human-readable citations. It adds overhead for
the authors so we'll have to weigh those tradeoffs.

@cgreene https://github.com/cgreene do you have an opinion about
creating a citation mapping file?

—
You are receiving this because you were mentioned.

Reply to this email directly, view it on GitHub
#2 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAhHsyxt2LfDNMD2AVezGXxetFqEDvd8ks5q4JWkgaJpZM4Jb6zp
.