A Machine Learning Bioinformatics Method to Predict Biological Activity from Biosynthetic Gene Clusters

Abstract

Research in natural products, the genetically encoded small molecules produced by organisms in an idiosyncratic fashion, deals with molecular structure, biosynthesis, and biological activity. Bioinformatics analyses of microbial genomes can successfully reveal the genetic instructions, biosynthetic gene clusters, that produce many natural products. Genes to molecule predictions made on biosynthetic gene clusters have revealed many important new structures. There is no comparable method for genes to biological activity predictions. To address this missing pathway, we developed a machine learning bioinformatics method for predicting a natural product’s antibiotic activity directly from the sequence of its biosynthetic gene cluster. We trained commonly used machine learning classifiers to predict antibacterial or antifungal activity based on features of known natural product biosynthetic gene clusters. We have identified classifiers that can attain accuracies as high as 80% and that have enabled the identification of biosynthetic enzymes and their corresponding molecular features that are associated with antibiotic activity.

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

“…We intend to add to this multi-omics approach: 1) BGC bioactivity (20); 2) MS/MS bioactivity, by creating a new machine learning tool to predict bioactivity straight from the MS/MS spectra, and; 3) MS/MS substructure predictions, by integrating tools like MS2LDA (37), CSI: FingerID /SIRIUS 4 (38) and MassQL (not published yet). NPClassScore (33) demonstrated that biosynthetic class can be predicted with a combination of CANOPUS and MolNetEnhancer, hence, the NPOmix users can already run the KNN version with similarity and biosynthetic class as features, a version that yielded a precision of 92.9% in the validation set.…”

“…We stress that a larger training dataset with more complete genomes is likely to increase the size of the validation set by adding more valid BGCs into the analysis. We were also able to combine NPOmix with in silico metabolomics tools like Dereplicator+ (20) to make new links between MS/MS spectra, BGCs, and molecular structures, as we will exemplify by brasilicardin A. This was accomplished by annotating cryptic MS/MS spectra (without a GNPS library hit and therefore not present in the GNPS database) to known BGCs (found in the MIBiG database).…”

“…In silico predictions using raw DNA sequence and unlabeled mass fragmentation data can be used to predict: bioactivity, either by gene content or docking experiments; full planar structures, by dereplication, homology, or de novo ; partial stereochemistry, and; novelty, based on the abundance from all bacteria sampled from nature so far. For example, some random forest classifiers by the Clardy lab (20) at Harvard medical center use only the BGC sequences to predict with about 80% precision if a BGC will produce an anticancer, antifungal, or antibacterial metabolite. Coelichelin was isolated using an in silico structure prediction that indicated the peptide to be a siderophore, which was confirmed by culturing the producer in an iron-deficient media and isolating the induced metabolite (overexpressed when compared to the control).…”

A Brief Review of Machine Learning-Based Bioactive Compound Research

“…String information, such as DNA nucleotides and/or protein amino acid sequences, can be used to build a ML model for a particular purpose. Walker et al [65] developed a machine learning bioinformatics method for predicting a natural product's antibiotic activity directly from the sequence of its biosynthetic gene cluster. In this study, they assembled a training dataset from the MiBIG (Minimum Information about Biosynthetic Gene Custer) database, comprising sequences of biosynthetic gene clusters (BGCs).…”

A systematically biosynthetic investigation of lactic acid bacteria reveals diverse antagonistic bacteriocins that potentially shape the human microbiome

“…Machine learning models reveal that most BGCs may encode antagonistic SMs Given the abundance and prevalence of LAB BGCs in the human microbiome, we next want to study the potential bioactivities of BGC-encoding SMs. The bioactivity of SMs encoded by BGCs was recently predicted using machine learning strategies based on chemical ngerprints of predicted compound structure, protein family (PFAM) domains, and other genetic features [32][33][34]. Here, we adapted four common machine learning classi ers (logistic regression, elastic net regression, random forest, and support vector machines) to predict the bioactivities of LAB-derived SMs.…”