Search

Additional file of Automatic reconstruction of metabolic pathways from identified biosynthetic gene clusters

Additional file 2. Detailed comparison of 8 BGCs for evaluation the accuracy of the BiGMeC pipeline.

Summary: Metabolism is the set of all chemical reactions responsible for the conversion of nutrients into the energy and cellular building blocks required for growth and cellular maintenance in a living organism. Because of our detailed knowledge of enzymes and the chemical reactions they catalyze, one can create a rather accurate representation of an organism’s metabolic network from the sequenced and annotated genome. However, in contrast to classical textbook depictions of individual metabolic pathways, these metabolic networks are often highly interconnected and can contain thousands of different reactions and metabolites. Due to this complexity, computational and mathematical algorithms are often required to predict the phenotypic outcome of genetic modifications or changes in the nutrient environment. When a metabolic network is combined with a representation of growth, cellular maintenance requirements, and available nutrients, it is called a genome-scale metabolic model. In this work we assemble and apply genome-scale metabolic models to study two rather different organisms. The first organism, Streptomyces coelicolor, is a complex, soil-dwelling bacterium that is of great interest within drug discovery as a cell factory for production of novel biopharmaceuticals. Through two consecutive publications we merge and improve existing S. coelicolor models into a consensus model that is hosted in an open-source environment to encourage contributions from the Streptomyces research community. We then apply the developed model to explore and understand how one should proceed with strain development to create a mutant strain that is optimal for heterologous expression of biosynthetic gene clusters. Another contribution in this direction is our development of a computational pipeline that automatically reconstructs metabolic pathways encoded by biosynthetic gene clusters. The second organism, Prochlorococcus, is the most abundant phototrophic marine bacterium, and thus a major player in the marine food web and global carbon fixation. We use random sampling and dynamic flux balance analysis to understand how its metabolism is affected by the day-night cycle and varying nutrient conditions, with a particular focus on glycogen allocation and release of organic compounds that become nutrients for marine heterotrophs. Furthermore, this study required method development extending the software COMETS to account for the periodicity of available daylight and light absorption. Together, this work contributes to an increased understanding of S. coelicolor and Prochlorococcus, in addition to updated and improved genome-scale metabolic models which are by themselves valuable tools in further research of these bacteria. Additionally, we have developed generic tools of great value for a broader audience, both towards drug development and for future studies of photosynthetic microbes. Sammendrag: Næringsstoffer omdannes til byggeklosser og energi i en bakterie (eller celle) ved hjelp av en mengde kjemiske reaksjoner som til sammen utgjør bakteriens metabolisme. De kjemiske reaksjonene katalyseres av enzymer, og metabolismen til hver enkelt art defineres av genene i bakteriens arvemateriale, og hvilke enzymer genene koder for. Med utgangspunkt i arvematerialet kan man derfor skissere opp modeller av hver arts spesifikke metabolisme som man kan bruke til f.eks. å forutsi hvordan en bakterie vil oppføre seg i et gitt vekstmiljø. I denne doktorgraden har jeg laget og anvendt metabolske modeller for å studere to ulike bakterier. Den første bakterien, Streptomyces coelicolor, tilhører en slekt som er opphavet til mange av antibiotikaene som brukes i dag. På grunn av den økende graden av antibiotikaresistente sykdomsfremkallende bakterier er det et akutt behov for å finne og produsere nye virkestoffer. Her kan S. coelicolor spille en viktig rolle ved å fungere som en mikrobiell fabrikk som kan uttrykke genmateriale overført fra andre bakterier som koder for produksjonen av nye virkestoffer. Ved å kombinere en metabolsk modell med eksperimentelle data har jeg undersøkt hvilke tiltak som kan optimalisere produksjon av nye virkestoffer i S. coelicolor. Jeg har brukt en lignende fremgangsmåte for å forstå hvordan ulike miljøer påvirker Prochlrococcus' evne til å lagre eller skille ut organisk materiale. Som den mest tallrike fotosyntetiske bakterien i havet er Prochlorococcus en hjørnestein i den marine næringskjeden, og det organiske materialet den produserer er viktige næringsstoffer for andre bakterier som ikke har evnen til å gjøre fotosyntese. Ved å simulere tusenvis av ulike næringsmiljøer har vi identifisert hvilke faktorer som har størst innvirkning på denne bakteriens metabolisme. Til sammen har dette arbeidet bidratt til utviklingen av kunnskap og modelleringsverktøy for to bakterier som er av stor interesse innenfor hvert sitt anvendelsesområde.

Additional file 1. Details on tailoring reactions and synthesis of the rare extender unit methoxymalonyl-ACP, as well as a description of the analysis used to develop the heuristics that indicate the presence of these reactions from smCOG annotations.

Many biosynthetic gene clusters (BGCs) in the genomes of environmental microorganisms require heterologous expression in order to realize their genetic potential, including cryptic and metagenomic BGCs. Streptomyces coelicolor M1152 is a widely used host strain for the heterologous expression of BGCs, as it has been genetically engineered for this purpose via the deletion of four of its native biosynthetic gene clusters (BGCs) and the introduction of a point mutation in the rpoB gene that encodes the beta subunit of RNA polymerase. This latter mutation was shown to have a strong positive impact on antibiotic biosynthesis via processes that remain poorly understood. Therefore, a systemic understanding of the consequences on cellular metabolism of the genomic changes of M1152 could greatly contribute to this understanding. Here we carried out a comparative analysis of M1152 and its ancestor strain M145, connecting observed phenotypic differences to changes in transcript and protein abundance. Measured protein abundance was used to constrain an amended genome-scale model (GEM) and to predict metabolic fluxes. This approach connects observed differences in growth rate and glucose consumption to changes in central carbon metabolism, accompanied by differential expression of important regulons. Our results suggest that precursor availability is not limiting the biosynthesis of secondary metabolites. This implies that alternative strategies could be beneficial for further development of S. coelicolor for heterologous production of novel compounds. Importance This study provides the first systems description of S. coelicolor M1152, an engineered host widely used for the heterologous expression of BGCs directing the synthesis of natural products. By combining time-series proteomics and transcriptomics, batch fermentation data and genome-scale modelling, we can connect observed phenotypes to known genetic modifications and find extensive metabolic rewiring in the M1152 strain compared to the wild-type stain M145. Our study indicates that the deletion of secondary metabolite biosynthetic pathways thought to enhance precursor availability, only has a minor impact on the ability of the modified strain to produced heterologous molecules. In contrast, the rpoB mutation is likely responsible for the most dramatic changes in regulatory features and precursor availability. The amended genome-scale model, reconstructed in an open-science framework, allowed us to contextualize the transcriptional changes. This framework facilitates further development by the research community in an organized manner, including version control, continuous integration and quality control and tracking of individual contributions.

Ultrasound mediated drug delivery is an important tool in the fight against cancer. A new concept called Acoustic Cluster Therapy (ACT ) is under development, and two pilot imaging studies have been performed on prostate cancer xenografts in mice. A large amount of raw ultrasound data has been recorded, but existing software can not perform the required image processing. The ACT concept is based on clusters of microbubbles and microdroplets. When exposed to diagnostic ultrasound, the microdroplets become microbubbles. This phase-shift from liquid to gas is followed by a microbubble growth to 30μm. These phase-shift bubbles get stuck in the small capillaries of the tumor vasculature. A complete program has been developed in MATLAB® to process the raw ultrasound data. The program is tailored to the unique properties of the phase-shift bubbles, and is able to reduce noise and motion artefacts, to visualize the contrast agent, and to count the number of ultrasound activated phase-shift bubbles. The program produces high quality videos, displaying both free flowing contrast agent and identified, stuck phase-shift bubbles. The program was validated against a synthesized data set, and we found that the program counted accurately up to 2 bubbles/mm2. A saturation was experienced above this threshold, and too few bubbles were counted. The program was applied to a data set of 16 tumors, divided into four groups based on different ACT cluster dose and activation ultrasound settings. A significant difference (p = 0.023) was found between the different doses, while no significant difference (p = 0.146) was found between the different activation ultrasound settings. There was neither a correlation between the tumor size and the number of stuck phase-shift bubbles. The results show very good correlation with the resultss obtained from manual counting.

Biosyntetiske genkluster (BGC-er) fasiliterer produksjonen av sekundære metabolitter i organismer. Disse forbindelsene er kjent for å ha nyttige egenskaper som antibiotisk, antiviral eller anti-tumor aktivitet. Imidlertid har bare en liten brøkdel av identifiserte BGC-er hatt sine produkter og metabolske stier eksperimentelt verifisert, der flertallet av dem kun er identifisert gjennom in silico informasjonsutvinning av genomdata. Den manuelle eksperimentelle analysen av disse ikke-karakteriserte BGC-ene er begrenset av treghet og av lave produksjonsutbytter for deres tilknyttede produkter i laboratorieforhold. For å effektivt studere BGC-er og veilede den eksperimentelle analysen av dem så trengs det et verktøy som kan modellere produksjonen av deres tilknyttede sekundære metabolitter med høy presisjon. Dette arbeidet presenterer en metode for å automatisk rekonstruere metabolske stier av biosyntetiske genkluster (BGC-er) fra annoterte gen-data. Denne metoden ble utviklet som en programvare (ARMRiPP) som bruker data fra BGC-genomdatautvinningsprogrammet, antiSMASH, og returnerer metabolske stier i en dataform som enkelt kan implementeres i genomskala metabolske modeller (GEM-er). For å veilede sti-rekonstruksjonen så ble ARMRiPP utviklet med evne til å generere prediksjoner av den kjemiske strukturen til BGC-assosierte forbindelser. Fokuset var på en av de større BGC-familiene, ribosomalt syntetiserte og post-translasjonelt modifiserte peptider (RiPP), og tre hovedklasser ble implementert: lanthipeptider, thiopeptider og lasso-peptider. ARMRiPP sin evne til å forutsi riktige strukturer og riktige stier ble begge evaluert. Tanimoto-score ble brukt som et mål for strukturell likhet, og vi kom frem til en gjennomsnittlig score på 0,9 for prediksjon av riktige strukturer ved et utvalg av 57 strukturer fra forskjellige RiPP-klasser, som indikerer høy nøyaktighet. Fire av de rekonstruerte stiene ble testet for deres evne til å forutsi korrekt produksjonsutbytte av de tilknyttede RiPP-forbindelsene sammenlignet med de kjente stiene fra litteraturen, og her kom vi frem til en gjennomsnittlig feilmargin på 7%. Begrensningene ved disse resultatene skyldtes i stor grad faktorer som manglende implementering av ulike RiPP biosyntetiske reaksjoner, begrenset mekanistisk kunnskap om biosyntesen til visse RiPP, begrensninger i antiSMASH-annoteringer og potensiell dataforskyvning. Nøyaktigheten av den strukturelle prediksjonen til programvaren ble vurdert i lys av ytelsen til en lignende programvare, PRISM. Som en casestudie så ble de rekonstruerte stiene brukt til å estimere den metabolske byrden for produksjonen av RiPP-forbindelser i forskjellige organismer, ved bruk av flere GEM-er. Små forskjeller i metabolsk byrde ble observert mellom RiPP-klassene, og større forskjeller mellom forskjellige fylogenetiske grupper. Lengden på RiPP-forgjengerpeptidet ble observert å ha en stor effekt på den metabolske byrden til RiPP-stien. Det ble også observert en generell trend at metabolsk byrde ville være lavere for RiPP-stier som ble satt inn i GEM-er av sin naturlige vert. Ved videre undersøkelse ble det avdekket statistisk signifikante forskjeller i metabolsk byrde hos RiPP-stier i deres naturlige verter og i heterologe verter, og denne forskjellen var mer signifikant for stier som ble satt inn i verter av forskjellig fylogeni enn deres naturlige. Observasjonene fra disse resultatene ble diskutert fra perspektivet av mikrobiell økologi, og for hvorvidt de har sammenheng med BGC-spesifikke evolusjonære effekter. Biosynthetic gene clusters (BGCs) facilitate the production of secondary metabolites in organisms. These compounds are known to have useful properties such as antibiotic, antiviral or anti-tumor activity. However, only a tiny fraction of identified BGCs have had their products and metabolic pathways experimentally verified, with the vast majority of them being identified in silico by genome mining tools. The manual experimental analysis of these uncharacterized BGCs is limited due to it being much slower than the discovery rate of BGCs and due to low production yields of their associated products in laboratory conditions. To effectively study BGCs and guide the experimental analysis of them, there is a need for a tool which can accurately model the metabolic production of their associated secondary metabolites. This thesis presents a method for automatically reconstructing the metabolic pathways of biosynthetic gene clusters (BGCs) from annotated gene data. This method was developed as a piece of software (ARMRiPP) that uses data from the BGC genome mining tool, antiSMASH, and outputs metabolic pathway data that can easily be implemented into genome-scale metabolic models (GEMs). As a way of guiding the pathway reconstruction, ARMRiPP was developed to generate structure predictions of the BGC associated compounds. The focus was on the major BGC family, ribosomally synthesized and post-translationally modified peptides (RiPPs), and three major classes were implemented: lanthipeptides, thiopeptides and lasso peptides. Both ARMRiPP’s ability to predict correct structures and correct pathways was evaluated. Using the Tanimoto score as a metric of structural similarity, we arrived at an average score of 0.9 across 57 BGCs of different RiPP classes, indicating a high accuracy. Four of the reconstructed pathways were tested for their ability to correctly predict production yield of the associated RiPP compounds when compared to their pathways in literature, and here we arrived at an average error of 7%. Limitations of these results related in large part to factors such as missing implementations for various modification reactions, limited mechanistic knowledge of certain RiPP pathways, limitations in antiSMASH annotations and potential sources of data skewness. Structural prediction accuracy was also considered in light of the performance of a similar software, PRISM. As a case study, the reconstructed pathways were used to estimate the metabolic burden of RiPP compound production in different organisms, using multiple GEMs. Slight differences in metabolic burden were observed between the RiPP classes, and larger differences between different phylogenetic groups. It was also observed that precursor length had a large impact on the associated metabolic burden of RiPPs. A general trend was observed that metabolic burden would be lower for RiPP pathways when put into GEMs of their native host. Investigating this further revealed statistically significant differences in the metabolic burden of pathways in their native hosts and in heterologous hosts, and this difference was more pronounced for pathways put into hosts of different phylogeny than their native. The observations from the metabolic burden results were discussed from the perspective of microbial ecology, and how it potentially relates to BGC specific evolutionary effects.

Vi har i dette arbeidet utviklet og implementert en algoritme som konverterer informasjon om predikerte biosyntetiske genklustere (BGC'er) som gitt av antiSMASH til metabolske reaksjonsveier for bruk i genomskalamodeller (GEM'er). Nøyaktigheten til algoritmen blir evaluert gjennom en detaljert sammenligning med eksperimentelt bestemte metabolske reaksjonsveier for åtte BGCer. Vi rapporterer 82 % gjennomsnittlig nøyaktighet for PKS - og NRPS-domener generelt, som følge av 78% nøyaktighet i substratspesifisitet for forlenger-enheter, og 84 % nøyaktighet for kofaktorassosierte reaksjoner. Med denne algoritmen har vi også konstruert metabolske veier for alle T1PKS, transAT-PKS og NRPS BGCer som finnes i MIBiG-databasen. Basert på smCOG-definisjoner, var vi i stand til å forutsi syntese av den uvanlige forlenger-enheten metoksymalonyl-ACP. Fra andre smCOG-definisjoner ble det etablert en sammenheng mellom antall påviste glykosyltransferaser i en BGC, og antallet glykosylgrupper som deltok i den metabolske reaksjonsveien til sekundærmetabolitten. To andre tilleggsreaksjoner kunne også forutses på samme vis. For tilleggsreaksjoner som ikke er inkludert i de konstruerte metabolske reaksjonsveiene, prøver vi å belyse konsekvensen av dette. Vi diskuterer også de forskjellige hindringene man står overfor når man prøver å konstruere metabolske reaksjonsveier fra BGCer, samt utfordringer man møter ved modellering av sekundærmetabolisme generelt. Vi avslutter med å foreslå at SubClusterBLAST-funksjonaliteten til antiSMASH utvides til å omfatte ytterligere kjente tilleggsreaksjoner som finnes for PKS/NRPS. I tillegg foreslår vi å oppdatere databasene som brukes til prediksjon av NRPS/PKS modulspesifisitet, slik at prediksjonene som antiSMASH gir - og dermed de metabolske reaksjonsveiene som algoritmen produserer - blir mer tro til sine reelle motparter. Prosjektet er tilgjengelig fra: https://github.com/FredrikFossheim/MasterThesis \noindent We have in this work developed, and implemented, an algorithm that converts information about predicted biosynthetic gene clusters (BGCs) as provided by antiSMASH into metabolic pathways for use in genome-scale metabolic models (GEMs). The accuracy of the algorithm is evaluated through a detailed comparison with experimentally determined pathways for eight BGCs. We report an overall 82% average accuracy for polyketide synthase (PKS) and nonribosomal peptide synthase (NRPS) domains in general, resulting from a 78 % accuracy in substrate specificity for extender units, and 84% accuracy for cofactor-associated reactions. With this algorithm, we have also constructed metabolic pathways for all T1PKS, transAT-PKS and NRPS BGCs that exist in the MIBiG database. Based on smCOG definitions, we were able to predict the synthesis of the uncommon extender unit methoxymalonyl-ACP. From other smCOG definitions, there was also established a relationship between the number of detected glycosyltransferases in a BGC, and the number of glycosyl groups that took part in the metabolic pathway of the secondary metabolite. Two other tailoring reactions were found to be predictable by the same means. For tailoring reactions that are not included in the constructed metabolic pathways, we attempt to elucidate the consequence of these. We also discuss the different obstacles one faces when attempting to construct metabolic pathways from BGCs, as well as those of modeling secondary metabolism in general. We end by suggesting that the SubClusterBLAST functionality of antiSMASH is expanded to include additional known tailoring reactions that are found for PKS/NRPS. In addition, we suggest updating the databases used for prediction of NRPS/PKS module specificity so that the predictions that antiSMASH makes - and in turn the metabolic pathways that the algorithm produces - are more true to their real life counterparts. The project is available from https://github.com/FredrikFossheim/MasterThesis

CSD-metoden er et rammeverk utviklet ved Institutt for Bioteknologi og Matvitenskap ved Norges Teknisk-Naturvitenskapelige Universitet (NTNU). Metoden brukes for å studere differensielle samuttrykk av gener i to grupper, og separerer de ulike formene av differensielle samuttrykk inn i de tre kategoriene konservert (C), spesifikk (S) og differensiert (D), derav CSD. I denne masteroppgaven ble CSD-metoden brukt til å generere tidsavhengige multiplekse CSD nettverk hvor MS-pasienter ble sammenlignet med friske kontroller. Et program som tar inn tre CSD-nettverk, laget med de to samme gruppene av pasienter og kontroller ved tre ulike tidspunkt, ble utviklet for å generere multiplekse nettverk som sammenlignes på node-, link- og nabolagsnivå. Det ble dannet to slike multiplekse nettverk, et hvor MS-pasienter ble sammenlignet med friske kontroller, og ett hvor MS-pasienter behandlet med IFN-b ble sammenlignet med den samme kontrollgruppen. De multiplekse CSD-nettverkene avslørte biologiske prosesser med genutrykksmønstre som ikke ble påvirket av sykdommen eller behandlingen. Eksempler på slike prosesser var hjerte-relaterte prosesser, prosesser involvert i utvikling av ekstracellulære vev, samt den spesifikke prosessen "Kollagenaktivert tyrosin kinase reseptor signalvei". En prosess som utmerket seg med et spesifikt (S) utrykksmønster mellom pasienter og kontroller var "NLS-bærende proteinimport til kjernen". En hypotese er at denne prosessen kan assosieres med genregulering hos MS-pasienter. Mange av de mest fremtredende hubbene i nettverkene kunne kobles til MS aller IFN-b. Både sykdommen multippel sklerose og behandlingen IFN-b kan kobles til immunsystemet, og å separere prosesser relatert til sykdomsforløpet og behandlingsmekanismer viste seg derfor å være utfordrende. CLCN3, LUZP6 og LGALS1 var gener som viste seg som tydelige hubber i det multiplekse nettverket bestående av ubehandlede MS-pasienter, og som kunne assosieres med MS fra tidligere studier. ARIH1, HERC6 og NUB1 var viktige hubber i nettverket dannet med IFN-behandlede pasienter. Dette er regulerende gener som kan assosieres med regulering av immunsystemet og blir påvirket av IFN-b. HADHB var også sentral i dette nettverket, og kan assosieres med MS. Konstruksjonen av multiplekse CSD-nettverk ga klare indikasjoner på genuttrykksmønstre bevart over tid mellom MS-pasienter og friske kontroller. Undersøkelse av overlappende linker og nabolag utfylte hverandre og ga et komplekst bilde av stabile genuttrykksmønstre. Differensielle (D) og spesifikke (S) samuttrykk av gener var mer stabile over tid gruppen med IFN-behandlede pasienter sammenlignet med ubehandlede pasienter. En mulig forklaring på denne tendensen er at mangfoldet i MS-typer og variasjonen i sykdomsforløpet kan resultere i stor variasjon i genuttrykksmønstre, mens effekten til et enkelt virkestoff kan gi et sterkere signaler. The CSD method is a framework developed at the Department of Biotechnology and Food Science at the Norwegian University of Science and Technology (NTNU). The method is used to study differential co-expression of genes in two groups, and separates the various forms of differential co-expression into three different types: Conserved (C), Specific (S) and Differentiated (D), thereby CSD. In this thesis, the CSD method was used to generate time-dependent multiplex networks that compared MS patients to healthy controls. A program, that uses three CSD networks generated from the same two groups at three different time-points, was developed to make multiplex networks that were investigated on node, link and community level. Two multiplex networks were constructed and studied, one comparing untreated MS patients to healthy controls, and one comparing IFN-b-treated MS patients to healthy controls. The multiplex CSD networks revealed biological processes with gene expression patterns that were not affected by the disease nor treatment. Examples of such processes were some cardiac processes, processes involved in the development of extracellular tissues and morphogenesis, in addition to the specific process "Collagen-activated tyrosine kinase receptor signalling pathway". A biological process that stood out with a preserved specific (S) co-expression pattern between MS-patients and healthy controls was "NLS- bearing protein import into nucleus". A hypothesis is that this process could be involved in gene regulation in MS-patients. Many of the prominent hubs in the network could be associated with MS of IFN-b. As the autoimmune disease multiple sclerosis and the medication IFN-b both interact with the immune system, separating the mechanisms of disease pathogenesis from treatment mechanisms was challenging. CLCN3, LUZP6 and LGALS1 were genes presented as distinct hubs in the multiplex CSD network of untreated MS patients that could be associated with MS based on published literature. ARIH1, HERC6 and NUB1 were genes presented as important hubs in the network of IFN-treated MS patients. These are all regulating genes that can be associated with immune regulation and possibly affected by IFN-b. HADHB was also central in this network and can be associated with MS. The generation of multiplex CSD networks provided clear indications of time preserved expression patterns between MS patients and controls. Investigation of overlapping links and communities provided, in combination, a complex picture on stable expression patterns. Differential (D) and specific (S) co-expression of genes were more stable over time in the group of IFN-treated patients compared to untreated MS patients. A possible explanation for this trend is the diversity of MS sub-types and variation in disease progression that can result in variation in gene expression patterns, while the effect of one simple therapeutic could provide a stronger expression pattern.