gregmedlock / ssl_ensembles Goto Github PK

This repository holds all code and data for "Guiding the refinement of biochemical knowledgebases with ensembles of metabolic networks and machine learning", Medlock & Papin, bioRxiv 2018, currently in press at Cell Systems. The manuscript and the accompanying code describe Automated Metabolic Model Ensemble-Driven Elimination of Uncertainty with Statistical learning (AMMEDEUS). Description of all analyses contained in this repository are expanded upon in the manuscript.

Our goal was to develop a system for guiding the curation of genome-scale metabolic network reconstructions (GENREs) by harnessing the variability across ensembles of GENREs. We generated ensembles of GENREs consistent with experimental data, performed gene essentiality simulations with those ensembles, and applied machine learning to identify differences between networks that are associated with differences in simulated gene essentiality. We applied this approach to 29 bacterial species with suitable genomic and phenotypic data available, generating prioritized curation targets for all species. We summarize these findings across all 29 species to identify reactions and subsystems in the biochemical database used in the study that, if curated, would substantially reduce uncertainty in gene essentiality simulations for many organisms.

Data links: [1] https://github.com/ModelSEED/ModelSEEDDatabase/blob/dev/Biochemistry/Pathways/ModelSEED_Subsystems.tsv [2] https://github.com/ModelSEED/ModelSEEDDatabase/blob/dev/Media/media_list_with_meta.txt

• deep_gapfill.ipynb has a loop that iterates over each species to generate an ensemble. For each species, this process took 30-120 minutes in my experience. As a result, notebook failures would often require restarting the process, so the code block checks for existing ensembles for each species. If fast generation of all ensembles is necessary, use of a cluster and reworking the notebook to act on individual species, rather than iterating over all species, is recommended. We opted not to take this approach to simplify the codebase and because I had access to reasonably hefty computing power.
• generate_recons.ipynb requires that a user creates a PATRIC account to access the ModelSEED resource via Mackinac. You can create an account on the PATRIC site (https://www.patricbrc.org) but we recommend following whatever the current instructions are for creating a PATRIC account from the ModelSEED website (http://modelseed.org). When executing the notebook, you must replace my account name in the mackinac.get_token() function with your own account name, and then you will be prompted to enter your password within the notebook. This cannot be automated, to my knowledge.
• Draft models generated with generate_recons.ipynb contain gapfilled reactions from ModelSEED during their complete media gapfilling step. During this step, flux through biomass is constrained to be positive and exchange reactions for any metabolite for which a transporter is annotated is enabled; reactions required for a feasible solution are added to the network. This process is sensitive to the version of the ModelSEED biochemistry used, thus the solutions returned maay change over time and be substantially different from the models I obtained for this manuscript.
• I attempted to generate figures as final as possible in these notebooks; multiple figures were assembled with Adobe Illustrator in the final publication. Chemical structures were drawn in chemdraw.

cobrapy >= 0.13 mackinac (devel branch at the time -- see commit 83eced233406c27ed0d749e6d9f76d41350e3b09 in https://github.com/gregmedlock/mackinac) medusa (devel branch -- see commit e7c7eb88aea0b4dd8d08b996c441ab5f1758cd07 in https://github.com/gregmedlock/Medusa)