Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GraSSRep: Graph-Based Self-Supervised Learning for Repeat Detection in Metagenomic Assembly

Ali Azizpour, Advait Balaji, Todd J. Treangen, Santiago Segarra

TL;DR

GraSSRep addresses repeat detection in metagenomic assembly by leveraging the assembly unitig graph with a self‑supervised graph neural network framework. It builds a unitig graph, extracts sequencing and graph features, generates high‑confidence pseudo‑labels, and combines a GNN embedding with a random forest classifier to propagate labels to unlabeled unitigs, followed by a fine‑tuning step that uses sequencing features to correct outliers. Across simulated and Shakya metagenomes, GraSSRep demonstrates robustness to repeat length, copy number, and coverage variations, and significantly outperforms existing tools in precision and recall, aided by learnable graph features and adaptive label refinement. The approach promises practical impact by enabling more accurate repeat detection, which can improve assembly quality and downstream comparative analyses, and it can be extended to real datasets and other genomic marker discovery tasks.

Abstract

Repetitive DNA (repeats) poses significant challenges for accurate and efficient genome assembly and sequence alignment. This is particularly true for metagenomic data, where genome dynamics such as horizontal gene transfer, gene duplication, and gene loss/gain complicate accurate genome assembly from metagenomic communities. Detecting repeats is a crucial first step in overcoming these challenges. To address this issue, we propose GraSSRep, a novel approach that leverages the assembly graph's structure through graph neural networks (GNNs) within a self-supervised learning framework to classify DNA sequences into repetitive and non-repetitive categories. Specifically, we frame this problem as a node classification task within a metagenomic assembly graph. In a self-supervised fashion, we rely on a high-precision (but low-recall) heuristic to generate pseudo-labels for a small proportion of the nodes. We then use those pseudo-labels to train a GNN embedding and a random forest classifier to propagate the labels to the remaining nodes. In this way, GraSSRep combines sequencing features with pre-defined and learned graph features to achieve state-of-the-art performance in repeat detection. We evaluate our method using simulated and synthetic metagenomic datasets. The results on the simulated data highlight our GraSSRep's robustness to repeat attributes, demonstrating its effectiveness in handling the complexity of repeated sequences. Additionally, our experiments with synthetic metagenomic datasets reveal that incorporating the graph structure and the GNN enhances our detection performance. Finally, in comparative analyses, GraSSRep outperforms existing repeat detection tools with respect to precision and recall.

GraSSRep: Graph-Based Self-Supervised Learning for Repeat Detection in Metagenomic Assembly

TL;DR

GraSSRep addresses repeat detection in metagenomic assembly by leveraging the assembly unitig graph with a self‑supervised graph neural network framework. It builds a unitig graph, extracts sequencing and graph features, generates high‑confidence pseudo‑labels, and combines a GNN embedding with a random forest classifier to propagate labels to unlabeled unitigs, followed by a fine‑tuning step that uses sequencing features to correct outliers. Across simulated and Shakya metagenomes, GraSSRep demonstrates robustness to repeat length, copy number, and coverage variations, and significantly outperforms existing tools in precision and recall, aided by learnable graph features and adaptive label refinement. The approach promises practical impact by enabling more accurate repeat detection, which can improve assembly quality and downstream comparative analyses, and it can be extended to real datasets and other genomic marker discovery tasks.

Abstract

Repetitive DNA (repeats) poses significant challenges for accurate and efficient genome assembly and sequence alignment. This is particularly true for metagenomic data, where genome dynamics such as horizontal gene transfer, gene duplication, and gene loss/gain complicate accurate genome assembly from metagenomic communities. Detecting repeats is a crucial first step in overcoming these challenges. To address this issue, we propose GraSSRep, a novel approach that leverages the assembly graph's structure through graph neural networks (GNNs) within a self-supervised learning framework to classify DNA sequences into repetitive and non-repetitive categories. Specifically, we frame this problem as a node classification task within a metagenomic assembly graph. In a self-supervised fashion, we rely on a high-precision (but low-recall) heuristic to generate pseudo-labels for a small proportion of the nodes. We then use those pseudo-labels to train a GNN embedding and a random forest classifier to propagate the labels to the remaining nodes. In this way, GraSSRep combines sequencing features with pre-defined and learned graph features to achieve state-of-the-art performance in repeat detection. We evaluate our method using simulated and synthetic metagenomic datasets. The results on the simulated data highlight our GraSSRep's robustness to repeat attributes, demonstrating its effectiveness in handling the complexity of repeated sequences. Additionally, our experiments with synthetic metagenomic datasets reveal that incorporating the graph structure and the GNN enhances our detection performance. Finally, in comparative analyses, GraSSRep outperforms existing repeat detection tools with respect to precision and recall.
Paper Structure (19 sections, 7 equations, 7 figures)

This paper contains 19 sections, 7 equations, 7 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Step 1: Unitig graph construction
  4. Step 2: Feature extraction
  5. Step 3: Selection of the training nodes
  6. Step 4: Unitig classification via self-supervised learning
  7. Step 5: Fine-tuning the labels
  8. Experimental setup
  9. Datasets
  10. Assembly
  11. Method design and hyperparameter choices
  12. Results
  13. Evaluation on varying repeat characteristics
  14. Ablation study of the steps of the algorithm
  15. Comparison with existing repeat detection methods
  16. ...and 4 more sections

Figures (7)

  • Figure 1: Repeat positions within the assembly graph.
  • Figure 2: Overview of GraSSRep. (a) Reads are assembled into unitigs forming the nodes of the unitig graph. Edges are constructed based on the read mapping information. Also, feature vectors are computed for each unitig. (b) Unitigs with distinctive sequencing features are selected as training nodes and labeled. (c) The unitig graph is input into a GNN. Embeddings are generated for each unitig and combined with the initial features. A random forest classifier predicts labels for all unitigs based on the augmented feature vectors. (d) Sequencing features are employed to identify outliers within each predicted class, leading to the reassignment of their class labels.
  • Figure 3: Assessing the method across various repeat characteristics. (a) The model remains stable in metrics even with increasing repeat length, especially beyond the outer distance of read pairs (500 base pairs). (b) The method is robust to the copy number variation, consistently achieving an F1-score above 90%. (c) Higher sequencing coverage improves the model's performance.
  • Figure 4: (a) Progression of the method's performance throughout the different steps, highlighting the effectiveness of each step in improving repeat detection. We also test the impact of excluding the GNN embeddings. (b) High importance of GNN-generated embeddings in RF classification.
  • Figure 5: GraSSRep compared to the other repeat detection methods, demonstrating superior performance.
  • ...and 2 more figures