Transporter annotations are holding up progress in metabolic modeling

Abstract

Mechanistic, constraint-based models of microbial isolates or communities are a staple in the metabolic analysis toolbox, but predictions about microbe-microbe and microbe-environment interactions are only as good as the accuracy of transporter annotations. A number of hurdles stand in the way of comprehensive functional assignments for membrane transporters. These include general or non-specific substrate assignments, ambiguity in the localization, directionality and reversibility of a transporter, and the many-to-many mapping of substrates, transporters and genes. In this perspective, we summarize progress in both experimental and computational approaches used to determine the function of transporters and consider paths forward that integrate both. Investment in accurate, high-throughput functional characterization is needed to train the next-generation of predictive tools toward genome-scale metabolic network reconstructions that better predict phenotypes and interactions. More reliable predictions in this domain will benefit fields ranging from personalized medicine to metabolic engineering to microbial ecology.

Keywords: flux balance analysis; functional genomics; metabolic modeling; microbial community modeling; transporter annotation.

FIGURE 1

The pitfalls of transporter annotations in community metabolic modeling. (A) Types of errors encountered when assigning a single putative transporter to a single substrate. An annotation may miss an assignment where there should be one, may create an assignment where there should not, or may get the direction(s) of transport wrong (either due to an incorrect orientation of an irreversible process, or due to a reversibility error). (B) Mappings from transporter genes to substrates are non-unique. A single gene may map to a single substrate or multiple substrates, a single gene may be a part of a complex with multiple genes which map to a single substrate or multiple substrates. (C) Microbial interactions are variously affected by transporter annotation errors. For example, a species might not grow with missing assignment errors, the community might accumulate or deplete extracellular metabolites by false assignment errors, or a mutualism might be broken by directionality errors. (D) Analysis of transport mappings in BiGG models (n = 108 models). Histograms showing the proportion of transporter reactions to total reactions (left), the proportion of transporter genes to total genes (second from left), the proportion of one-to-many gene-to-transporter mappings to total transporter genes (second from right), and the proportion of one-to-many exometabolite-to-transporter gene mappings to total exometabolites (right). The large peaks correspond, mostly, to models of Escherichia coli, which are overrepresented in the BiGG database.

FIGURE 2

Summary of transporter annotations retrieved from TransportDB 2.0. (A) Distributions of the proportion of transporters annotated to different levels of specificity across all organisms. Vertical dashed lines correspond to the mean of each distribution, and an example of each category is provided. (B) Distributions of the proportion of transporters of the top 3 most abundant [super-] families across all organisms. ABC–ATP binding cassette; MFS–major facilitator superfamily; PTS–phosphotransfer-driven group translocators.

FIGURE 3

A proposed computational workflow to progressively narrow the search space for experimental validation of transporter functional annotations. Red lines correspond to paths followed for a single transporter and are repeated for all un-annotated transporters, while black lines correspond to paths taken (once) for the whole genome. The pipeline begins (1) with alignment of transporter genes to the TCDB, retrieving a list (horizonal bars) of all children metabolites associated with the lowest common ancestor ontology term. In another path (2), a draft GEM is reconstructed to generate a list of all intracellular metabolites synthesized or degraded in the metabolic network. The intersection of both lists (cyan bars) is passed to a third path (3) as candidates for docking simulations using the predicted protein structure. Predicted binding affinities that exceed some threshold are finally passed as candidates for experimental validation.