Skip to main content
. 2022 May 4;13:2455. doi: 10.1038/s41467-022-29867-4

Fig. 3. HMO biosynthesis models constructed with flux analysis and ranked by expression concordance perform consistently across cohorts. Overview of computational methods for model assembly (A–F) and assessment (G–I).

Fig. 3

A To build the candidate models of HMO biosynthesis, reaction rules were defined to specify all possible monosaccharide additions. B The Complete Network includes all oligosaccharides and irreversible reactions resulting from the iterative addition of monosaccharides to a root lactose. C Using Flux Variability Analysis (see Supplementary Methods 5.4), the Complete Network was trimmed, removing reactions that cannot reach experimentally-measured HMOs, to produce a Reduced Network D (Supplementary Fig. 10); red triangles are observed HMOs blue lines are “sink reactions” joining alternative isomers (Supplementary Fig. 5). E From the Reduced Network, Mixed Integer Linear Programming (MILP) was used to extract candidate models, each representing a subnetwork capable of uniquely synthesizing the observed oligosaccharide profile using a minimal number of reactions; black lines are reactions retained in a candidate model. F Flux Balance Analysis estimated flux through each reaction necessary to simulate the measured relative oligosaccharide abundance (Supplementary methods 5.4). G Model scores were computed as the average maximum correlation between linkage-specific candidate genes and normalized flux through that linkage (Supplementary Fig. 7 and Supplementary Methods 4.4). H Model scores were parameterized on cohort 1 (left) and cohort 2 (right) data (see Methods section). High-performing models, 95th percentile of scores, are highlighted in red. I Of the >40 million models considered (blue), 2.66 and 2.32 million models were high-performing when parameterized on data from cohort 1 or cohort 2, respectively. Nearly 250,000 models consistently explained the relationship between predicted flux and expression data from both cohort 1 and cohort 2. These commonly selected models were analyzed for common structural features.