plastiC: A pipeline for recovery and characterization of plastid genomes from metagenomic datasets

Version Changes

Revised. Amendments from Version 1

In this revision, we have focused on updating the usability of our workflow for others in the community. Specific changes can be found in our GitHub repository and include:  1) Improved documentation  2) Switching software dependencies from a conda environment to singularity to simplify the installation process.  In addition to this, we also map our recovered plastid genomes in our case study to a reference  Trebouxia sp. plastid genome.

Abstract

The use of culture independent molecular methods, often referred to as metagenomics, have revolutionized the ability to explore and characterize microbial communities from diverse environmental sources. Most metagenomic workflows have been developed for identification of prokaryotic and eukaryotic community constituents, but tools for identification of plastid genomes are lacking. The endosymbiotic origin of plastids also poses challenges where plastid metagenomic assembled genomes (MAGs) may be misidentified as low-quality bacterial MAGs. Current tools are limited to classification of contigs as plastid and do not provide further assessment or characterization of plastid MAGs. plastiC is a workflow that allows users to identify plastid genomes in metagenome assemblies, assess completeness, and predict taxonomic association from diverse environmental sources. plastiC is a Snakemake workflow available at https://github.com/Finn-Lab/plastiC. We demonstrate the utility of this workflow with the successful recover of algal plastid MAGs from publicly available lichen metagenomes.

Introduction

Culture-independent sequencing of environmental samples, often referred to as metagenomics, has revolutionised the field of microbial ecology, allowing taxonomic and functional characterization of microbial communities from diverse environmental sources. A variety of bioinformatic tools and pipelines are available for the generation and quality estimation of prokaryotic and eukaryotic metagenomic assembled genomes (MAGs) ( Parks et al., 2015; Saary et al., 2020). Organellar genomes are essential to the functioning of eukaryotic organisms, but are poorly assessed in current workflows. Plastid (or chloroplast) genomes are of particular interest because of their roles in primary production, yet the diversity of plastids present in metagenomic data is usually ignored and constrains the full understanding of community structure and function. In standard metagenomic workflows, plastid MAGs may be mistakenly flagged as low-quality bacterial genomes as they lack all of the single copy bacterial marker genes due to the endosymbiotic origin of organelles resulting in potential discard from downstream analyses.

Most existing MAG quality assessment methods are rooted in single-copy marker gene (SCMG) approaches. However, SCMG approaches are not easily transferred to the genomes of plastid organelles due to their reduced genome size, variable gene content (especially driven by organellar-nuclear gene transfer) and large repetitive regions. Thus, alternative methods for quality evaluation are required. Here, we present plastiC, a Snakemake ( Mölder et al., 2021) workflow for the identification and evaluation of plastids in metagenomic samples - see Software availability ( Cameron, 2023). and demonstrate the utility of this workflow to successfully recover algal plastid genomes from lichen metagenomes.

Methods

Implementation

plastiC is designed to analyse contigs generated from metagenomic assembly workflows or high fidelity long-reads ( Figure 1). Contigs from these assemblies can be sorted into bins using intrinsic sequence and coverage features, and these bins processed to identify those that putatively correspond to prokaryotic and/or eukaryotic MAGs. These MAGs are typically processed to identify likely taxonomic affiliation derived from marker gene placement, and their quality ( i.e. completeness and contamination) assessed. While plastiC can be run on raw assemblies, we recommend removing contigs that are assigned to “high quality” bins/MAGs before analysis. Putative plastid contigs are identified using Tiara ( Karlicki et al., 2022). As plastid genomes are expected to have consistent coverage and tetranucleotide frequency, the assembly is subsequently (re-)binned using metaBAT2 (RRID: SCR_019134) ( Kang et al., 2019) with a reduced bin size of 50 kb to ensure that putative plastid bins are retained. The resulting bins are subsequently scanned for plastid contigs as identified by Tiara, and bins containing more than a user-specified threshold (default: 90%) of plastid sequences are retained as probable plastid bins.

Figure 1. Data preparation and overall workflow for plastic.

Users provide metagenomic assemblies or high fidelity long reads from which putative plastid bins are identified. A summary report and FASTA files of identified marker loci are generated for each sample.

Machine learning algorithms trained on Kyoto Encyclopedia of Genes and Genomes (KEGG) module completeness were recently demonstrated as an effective approach for estimating metagenomic bin completeness ( Chklovski et al., 2022) and we adapted this approach in plastiC. Prodigal (RRID: SCR_021246) ( Hyatt et al., 2010) is used to predict genes in the likely plastid bins, and these predictions are compared to the UniRef100 database ( Bateman et al., 2021) using DIAMOND blastp (RRID: SCR_016071) ( Buchfink et al., 2021) to obtain KEGG annotations ( Kanehisa et al., 2016). plastiC then uses the KEGG annotations to calculate KEGG module coverage, and thereby estimate plastid completeness. Due to the presence of shared metabolic pathways and genes between plastid and bacterial chromosomes, bacterial bins may also score as highly complete using this approach, hence the requirement to remove bacterial MAGs before running plastiC.

Importantly, plastiC also identifies the potential eukaryotic source of the plastid. plastiC initially performs a taxonomic classification on the plastid contigs using CAT ( von Meijenfeldt et al., 2019) and additionally searches for taxonomic marker genes using hmmer (for rbcL) (RRID: SCR_005305) ( Eddy, 2022) and barrnap (for rRNA loci) (RRID: SCR_015995) ( Seeman, 2022). These loci can be used for further taxonomic analyses. As outputs, plastiC produces report files summarising bin span, estimated completeness, and taxonomic allocation for each putative plastid bin, as well as FASTA files of the marker loci identified.

Model training, testing and validation

Reference protein sequences from complete plastid genomes available in RefSeq (n = 2633) using Entrez Direct ( Kans, 2023) were downloaded on 2022-11-2 (see Underlying data). One tenth (261) of these references were randomly removed to generate an independent testing set. The training set (n = 2372) was further refined to exclude any reference genomes that were smaller than the median plastid genome size of protists (94 kb) to ensure that outlier small genomes do not impact the training of the model for a final training set of 2279 plastid genomes. DIAMOND blastp was run on both training and testing datasets with UniRef100 to obtain KEGG annotations, which were used to calculate KEGG module completeness based on identified KEGG hits. For training sets, reference genome KEGG annotation counts were subsampled without replacement to create simulated examples of plastid genomes with lower levels of completeness ranging from 0 – 100% in increments of 5%. Test reference genomes were also subsampled to create simulated examples to validate the robustness of the completeness estimates ranging from 10 – 100% in increments of 10%.

Scikit-learn (RRID: SCR_002577) was utilised to develop and validate machine learning models for estimating metagenomic plastid genome completeness. Specifically, Ada boosting, gradient boosting and random forest regressions were evaluated to determine the effectiveness for estimating completeness of plastid genomes. Training data were split 90 training set:10 test set. K-folds (n=5) cross-validation with shuffling was performed to cross-validate the model.

Each model was tested on reference plastid genomes that were not used in the training set. In addition to the test plastid genome set, KEGG module completeness was predicted for a set of mitochondrial genomes (n = 142) to evaluate whether the model could accurately differentiate between different organellar genomes. Across the cross-validated model set, all three regression models were able to differentiate between plastid and mitochondrial completeness ( i.e., predict low completeness for mitochondria; high completeness for plastids; Table 1). However, the Ada boosting regression had the lowest plastid completeness estimate (91.79%) and highest mitochondrial estimate (5.62%) of tested models. In combination with the higher mean-squared error, this suggests that the Ada boosting regression model is not the best performing model for application in plastid completeness estimates.

Table 1. Median predicted completeness values ± standard deviation on whole plastid and mitochondrial reference genomes with k-fold cross validation (n = 5; shuffled) to ensure differentiation between organellar genome completeness scores.

	Ada boosting regression		Gradient boosting regression		Random forest regression
	Plastids	Mitochondria	Plastids	Mitochondria	Plastids	Mitochondria
CV1	91.79 ± 6.36	5.6 ± 0.00	98.11 ± 7.27	0.43 ± 0.00	99.06 ± 6.97	0.06 ± 0.00
CV2	91.98 ± 6.42	5.40 ± 0.00	97.97 ± 7.49	0.52 ± 0.00	99.00 ± 7.17	0.05 ± 0.00
CV3	92.15 ± 6.43	4.83 ± 0.00	97.99 ± 7.11	0.47 ± 0.00	99.08 ± 6.97	0.05 ± 0.00
CV4	91.54 ± 6.31	6.81 ± 0.00	98.10 ± 7.60	0.45 ± 0.00	98.98 ± 7.21	0.05 ± 0.00
CV5	91.51 ± 6.55	5.49 ± 0.00	97.95 ± 7.39	0.44 ± 1.73	99.05 ± 7.09	0.05 ± 0.00
Mean	91.79 ± 6.41	5.62 ± 0.00	98.02 ± 7.36	0.45 ± 0.35	99.03 ± 7.08	0.053 ± 0.00

The performance of models with plastids of varying completeness was assessed by subsampling the test genomes to simulate varying levels of completeness. Predicted completeness values were compared to expected values at the subset levels to determine efficacy of the model on accurate estimation. Median prediction values and standard deviation was calculated for each iteration of the model produced through cross-validation ( Table 2). A linear regression was performed ( Figure 2; Table 3) on the predicted completeness compared to expected value and Pearson’s correlation R ² was calculated to identify similarity between predicted and expected completeness scores. All regressions and correlation coefficients were statistically significant (p < 2e-16) but the gradient boosting regression model showed the highest correlation between expected and predicted values. To conclude evaluation of the effectiveness of each model, differences between the expected and predicted values for each cross-validated model iteration were calculated ( Figure 3). The random forest regression model had the best performing mean-squared error (0.0001), highest completeness estimates for whole plastid genomes (99.03%) and lowest completeness estimates for whole mitochondrial genomes (0.053%). However, it frequently overestimated completeness (median difference = -7.5). Based on the correlation between expected and predicted and median value of discrepancy between predicted and expected values (n = 0.37), the gradient boosting regression model was chosen as the best-performing model for plastid genome completeness estimates.

Figure 2. Predicted completeness estimates on test plastid genomes (not used in model training) with three different regression models: Ada boosting, gradient boosting and random forest.

The resulting linear equation on linear regression of predicted ~ expected values is plotted to demonstrate the predictive performance of the different models.

Figure 3. Differences in expected completeness of subsampled plastid reference genomes and predicted completeness.

Positive values in difference indicate underestimation of plastid predicted completeness while negative values represent overestimation. Gradient boosting regression consistently had the smallest median discrepancy between predicted values and expected value (0.37), while random forest had the largest discrepancy frequently resulting in the overestimation of predicted plastid completeness (median = -7.5). Ada boosting regression performed moderately (median = -2.12).

Table 2. Median predicted completeness values for plastid reference genomes in independent testing validation set (n = 261).

Test set plastid references were subsampled down to 10% to evaluate the quality of completeness estimates provided to non-complete genomes.

Expected Completeness	Ada Boosting Regression	Gradient Boosting Regression	Random Forest Regression
100	92.56 ± 6.56	99.28 ± 7.41	100.00 ± 7.11
90	92.33 ± 6.58	95.78 ± 7.34	95.05 ± 6.72
80	85.07 ± 8.13	83.99 ± 8.39	87.50 ± 8.40
70	64.92 ± 7.85	69.63 ± 6.51	82.25 ± 7.95
60	49.21 ± 6.28	51.05 ± 5.22	60.00 ± 4.69
50	74.42 ± 7.65	68.87 ± 7.50	75.00 ± 7.07
40	49.04 ± 5.21	38.66 ± 5.03	57.10 ± 6.49
30	27.21 ± 3.88	36.73 ± 5.08	57.35 ± 8.76
20	28.68 ± 11.20	25.89 ± 6.73	34.80 ± 8.99
10	11.63 ± 4.71	8.34 ± 2.66	6.79 ± 6.18

Table 3. Linear regression and Pearson correlation as calculated comparing the predicted completeness to expected completeness on the independent test plastid genome estimations.

	Ada Boosting Regression	Gradient Boosting Regression	Random Forest Regression
Linear Equation	y = 0.88x + 9.19	y = 0.95x + 4.58	y = 0.87x + 16.85
Adjusted R 2	0.83	0.89	0.83
p-value (Linear regression)	<2.20e-16	<2.20e-16	<2.20e-16
Pearson’s Correlation	0.91	0.95	0.91
p-value (Pearson’s)	<2.20e-16	<2.20e-16	<2.20e-16

Operation

plastiC is a Snakemake workflow which automates the process of identification, assessment and taxonomic allocation of plastid genomes in metagenomic samples written using Python and bash. Specifically, this workflow contains eight rules (i.e., steps), including an optional mapping step. Users are required to populate the config. yaml providing information on paths to directories and files, and options for workflow customization (e.g., filter thresholds, inclusion of mapping rule). plastiC requires one of the following for each sample to be analysed: i) a metagenomic assembly, ii) high fidelity long-reads. If available, users also have the option to provide a bam file for mapped reads to the assembly. Users may opt to decrease the minimum percentage of plastid sequence required to be classified as a probable plastid bin ( min_plastid_content) to increase or decrease stringency as appropriate for individual use cases.

Since plastiC runs on previously generated metagenomic assemblies, memory requirements are typically low but may vary depend on the nature of the sample. In development, the majority of rules required less than 1 GB RAM for successful completion with the exception of i) contig classification using Tiara (<5 GB RAM) and ii) taxonomic source classification using CAT (<8 GB RAM). These memory requirements may vary depending on sample size and complexity of the community. Users may specify computational resource allocations for individual rules and execution on a high-performance compute cluster in the cluster.yaml.

Use cases

To demonstrate the effectiveness of plastiC on metagenomic data, lichen metagenomes were downloaded from the project accession PRJNA646656 (n = 13; Smith et al., 2020) in the European Nucleotide Archive (see Underlying data). Lichens are composite organisms composed of a mutualist symbiotic association between a primary fungal partner (mycobiont) and algal partner (photobiont). Chlorophyta (green alga) taxa are the photobionts in many lichen species and thus plastids should be present in lichen metagenomic samples. The analysed samples were derived from 10 species of lichen spanning six genera ( Table 4) which are all expected to contain Trebouxia, a green algal genus, as their primary photobiont. Metagenomic assemblies for these samples were generated as part of a large survey of lichen microbial symbionts ( Tagirdzhanova et al., 2024). In brief, metagenomic datasets were filtered using fastp and human contamination was removed using BMTagger. Quality-controlled reads were assembled using metaSPAdes. These assemblies were used with plastiC to recover plastid genomes.

Plastid contigs were identified in all 13 samples using Tiara ( Table 4). Metagenomic assemblies were binned using metaBAT2 with the reduced bin size threshold of 50 kb. These bins were then classified as being putative plastid genomes based on the fraction of plastid contigs. Bins that were comprised of >90% plastid nucleotides were retained as probable plastid bins for further analysis. Of the 13 lichen metagenomes analysed, a single plastid bin was identified in 7 of them ( Table 5). For the remaining six samples, plastid contigs were not successfully binned and were retained in the unbinned portion with other sequences.

Table 4. Lichen metagenome assembly and binning information.

Metagenomes were assembled into contigs and these assemblies were used to identify plastid contigs using Tiara and for binning with metaBAT2. Probable plastid bins were identified based on the distribution and location of plastid contigs within bins, with a threshold of >90% to be retained for downstream analyses.

Run Accession	Scientific Name	Total Contigs	Plastid Contigs	Total Bins	Probable Plastid Bins
SRR12240187	Xanthoparmelia chlorchroa	190210	12	15	1
SRR12240188	Xanthoparmelia chlorochroa	115748	12	18	1
SRR12240177	Xanthoparmelia maricopensis	364371	20	23	1
SRR12240174	Xanthoparmelia neocumberlandia	253456	108	23	0
SRR12240175	Xanthoparmelia neocumberlandia	171912	93	13	1
SRR12240180	Xanthopermelia chlorochoa	164812	14	20	1
SRR12240179	Xanthopermelia mexicana	253280	3	24	1
SRR12240178	Xanthopermelia plittii	190261	81	17	1
SRR12240185	Mobergia calculiformis	129614	505	28	1
SRR12240183	Physcia biziana	330121	64	27	0
SRR12240182	Physciella chloantha	650113	31	26	0
SRR12240181	Rinodina sp	170759	207	28	0
SRR12240184	Oxernella safavidorum	455310	587	22	0

Table 5. Probable plastid bin characteristics including bin span, total contig count and number of plastid contigs.

Completeness estimates were performed based on KEGG module coverage and gradient boosting regression, and taxonomic association of plastid genomes performed with CAT, and coverage of a reference chloroplast genome from a Trebouxia chloroplast genome.

Run Accession	Bin Span	Total Contig Count	Number of Plastid Contigs	Completeness Estimate	Taxonomic Association	Coverage of Ref. NC_ 045839.1
SRR12240187	231882	1	1	96.1	Trebouxia	23.6
SRR12240188	231734	2	2	96.1	Trebouxia	23.7
SRR12240177	238137	10	10	95.9	Trebouxia	26.6
SRR12240174	N/A	N/A	N/A	N/A	N/A	N/A
SRR12240175	N/A	N/A	N/A	N/A	N/A	N/A
SRR12240180	254207	5	5	96.1	Trebouxia	27.3
SRR12240179	258520	1	1	96.4	Trebouxia	30.0
SRR12240178	64547	11	10	10.8	Trebouxia	7.4
SRR12240185	77893	9	8	19.1	Trebouxia	12.1
SRR12240183	N/A	N/A	N/A	N/A	N/A	N/A
SRR12240182	N/A	N/A	N/A	N/A	N/A	N/A
SRR12240181	N/A	N/A	N/A	N/A	N/A	N/A
SRR12240184	N/A	N/A	N/A	N/A	N/A	N/A

Taxonomic source prediction was performed on the putative plastid bins. All plastid bins identified in the sample were attributed to the genus Trebouxia, which corresponds with expectations of the photobiont in these lichens being a trebouxoid green alga. Plastid bins ranged in estimated completeness from 10.8 to 96.4% and completeness was positively correlated to the bin span in these example samples ( Figure 4). Identified plastid bins were mapped to a reference Trebouxia lichen photobiont genome ( Martínez-Alberola et al., 2020) with minimap2 ( Li, 2018) and coverage calculated with bedtools ( Quinlan & Hall, 2010).

Figure 4. Recovered Trebouxia plastid metagenomic assembled genomes (MAGs) from lichen metagenomes.

Samples with lower completeness plastid MAGs correlate with smaller bin sizes suggesting incomplete plastid genome content present in metagenomic data.

Discussion

Currently, metagenomic characterization of plastid genomes is limited to the identification of plastid contigs using Tiara. However, only identifying plastid contigs prevents full disentanglement of metagenomic datasets, including the potential presence of multiple plastid genomes arising from different sources. plastiC uniquely provides users with the opportunity to characterize plastid genomes including separation into independent plastid genomes and source classification to inform on present eukaryotic photosynthetic organisms. Notably, this is the first tool to provide a completeness estimate for plastid genomes from metagenomic sources. The high copy number and small genome size of plastids may permit easier recovery and serve as a proxy for the identification of the presence of eukaryotes which may not have sufficient genomic signal to be recovered as a high-quality eukaryotic MAG, and facilitate targeted coassembly of samples, based on common plastid composition. Furthermore, plastiC offers the potential to expand reference databases with metagenomic derived plastids. plastiC will support characterization of plastids in diverse ecological contexts including, for example, symbiotic interactions (e.g., lichens as demonstrated in the Use cases), complex marine ecosystems and characterisation of components of herbivore diets.

Funding Statement

This work was supported by EMBL (European Molecular Biology Laboratory) core funds. The authors were supported by Wellcome [2205040].

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

Underlying data

RefSeq accession numbers for model training and testing are available on the plastiC GitHub repository: https://github.com/Finn-Lab/plastiC/tree/main/scripts/model_training/accession_lists.

To demonstrate the effectiveness of plastiC in the recovery and characterisation of plastid genomes from metagenomic samples, we used publicly available lichen reads from the European Nucleotide Archive (ENA), available under the project accession: PRJNA646656.

Software availability

Source code available from: https://github.com/Finn-Lab/plastiC

Archived source code at time of publication: https://doi.org/10.5281/zenodo.7950626 ( Cameron, 2023).

License: GPL-3.0-only