PaperlayerView exact revision on arXiv

Source-linked AI summary

Bacteriophage classification for assembled contigs using Graph Convolutional Network

Jiayu Shang, Jingzhe Jiang, Yanni Sun

arXiv:2102.03746v2q-bio.GNcs.LG

TL;DR

Phage contigs are accumulating faster than taxonomic classification can keep pace, amid limited labeled genomes and ambiguous sequence matches. PhaGCN combines CNN sequence features and protein-similarity graph edges in a semi-supervised GCN. The paper reports improved performance for novel phage classification and favorable comparisons with other tools.

  • Problem

    Rapidly growing phage-like contig collections are difficult to classify because known taxa are limited and protein homologs can cross taxonomic families.

    Source: §1 · p. 2

  • Method

    PhaGCN combines CNN-encoded contig sequences and protein-similarity and gene-sharing graph edges, then trains a GCN using labeled and unlabeled samples.

    Source: §1 · p. 2, §2 · pp. 3–4

  • Results

    PhaGCN outperforms other popular phage-classification methods and renders better performance for novel phage classification.

    Source: §1.1 · p. 3, §4 · p. 14

  • Takeaways & Limitations

    The authors present PhaGCN as a method for improving classification of novel phages.

    Source: §4 · p. 14

    PhaGCN remains difficult to apply to classes with fewer than 10 training samples.

    Source: §4 · p. 14

Abstract

from arXiv · show

Motivation: Bacteriophages (aka phages), which mainly infect bacteria, play key roles in the biology of microbes. As the most abundant biological entities on the planet, the number of discovered phages is only the tip of the iceberg. Recently, many new phages have been revealed using high throughput sequencing, particularly metagenomic sequencing. Compared to the fast accumulation of phage-like sequences, there is a serious lag in taxonomic classification of phages. High diversity, abundance, and limited known phages pose great challenges for taxonomic analysis. In particular, alignment-based tools have difficulty in classifying fast accumulating contigs assembled from metagenomic data. Results: In this work, we present a novel semi-supervised learning model, named PhaGCN, to conduct taxonomic classification for phage contigs. In this learning model, we construct a knowledge graph by combining the DNA sequence features learned by convolutional neural network (CNN) and protein sequence similarity gained from gene-sharing network. Then we apply graph convolutional network (GCN) to utilize both the labeled and unlabeled samples in training to enhance the learning ability. We tested PhaGCN on both simulated and real sequencing data. The results clearly show that our method competes favorably against available phage classification tools.

1 Introduction

Metagenomic sequencing is rapidly revealing phage-like contigs, but their taxonomic classification remains difficult because known taxa are limited and protein homologs can be shared across families. PhaGCN combines sequence and protein-similarity features in a semi-supervised graph model to classify contigs.

1 Introduction

  • RefSeq phage records rose from 1,468 in 2015 to 3,852 in 2020, yet known phages remain only a fraction of the virome and classification is challenging.Metagenomic sequencing and assembly have produced phage-like contigs from varied samples.Source: §1 · pp. 1–2
  • Limited labeled genomes and cross-family protein homologs complicate classification, while more than 1,800 new Caudovirales sequences found in 2020 lacked family assignments.The paper reports over 7,616 proteins shared across Caudovirales families and over 18,970 highly similar cross-family protein pairs.Source: §1 · p. 2
  • PhaGCN classifies assembled contigs using CNN-learned sequence features, protein similarity, and gene-sharing features in a semi-supervised GCN.Both labeled and unlabeled sequences are used during training.Source: §1 · p. 2

1.1 Related work

  • Alignment-based tools use nucleotide or protein homology, but shared genes can yield ambiguous or higher-rank labels, and novel proteins or weak reference alignments limit classification.The paper notes that sequence similarity alone may not provide ideal taxonomic resolution.Source: §1.1 · p. 2
  • Gene-sharing network methods such as vConTACT 2.0 lose accuracy on short contigs because few proteins may yield too few valid network edges for clustering.Unclustered contigs may receive no labels.Source: §1.1 · p. 2
  • PhaGCN combines alignment-based and learning-based strengths, and the reported experiments show it outperforms other popular phage-classification methods.The approach uses DIAMOND-derived similarities for edge construction, CNN-encoded nucleotide information, and labeled and unlabeled samples in GCN training.Source: §1.1 · p. 3

2 Methods

PhaGCN formulates phage-contig classification as semi-supervised learning and combines CNN-encoded sequence features with protein-similarity and gene-organization edges in a knowledge graph for GCN classification.

2 Methods

  • Semi-supervised learning combines labeled reference genomes with unlabeled contigs, while GCN uses graph structure to learn features for classifying unlabeled samples.The graph convolution aggregates information from neighboring nodes.Source: §2 · p. 3

2.1 Using CNN to encode input sequences in the knowledge graph

  • A CNN trained on reference genomes encodes contig segments as numerical vectors that become knowledge-graph node features.The CNN uses skip-gram embeddings, convolution, and dense layers; contigs with multiple 2-kbp segments are represented by the average of their segment outputs.Source: §2.1 · pp. 4–5

2.2 Construction of the edge in GCN

  • Knowledge-graph edges connect sequences likely to share a taxonomic group, using shared protein clusters and average protein similarity.Protein clusters are built from six-frame contig translations aligned with reference proteins, then identified with Markov clustering.Source: §2.2 · p. 5, §2.2.1 · p. 5

2.2.1 Construction of the protein cluster

  • DIAMOND alignments of six-frame contig translations to reference proteins are clustered with MCL, retaining clusters containing at least two proteins.When a contig has alignments in multiple reading frames, only the best frame is kept.Source: §2.2.1 · p. 5

2.2.2 Definition of the edges

  • Pweight measures the expected number of sequence pairs sharing at least c protein clusters, while Eweight is the mean E-value of shared-protein alignments.Edges are created when Pweight plus Eweight exceeds the default threshold τ = 1; contigs without a graph edge receive no prediction.Source: §2.2.2 · p. 6

2.3 The GCN model

  • The GCN trains on graph features from labeled and unlabeled samples but calculates loss only on labeled samples, then uses SoftMax outputs to label unlabeled contigs.The model uses two graph convolutional layers and a dense layer.Source: §2.3 · p. 6, Figure 3 · p. 7

3.1 Data and performance metrics

PhaGCN is evaluated for family-level classification within Caudovirales, using simulated and real sequencing data and metrics that weight classes equally.

3.1 Data and performance metrics

  • The evaluation focuses on family-level classification among 3,639 Caudovirales genomes from eight families, since lower ranks have few genomes per group.Caudovirales accounts for 95.8% of RefSeq phage reference genomes.Source: §3.1 · p. 7

3.1.1 Data and experiment design

  • Experiments use randomly sampled reference-genome contigs, contigs assembled from simulated reads, and assembled contigs from two real sequencing datasets.The real-data evaluation includes virus-like contigs from 71 oyster metagenomic datasets.Source: §3.1.1 · p. 7
  • Macro-accuracy, macro-recall, and macro-precision are averaged independently across classes, giving each class equal weight.The compared pipelines are vConTACT 2.0, POGs, ClassiPhage, and PhaGCN's CNN model.Source: §3.1.1 · p. 8
  • Genome masking removes test genomes from training so test contigs represent phages without reference genomes in the training data.The design is intended to prevent shared genomes from making performance reflect memorization rather than learning.Source: §3.1.1 · p. 8

3.2 Experiments using simulated contigs

On randomly sampled contigs, PhaGCN has the highest classified-contig count and classification accuracy, and its precision and recall remain strong across contig lengths.

3.2 Experiments using simulated contigs

  • PhaGCN outperforms other tools across contig-length ranges, and its performance is stable as contig length changes.All pipelines improve with longer contigs; the authors attribute this to longer contigs containing more proteins.Source: §3.2 · p. 8
  • Both precision and recall are better with PhaGCN's GCN than with its CNN alone, which the authors say shows the knowledge graph enhances learning.Source: §3.2 · p. 8
  • PhaGCN produces the most predictions with the highest classification accuracy on randomly sampled contigs.The comparison includes contigs across three length ranges and complete genomes, using the same train/test construction for the other pipelines.Source: §3.2 · p. 8
  • vConTACT 2.0 predicts only a small number of contigs in Podoviridae and Siphoviridae, reducing its macro-precision.Families with no predictions receive precision 0.Source: §3.2 · p. 8

3.3 Experiments using simulated reads

PhaGCN outperforms the other tools on contigs assembled from simulated reads, including short contigs; the evaluation uses newly released Escherichia coli phages as test species.

3.3 Experiments using simulated reads

  • The read-based test simulates reads from 99 newly released 2020 Escherichia coli phage species, then assembles them with MEGAHIT.Contigs over 2 kbp with taxon-specific alignments and query coverage above 85% were retained, yielding 301 comparison inputs.Source: §3.3 · pp. 8–9
  • The random-contig figures compare precision and recall by contig-length range and show classified-contig percentage alongside macro-accuracy.Each length range uses 1,600 randomly sampled contigs from 20 genomes across eight families, averaged over three sets.Source: Figure 4 · p. 9, Figure 5 · p. 9
  • PhaGCN outperforms other tools across contig lengths, with over 80% accuracy at 2–4 kbp and 100% accuracy above 8 kbp.These contigs were assembled from simulated reads containing sequencing errors.Source: §3.3 · p. 9, Figure 6 · p. 10, Figure 7 · p. 10
  • The reported performance is slightly better than in the earlier experiment, with the authors noting that this test includes only newly released Escherichia coli phages.Source: §3.3 · p. 10

3.4 Running time comparison

PhaGCN is not the fastest compared method; pairwise sequence alignments are its most resource-demanding component, and the authors identify a possible optimization.

3.4 Running time comparison

  • PhaGCN is not the fastest program, and sequence alignment is its most resource-demanding component.Alignments produce protein clusters and define edges in the knowledge graph.Source: §3.4 · pp. 10–11
  • The authors suggest reducing pairwise alignments by producing a database of protein clusters.Source: §3.4 · p. 11
  • The reported elapsed-time comparison measures each tool's average time to classify 100 contigs on an eight-core Intel Xeon Gold 6258R CPU.Source: Table 2 · p. 11

3.5 Experiments on real sequencing data

The authors evaluated PhaGCN on assembled contigs from two real NCBI SRA datasets, comparing its classifications with three tools. They also examined an unaligned contig and the effect of PhaGCN on dataset composition analysis.

3.5 Experiments on real sequencing data

  • MEGAHIT contigs were labeled using NCBI SRA species; one bacterial contig per dataset was excluded, leaving 43 contigs per dataset for classification.Table 3 lists the reference species used as ground truth, and Table 4 lists the assembled contigs and their derived family labels.Source: §3.5 · p. 11, Table 3 · p. 11, Table 4 · p. 11
  • PhaGCN achieved 100% accuracy on both datasets, outperforming vConTACT 2.0, POGs, and ClassiPhage.The comparison is shown in Fig. 8.Source: §3.5 · p. 11, Figure 8 · p. 12
  • For a contig lacking ground truth, only PhaGCN assigned the correct Siphoviridae label, although that reference genome was not in its training data.BLAST against the Nucleotide Collection identified the contig as a Siphoviridae phage.Source: §3.5 · p. 11
  • PhaGCN improved composition analysis of dark matter in both datasets, as shown by the before-and-after comparison in Fig. 9.Source: §3.5 · p. 12, Figure 9 · p. 12

3.6 Phage classification in contigs produced from oyster metagenomic data

PhaGCN was applied to contigs from oyster metagenomic samples and compared with vConTACT 2.0. It predicted substantially more contigs, while ground truth was unavailable for this large-scale dataset.

3.6 Phage classification in contigs produced from oyster metagenomic data

  • Of 22,966 contigs longer than 4 kbp retained as input, PhaGCN assigned 17,199 to Caudovirales.The contigs were filtered to remove those alignable to bacteria, archaea, or eukaryota.Source: §3.6 · p. 13
  • About 74.8% of contigs were predicted by PhaGCN versus about 1.1% by vConTACT 2.0; vConTACT predictions were a subset of PhaGCN’s.Table 5 summarizes the prediction results.Source: §3.6 · p. 13, Table 5 · p. 13
  • Ground truth was unavailable for this large-scale dataset, although the authors report over 92% accuracy for contigs longer than 4 kbp.Source: §3.6 · p. 13

3.7 Extension of PhaGCN

PhaGCN was extended to three families outside Caudovirales and tested for rejecting non-Caudovirales contigs. The results support family extension and targeted detection within the tested settings.

3.7 Extension of PhaGCN

  • With at least 10 genomes per added family, leave-one-genome-out tests classified almost all contigs correctly across Rudiviridae, Microviridae, and Inoviridae.The classification results are shown in Fig. 10(A).Source: §3.7 · p. 13, Figure 10 · p. 13
  • Only 3 of 42,180 non-Caudovirales contigs were accepted by PhaGCN, supporting its use for targeted phage detection.The rejection result is shown in Fig. 10(B).Source: §3.7 · p. 13, Figure 10 · p. 13
  • Phages outside the training families usually lack graph edges to its nodes, so PhaGCN does not misclassify them into Caudovirales.Source: §3.7 · p. 13

4 Discussion

The discussion attributes PhaGCN’s performance on novel phages to combining learned sequence features with protein-based graph edges and semi-supervised learning. It also identifies edge weighting and small families as areas for improvement.

4 Discussion

  • Ambiguous or absent reference matches can lower alignment-based classification accuracy for highly divergent or novel phages.The discussion also notes that alignment-based approaches depend heavily on the reference database.Source: §4 · pp. 13–14
  • PhaGCN combines automatically learned nucleotide features in graph nodes with protein-based alignment edges, then uses semi-supervised GCN training on labeled and unlabeled data.Source: §4 · p. 14
  • The authors identify equal-probability edge weighting as a simplifying assumption and note that classification remains difficult for families with fewer than 10 training samples.Future work also includes extending phage detection and incorporating bacteria into the knowledge network for host detection.Source: §4 · p. 14
Loading 2102.03746v2…