Proof-of-concept for reasoning over the SemMedDB knowledge base, using miniKanren + heuristics + indexing.

[renesugar]

BIOMEDICAL DATA TRANSLATOR TECHNICAL FEASIBILITY ASSESSMENT OF REASONING TOOL: UNIVERSITY OF ALABAMA AT BIRMINGHAM

Our plan remains largely the same, except that we reserve the right to go beyond standard miniKanren to any logical / probabilistic logical programming language for deep causal reasoning. For example, for less complex queries that need greater performance, we may utilize SPARQL as a back-end. We have also found classical logical reasoning to be more effective than anticipated, so we plan to go as far as possible with classical logic programming before exploiting probabilistic logic programming.

https://github.com/webyrd/mediKanren

https://researchsoftwareinstitute.github.io/data-translator/

https://taggs.hhs.gov/Detail/AwardDetail?arg_AwardNum=OT2TR002517&arg_ProgOfficeCode=264

Summary:

mediKanren can be used to translate patient information and databases of biomedical research data into treatments using drugs that have already been approved for a different purpose.

For example, MOON software from Diploid Genomics takes five minutes to suggest the causal mutation out of the 4.5 million variants in a whole genome.

"Gene sequencing might reveal that a patient’s genetic mutation is causing overproduction of a specific protein, for example."

mediKanren can be used to find any FDA-approved drugs that inhibit that protein.

References:

1) https://www.uab.edu/news/research/item/10382-a-high-speed-dr-house-for-medical-breakthroughs

2) https://www.youtube.com/watch?v=RVDCRlW1f1Y

3) http://minikanren.org/minikanren-and-prolog.html

4) https://www.sciencedaily.com/releases/2019/04/190424153613.htm

5) http://www.radygenomics.org/2019/04/24/rady-childrens-institute-for-genomic-medicine-uses-artificial-intelligence-to-diagnose-genetic-diseases/

[evancofer]

@renesugar Is there a paper for this? If so, can you please include the entire abstract and a doi-based link to the paper? I think this issue is a good example of how to format it.

[renesugar]

Software only at this point.

It is part of the larger "Biomedical Data Translator" project.

[evancofer]

I am wondering how we should standardize the formatting. @cgreene do we have some existing issue examples for non-publications? I do not recall any.

[agitter]

837 is an example issue for a website that (at the time) did not have an associated paper. However, I don't see a major need to define a standard issue format for non-publications. The preference for DOI-based links for papers was to help with Manubot-style citations.

[evancofer]

@agitter That makes sense. @renesugar Could you add some sort of summary of each of the links you've included? The current one is informative for computer scientists, but it might help to clarify the biomedical application and highlight the relevance to deep learning more.

[renesugar]

As the software relates to deep learning, you could characterize the software as tools.

How they could be used is open-ended.

@agitter mentioned MoleculeChef on Twitter.

https://arxiv.org/abs/1906.05221

"Deep generative models are able to suggest new organic molecules by generating strings, trees, and graphs representing their structure."

Once you come up with a molecule that you believe has the properties you need, with the right databases, you could ask mediKanren if there are one or more existing FDA-approved drugs that have similar effects.

Deep learning could be used for gene-disease relation extraction from literature for use in mediKanren.

See "Unsupervised word embeddings capture latent knowledge from materials science literature" as a way of generating new knowledge from biomedical research data and in turn use that for gene-disease relation extraction for use in deep learning software, mediKanren, DeSigN, etc.

Or, using the same patient information and databases of past mediKanren cases where the outcome is known, what would software implemented using deep learning recommend as treatments?

mediKanren could be updated to use gene expression in treatment recommendations like DeSigN.

References:

1) https://techxplore.com/news/2019-07-machine-learning-algorithms-uncover-hidden-scientific.amp

2) https://www.nature.com/articles/s41586-019-1335-8

3) https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6061985/

4) https://sagebionetworks.org/

[evancofer]

@renesugar This is great. Thanks!