BinChill: A Metagenomic Binning Ensemble Method

Oliver S. Bak, Marcus D. Jensen, Frederik M. Trudslev, Andreas Windfeld, Andre LamuriasIEEE Access 2023 · OpenReview

Abstract

The goal of metagenomic binning is to reconstruct genomes from a mixture of DNA sequences into genomic bins, which can be considered a clustering task. Multiple methods have been proposed for this task, such as distance-based metrics, machine learning, and ensemble approaches. We propose BinChill, a metagenomic ensemble method, based on the generic co-occurrence ensembler method, ACE. BinChill incorporates domain information in the form of Single-Copy Genes (SCG) with a co-occurrence strategy. This strategy combines multiple clustering partitions according to how often two items co-occur in the same cluster. BinChill was able to reconstruct more or equally as many high- and medium quality while having an equal or faster runtime than other metagenomics-specific methods on a smaller simulated dataset. On larger datasets, both simulated and real-world, BinChill outperformed other methods in reconstructing high-quality bins, at the cost of an increased processing time when compared to generic ensemble clustering algorithms. This is due to the domain-specific steps that our method implements. Our results show that the strengths of multiple partitions can be combined to generate a partition of higher quality.

Introduction: The Challenge of Metagenomic Binning

Metagenomics enables the direct sequencing of DNA from complex environmental samples, such as soil or gut microbiomes. However, the resulting data is a fragmented collection of DNA sequences known as contigs. The critical task of "binning" involves grouping these fragments back into their original organism-specific genomes.

While numerous binning algorithms exist, their performance varies significantly across different datasets. Researchers often turn to ensemble methods to synthesize results from multiple binners into a consensus. However, generic ensemble methods lack biological context, and existing metagenomic-specific ensemblers often struggle to balance genome completeness with contamination. BinChill addresses this by integrating biological constraints—specifically Single-Copy Genes (SCGs)—directly into the ensembling process.

The BinChill Workflow

BinChill is built upon the Adaptive Cluster Ensemble (ACE) framework but is heavily customized for biological applications. The workflow operates in three distinct stages:

The three-stage BinChill workflow: Relaxation (Stage 1), Bin Assignment (Stage 2), and Bin Refinement (Stage 3).
  1. Relaxation (Stage 1): The method takes multiple input partitions and calculates a similarity matrix. It iteratively merges the most similar clusters across different partitions to form "ensemble clusters."
  2. Bin Assignment (Stage 2): Contigs are assigned to these clusters based on membership similarity. A custom scoring function ($S_b$) utilizing SCGs is employed here. Because a single genome should ideally contain exactly one copy of essential genes, BinChill penalizes bins with multiple copies of the same SCG to minimize contamination.
  3. Bin Refinement (Stage 3): This novel stage evaluates the final bins to determine if they should be split, merged further, or maintained based on SCG scores and co-occurrence patterns.

Experimental Evaluation

The authors evaluated BinChill using both simulated and real-world datasets:

  • Strong100: A small simulated dataset (1,150 contigs) for parameter tuning.
  • CAMI2 Oral: A large-scale simulated dataset (201k contigs) representing a human oral microbiome.
  • Real-World Samples: Seven diverse datasets from wastewater treatment plants and soil samples, sequenced using Nanopore technology.

Key Insights and Results

BinChill demonstrates that domain knowledge serves as both a filter and a guide. By using SCGs as a "ground truth," the algorithm's scoring function prevents the creation of contaminated "megabins." A significant finding was that Stage 3 (Refinement) accounted for a 46% average improvement in the number of high-quality (HQ) bins, highlighting the necessity of a biologically-aware cleanup phase.

Performance Highlights:

  • Accuracy: On the CAMI2 oral dataset, BinChill reconstructed 333 HQ bins, significantly outperforming DAS-Tool, which identified only 31 HQ bins in the same test environment.
  • Real-World Efficacy: In wastewater samples (e.g., the "Aale" dataset), BinChill found 110 HQ bins, surpassing standalone binners like VAMB (64 HQ bins) and other ensemblers like MetaWrap.
  • Efficiency: While domain-specific steps increase processing time on massive datasets compared to generic algorithms, BinChill remains competitive or faster than other metagenomics-specific methods on smaller scales.

Limitations and Future Work

Despite its success, BinChill has areas for improvement. Currently, the similarity threshold ($\alpha_{1min}$) requires manual tuning. Future iterations aim to automate this based on dataset characteristics. Additionally, while SCGs are central to the method, integrating the SCG calculation (currently requiring external tools like CheckM) directly into the pipeline would enhance usability. Finally, incorporating expected genome size into the scoring function could further prevent over-merging in complex samples.

Summary updated

Abstract

The goal of metagenomic binning is to reconstruct genomes from a mixture of DNA sequences into genomic bins, which can be considered a clustering task. Multiple methods have been proposed for this task, such as distance-based metrics, machine learning, and ensemble approaches. We propose BinChill, a metagenomic ensemble method, based on the generic co-occurrence ensembler method, ACE. BinChill incorporates domain information in the form of Single-Copy Genes (SCG) with a co-occurrence strategy. This strategy combines multiple clustering partitions according to how often two items co-occur in the same cluster. BinChill was able to reconstruct more or equally as many high- and medium quality while having an equal or faster runtime than other metagenomics-specific methods on a smaller simulated dataset. On larger datasets, both simulated and real-world, BinChill outperformed other methods in reconstructing high-quality bins, at the cost of an increased processing time when compared to generic ensemble clustering algorithms. This is due to the domain-specific steps that our method implements. Our results show that the strengths of multiple partitions can be combined to generate a partition of higher quality.
BibTeX
@inproceedings{bak2023binchill,
  author = {Oliver S. Bak and Marcus D. Jensen and Frederik M. Trudslev and Andreas Windfeld and Andre Lamurias},
  title = {{BinChill: A Metagenomic Binning Ensemble Method}},
  booktitle = {IEEE Access},
  year = {2023},
  url = {https://openreview.net/forum?id=CCHJvv0hwj},
}
Figure