Enhanced correlation-based linking of biosynthetic gene clusters to their metabolic products through chemical class matching

Abstract

Background It is well-known that the microbiome produces a myriad of specialised metabolites with diverse functions. To better characterise their structures and identify their producers in complex samples, integrative genome and metabolome mining is becoming increasingly popular. Metabologenomic co-occurrence-based correlation scoring methods facilitate the linking of metabolite mass fragmentation spectra (MS/MS) to their cognate biosynthetic gene clusters (BGCs) based on shared absence/presence patterns of metabolites and BGCs in paired omics datasets of multiple strains. Recently, these methods have been made more readily accessible through the NPLinker platform. However, co-occurrence-based approaches usually result in too many candidate links to manually validate. To address this issue, we introduce a generic feature-based correlation method that matches chemical compound classes between BGCs and MS/MS spectra. Results To automatically reduce the long lists of potential BGC-MS/MS spectrum links, we match natural product (NP) ontologies previously independently developed for genomics and metabolomics and developed NPClassScore: an empirical class matching score that we also implemented in the NPLinker platform. By applying NPClassScore on three paired omics datasets totalling 189 bacterial strains, we show that the number of links is reduced by on average 63% as compared to using a co-occurrence-based strategy alone. We further demonstrate that 96% of experimentally validated links in these datasets are retained and prioritised when using NPClassScore. Conclusion The matching genome-metabolome class ontologies provide a starting point for selecting plausible candidates for BGCs and MS/MS spectra based on matching chemical compound class ontologies. NPClassScore expedites genome/metabolome data integration, as relevant BGC-metabolite links are prioritised, and researchers are faced with substantially fewer proposed BGC-MS/MS links to manually inspect. We anticipate that our addition to the NPLinker platform will aid integrative omics mining workflows in discovering novel NPs and understanding complex metabolic interactions in the microbiome.

iPRESTO: Automated discovery of biosynthetic sub-clusters linked to specific natural product substructures

“…Combining such solutions with the integration of more diverse species, new annotations, and improved correlation scoring methods like the one developed in Hjörleifsson Eldjárn et al [ 35 ] would improve such analyses drastically. Furthermore, we expect that combining co-occurrence based scores (such as the standardised Metcalf one) with feature-based scores, such as NPClassScore [ 36 ], and the here developed iPRESTO, will further help to prioritise plausible BGC-MS/MS spectral links [ 12 , 13 ]. Indeed, we expect that tools like iPRESTO could in the future be built into frameworks like NPLinker [ 35 ].…”

NPOmix: A machine learning classifier to connect mass spectrometry fragmentation data to biosynthetic gene clusters

“…3B , score with co-occurrence threshold) because this very high value was obtained using only the similarity and the biosynthetic class features and three nearest neighbors, a good number of candidates for genome mining. Moreover, the use of just the similarity and biosynthetic class is very appealing because these kinds of features can be well predicted for metabolites with known or unknown metabolites, for example, by using GNPS for similarity and CANOPUS and/or MolNetEnhancer for biosynthetic class predictions, as demonstrated by NPClassScore ( 39 ). It is much more challenging to predict substructures for unknown/cryptic metabolites, which is a topic of ongoing research by our group and others.…”

“…The latter is mainly due to the fact that NPOmix selects which metabolites (MS/MS spectra) should be targeted for classification in opposition to NPLinker that attempts to link all metabolites present in the analyzed metabolomes; (iii) NPOmix is a tool for connecting metabolites to BGCs, NPLinker is a framework that facilitates integrative omics analysis, and in the future, NPOmix will likely become part of the NPLinker framework. We stress that “benchmarking NPLinker” effectively means benchmarking the (standardized) Metcalf strain-correlation and Rosetta scoring systems; in the future, it is likely a combination of various strain-correlation-based and feature-based scoring systems (such as NPClassScore) ( 39 ) that will enable the most effective integrative omics mining. Both precisions for top-1, and top-1 shared ranks are provided in Fig.…”

“…4B ) and the recall displayed is optimal (100%, indicating the absence of false negatives). Of note, the prediction power of machine learning tools like NPOmix can be even higher with more top candidates (for this comparison we only used top-1) and better features, such as demonstrated by the implementation of biosynthetic class with NPClassScore ( 39 ) to improve NPLinker links and the use of extra features other than similarity by NPOmix. Additionally, the precision of NPOmix can improve with a bigger training set.…”

“…The links for these MS/MS files can also be found in the “NPOmix_SI-installation_and_running” notebook for downloading (GitHub repository https://github.com/tiagolbiotech/NPOmix ). The testing set included MS/MS spectra from the PoDP database, spectra from the GNPS ( 14 ), and also spectra also used in the NPLinker publication ( 39 ). If the potential users find the tool challenging to run, we have our contact information on the main webpage ( https://www.tfleao.com/npomix1 ) to submit samples and we expect that promising results will lead to fruitful collaborations.…”