Metagenomics Binning of Long Reads Using Read-Overlap Graphs

Abstract

Metagenomics sequencing enables the direct study of microbial communities revealing important information such as taxonomy and relative abundance of species. Metagenomics binning facilitates the separation of these genetic materials into different taxonomic groups. Moving from second-generation sequencing to third-generation sequencing techniques enables the binning of reads before assembly thanks to the increased read lengths. The limited number of long-read binning tools that exist, still suffer from unreliable coverage estimation for individual long reads and face challenges in recovering low-abundance species. In this paper, we present a novel binning approach to bin long reads using the read-overlap graph. The read-overlap graph (1) enables a fast and reliable estimation of the coverage of individual long reads; (2) allows to incorporate the overlapping information between reads into the binning process; (3) facilitates a more uniform sampling of long reads across species of varying abundances. Experimental results show that our new binning approach produces better binning results of long reads and results in better assemblies especially for recovering low abundant species. The source code and a functional Google Colab Notebook are available at https://www.github.com/anuradhawick/oblr.

Appendices

A Dataset Information

Tables 5 and  7 demonstrate the simulated and real dataset information respectively. Note that Table 5 and 6 tabulate the coverages used for simulation using SimLoRD [29] while Table 7 indicate abundances from the dataset sources.

Table 5. Information of simulated datasets.

Table 6. Information of simulated dataset containing 50 species.

Table 7. Information of real datasets.

B Interpretation of AMBER Per-bin F1-Score

The binning evaluations are presented using Precision, Recall and F1 score. Furthermore, stricter evaluations are presented using AMBER [21] for MetaBCC-LR, LRBinner and OBLR. This section explains the evaluation metrics in detail and discuss as to why AMBER evaluations are poor in some cases where number of bins predicted is further away from the actual number of species in the dataset. Note that the bin assignment matrix a can be presented as $M\times N$, illustrated in Table 8. Note that $N=5$ and $M=7$.

Table 8. Binning matrix

Recall is computed for each species, by taking the largest assignment to a bin. Precision is computed per bin taking the largest assignment of the bin to a given species. In contrast, AMBER uses purity and completeness to compute the per-bin F1 score using the following equations, for each bin b.

The true positives are computed using the majority species in a given bin. Because of this, if a bin appears as a result of a false bin split (1% reads), the Completeness of the bin will be very low as the majority of it (approximately 1%) according to AMBER evaluation. In comparison, the recall of the species using Eq. 5 will report 99% since 99% of the reads are in a single bin despite having the false bin split. Similarly, the false split of the bin will report a greater precision as long as the bin has no other species mixed according to Eq. 4. Consider the following running example.

Example 1. Suppose Species 1 has $a_{11}=99$ and $a_{16}=1$ with rest of the row having no reads and Bin 1 and 6 has no reads from another species. Purity in this case will be 100% for both bins 1 and 6 while completeness will be 99% and 1% respectively. F1-score will be 99.5% and 1.98% with average being very low at 50.7%. Recall will be 99% for Species 1 with 100% precision on both bins 1 and 6 since there are no impurities in each bin, thus, F1-score is 99.5% for each bin.

Example 2. Suppose Bin 2 has $a_{22}=100$ and $a_{32}=20$, with two species 2 and 3, with no other contaminants and species 2 and 3 are fully contained in the bin. Now, the purity of the bin is 83.33% and completeness is 83.33%, hence, F1-score is 83.33%. Recall for species 2 and 3 will be 100% since it is not broken into multiple bins. However, the precision for Bin 2 will be 83.33%, hence a F1 score of 90.91%.

This means, AMBER penalize whenever a species is broken into pieces across bins while not significantly penalizing bin mergers between large bins and smaller bins. This is because, dominant species in the bin will determine the purity and completeness.