Photosynthetic protein classification using genome neighborhood-based machine learning feature

Abstract

Identification of novel photosynthetic proteins is important for understanding and improving photosynthetic efficiency. Synergistically, genome neighborhood can provide additional useful information to identify photosynthetic proteins. We, therefore, expected that applying a computational approach, particularly machine learning (ML) with the genome neighborhood-based feature should facilitate the photosynthetic function assignment. Our results revealed a functional relationship between photosynthetic genes and their conserved neighboring genes observed by ‘Phylo score’, indicating their functions could be inferred from the genome neighborhood profile. Therefore, we created a new method for extracting patterns based on the genome neighborhood network (GNN) and applied them for the photosynthetic protein classification using ML algorithms. Random forest (RF) classifier using genome neighborhood-based features achieved the highest accuracy up to 87% in the classification of photosynthetic proteins and also showed better performance (Mathew’s correlation coefficient = 0.718) than other available tools including the sequence similarity search (0.447) and ML-based method (0.361). Furthermore, we demonstrated the ability of our model to identify novel photosynthetic proteins compared to the other methods. Our classifier is available at http://bicep2.kmutt.ac.th/photomod_standalone, https://bit.ly/2S0I2Ox and DockerHub: https://hub.docker.com/r/asangphukieo/photomod.

Wide range of metabolic adaptations to the acquisition of the Calvin cycle revealed by comparison of microbial genomes

“…The random forest analysis yielded a machine learning classifier that labeled genomes as CBB-positive or CBB-negative with 72.9% accuracy using ECs and 76.5% accuracy using Pfams, compared to the expected accuracy of 50% if picking labels at random. For comparison, it has been shown that a random forest can achieve 88% accuracy in predicting photosynthetic proteins based on gene neighborhood [ 49 ], a tree-based classifier can achieve 86% accuracy (94% with a k-nearest neighbor model) in predicting the recombination status of HIV genomes [ 50 ], and a support vector machine can achieve 87% accuracy in classifying bacteria as pathogenic or not based on their proteomes [ 51 ]. The accuracy achieved in the present study was likely limited by false negative genomes, genomes in the process of adapting to recent loss or gain of the Calvin cycle, a low amount of training data, and limited ability of the random forest to focus on relevant aspects of the data.…”

PhotoModPlus: A web server for photosynthetic protein prediction from genome neighborhood features

“…15,191 protein sequences consisting of at least one of 61 photosynthesis-specific GO terms identified by Ashkenazi et al [ 12 ] were included in the dataset. To avoid incomplete gene neighborhood identification, only photosynthetic proteins from 154 photosynthetic prokaryotes with complete genomes were included in the dataset, as previously reported [ 15 ]. To reduce sequence redundancy, we used the USEARCH analytical tool [ 22 ] to cluster similar sequences (sequence identity ≤ 50% as a diverse dataset and ≤ 70% as an easy dataset), using the command: where the input file is in FASTA format, identity is the percent sequence identity cutoff for the cluster, and output is the selected representative sequence in FASTA format.…”

“…Two gene clusters were merged into the same neighborhood gene cluster if they were in a range of 200-1000 bp in the divergent direction, following the operon interaction concept as demonstrated in S1 Fig. The homologous relationship between protein sequences was determined by the protein clustering method with three stringent criteria: 1E-10, 1E-50, and 1E-100, according to a previous study [15]. The genome neighborhood conservation scores (Phylo scores) were calculated based on the phylogenetic tree of organisms that conserve those gene neighborhoods, as described previously [15]. The quantile cutoff points were determined and used for converting the raw Phylo scores to simple numeric forms (i.e., 0, 1, 2, and 3).…”

“…SCMPSP performs better than the sequence similarity search method but is not significantly better than other machine learning models, indicating a limitation of the sequence-based feature. Therefore, a model called PhotoMod, which employs the genome neighborhood network (GNN) as a feature, was developed [ 15 ]. PhotoMod combines protein clustering, genome neighborhood conservation scoring, and the random forest (RF) algorithm to classify photosynthetic proteins.…”

“…The web server ( Fig 1 ) contains three main applications. The first application, PhotoMod, is our previously developed machine learning model that serves for photosynthetic protein classification [ 15 ]. The second, PhotoModGO, is a new machine learning model utilizing multi-label learning and genome neighborhood profile as a feature to predict 24 sub-functional classes of a photosynthetic function.…”