CLMB: Deep Contrastive Learning for Robust Metagenomic Binning

Abstract

The reconstruction of microbial genomes from large metagenomic datasets is a critical procedure for finding uncultivated microbial populations and defining their microbial functional roles. To achieve that, we need to perform metagenomic binning, clustering the assembled contigs into draft genomes. Despite the existing computational tools, most of them neglect one important property of the metagenomic data, that is, the noise. To further improve the metagenomic binning step and reconstruct better metagenomes, we propose a deep Contrastive Learning framework for Metagenome Binning (CLMB), which can efficiently eliminate the disturbance of noise and produce more stable and robust results. Essentially, instead of denoising the data explicitly, we add simulated noise to the training data and force the deep learning model to produce similar and stable representations for both the noise-free data and the distorted data. Consequently, the trained model will be robust to noise and handle it implicitly during usage. CLMB outperforms the previous state-of-the-art binning methods significantly, recovering the most near-complete genomes on almost all the benchmarking datasets (up to 17% more reconstructed genomes compared to the second-best method). It also improves the performance of bin refinement, reconstructing 8–22 more high-quality genomes and 15–32 more middle-quality genomes more than the second-best result. Impressively, in addition to being compatible with the binning refiner, single CLMB even recovers on average 15 more HQ genomes than the refiner of VAMB and Maxbin on the benchmarking datasets. On a real mother-infant microbiome dataset with 110 samples, CLMB is scalable and practical to recover 365 high-quality and middle-quality genomes (including 21 new ones), providing insights into the microbiome transmission. CLMB is open-source and available at https://github.com/zpf0117b/CLMB/.

Appendices

5 Appendix

A Figures

Fig. 7.

Performance of different clustering algorithms based on five datasets. Orange: DBSCAN Algorithm. Green: Exclude the outlier using DBSCAN first and cluster the others points using minibatch k-means algorithm. Red: Iterative medoid algorithm, which is developed by [10] and used by CLMB. (Color figure online)

Fig. 8.

Performance of CLMB with different samples. For any given number of samples, samples were randomly drawn 3 times and executed independently. For “single-sample”, all the samples were run independently. We note that for increasing number of samples, the random subsets chosen is not independent, due to only having 9 (Urog) or 10 (Airways, GI, Skin, Oral) samples in total. Orange: Multi-split workflow of CLMB, Green: Single sample workflow of CLMB. (Color figure online)

Fig. 9.

Performance of CLMB with different k-mer length on different datasets. It is assessed by the number of reconstructed NC strains. The performance varies among the datasets.

Table 1. Number of genomes at the strain level reconstructed with a precision of at least 95%

Table 2. Number of genomes at the strain level reconstructed with a precision of at least 95%

Table 3. Number of genomes at the species level reconstructed with a precision of at least 95%

B Tables

Table 4. Number of genomes at the genus level reconstructed with a precision of at least 95%

C Methods

In this section, we show the methods and experiments in our research.

1.1 C.1 Feature Calculation of TNFs and Abundance

We use the same approach to calculate TNFs and abundance as the previous work [10]. For each contig, we count the frequencies of each tetramer with definite bases, and, to satisfy statistical constraints, project them into a 103-dimensional independent orthonormal space to obtain TNFs [6]. As a result, the TNFs for each contig are a 103-dimensional numerical vector. We also count the number of individual reads mapped to each contig. More specifically, a read mapped to n contigs counts 1/n towards each. The read counts are normalized by sequence length and total number of mapped reads, which generates the abundance value in reads per kilobase sequence per million mapped reads (RPKM). The resulted abundance for each contig is a s-dimensional numerical vector, where s is the number of samples. TNFs are normalized by z-scaling each tetranucleotide across the sequences, and abundance are normalized across samples.

1.2 C.2 Benchmarking

CLMB and VAMB [10] were run with default parameters with multi-split enabled. MetaBAT2 [8] was run with setting minClsSize = 1 and other parameters as default. MaxBin2 [9] was run with default parameters. The benchmarking results were calculated using benchmark.py script implemented by [10]. The mapping of the recovered genomes to the reference genomes was the intermediate result⁶ of benchmark.py script. FastANI [45] with default parameters was used to calculate ANI between the reference genomes. For the binning refinement experiment, we use metaWRAP bin_refinement API [32, 33] with parameters –c 50 and –x 10, indicating we keep the genomes qualifying $completeness>50\%$ and $contamination<10\%$. The completeness and contamination of the genomes recovered by the bins are calculated using CheckM [34] with default parameters. We use the pipeline integrated in MetaGEM [11] for binning refinement experiment.

1.3 C.3 Data Fusion Experiment

We define the feature data as the raw data, and obtained the projected data by projecting the feature data to 32-dimension space using PCA. For the CLMB-encoded data, we obtained them by encoding the feature data to 32-dimension space with the deep contrastive learning framework. We assess the performance of these data by clustering them with the iterative medoid clustering and obtained the benchmarking results. All the experiments on CAMI2 datasets were run with default parameters with multi-split enabled, and the experiments on MetaHIT datasets was run with default parameters with multi-split disabled. For comparison to other clustering methods, we use MiniBatchKMeans (n_clusters = 750, batch_size = 4096, max_iter = 25, init_size = 20000, reassignment_ratio = 0.02) and DBSCAN (eps = 0.35, min_samples = 2) implemented by scikit-learn.

1.4 C.4 Binning of the Mother-Infant Transmission Dataset

We downloaded the sequencing datasets of selected mother-infant pairs (marked as 10001, 10002, 10003, 10005, 10006, 10007, 10008, 10009, 10015, 10019) using SRA Toolkit and filtered them based on quality using fastp [46]. Then, we assembled the short sequence reads into contigs using MEGAHIT [47, 48] and mapped the reads to the contigs using kallisto [49] in order to speed up this process for large datasets. The coabundance across samples can be subsequently calculated using kallisto quantification algorithm. With the assemblies and coabundances, we ran CLMB with default parameters and multi-split enabled. Then, we splited the fasta file into bins based on the result of clustering using create_fasta.py script. CheckM [34] on lineage specific workflow with default parameters was applied to the resulting bins to calculate the completeness and contamination, and only those with sufficient quality ($completeness\ge 50\%$, $contamination\le 5\%$) were considered for further analysis. Then, we use GTDB-tk [38] on for taxonomic assignment of each bins and phylogeny inference. We visualized the tree with iTOL [50].