Can a graph neural network sort unknown phage DNA into families?
Linking DNA fragments into a graph by shared proteins and training a graph network on labelled and unlabelled fragments together classified novel phages more accurately than existing tools, especially for short fragments.
Source
Bacteriophage classification for assembled contigs using graph convolutional network
Study at a glance
- Design
- Computational / modelling — Semi-supervised GCN over a knowledge graph of contigs and reference genomes, compared with vConTACT 2.0, POGs, ClassiPhage and a CNN-only baseline on simulated and real contigs.
- N
- No single N: 3639 Caudovirales reference genomes; simulated tests of 4800 contigs per length range; 301 contigs from simulated reads; 43 contigs from two real datasets; 22 966 oyster contigs.
- Population
- Bacteriophage genomes and assembled metagenomic contigs from 8 Caudovirales families
- Outcome
- Family-level macro-accuracy, macro-precision and macro-recall
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors built PhaGCN: a CNN turns each DNA contig into a feature vector (a node), and edges link contigs and reference genomes that share many similar proteins. A graph convolutional network trained on this graph uses both labelled references and unlabelled contigs. Test genomes were always removed from training ('genome masking') so every test phage was effectively new, and performance was compared with three alignment- or clustering-based tools and a CNN-only model.
What they found
PhaGCN beat the other tools at every contig length, and adding the graph improved both recall and precision over using the CNN alone. On contigs assembled from simulated reads it stayed above 80% accuracy even for 2-4 kbp contigs and reached 100% for contigs over 8 kbp. On two real datasets it had 100% accuracy on 43 contigs, and in oyster metagenomes it labelled about 74.8% of contigs versus about 1.1% for vConTACT 2.0.
The limits
What it doesn't show
Only family-level classification was tested, mostly within one order, because lower ranks have too few genomes; families with fewer than 10 training genomes remain hard. Contigs with no edge to the graph get no prediction at all, and the real-data tests are small (43 contigs) with read-level NCBI labels as ground truth, while the oyster data had no ground truth. PhaGCN is also not the fastest tool because of its heavy alignment step.
Key terms
- Graph convolutional network (GCN)
- A neural network layer that updates each node's features by combining them with its neighbours' features in a graph.
- Semi-supervised learning
- Training that uses a small labelled set plus a large unlabelled set, letting structure in the unlabelled data help classification.
- Contig
- A continuous stretch of DNA assembled from overlapping sequencing reads.
- Genome masking
- Removing test genomes from the training data so a model is scored on genuinely unseen organisms rather than memorised ones.
- Macro-averaged metrics
- Accuracy, precision or recall computed per class and then averaged, so every class counts equally regardless of size.
Flashcards
0 of 10 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What was PhaGCN's key design choice?
Common questions
Why not just use a plain CNN on the DNA sequence?
The paper shows the CNN alone does worse; the graph adds protein-sharing relationships between contigs, which raised both precision and recall.
Why does vConTACT 2.0 score so low on precision?
It gives predictions for very few short contigs, often leaving whole families with no predictions, which pulls the macro-averaged precision down.
Could PhaGCN be fooled into labelling non-target phages?
Rarely: of 42 180 contigs from non-Caudovirales phages, only 3 were accepted, because unrelated phages seldom form edges to the graph.
More on Graph neural networks