VirRep: a hybrid language representation learning framework for identifying viruses from human gut metagenomes

Abstract

Identifying viruses from metagenomes is a common step to explore the virus composition in the human gut. Here, we introduce VirRep, a hybrid language representation learning framework, for identifying viruses from human gut metagenomes. VirRep combines a context-aware encoder and an evolution-aware encoder to improve sequence representation by incorporating k-mer patterns and sequence homologies. Benchmarking on both simulated and real datasets with varying viral proportions demonstrates that VirRep outperforms state-of-the-art methods. When applied to fecal metagenomes from a colorectal cancer cohort, VirRep identifies 39 high-quality viral species associated with the disease, many of which cannot be detected by existing methods.

Keywords: Human gut metagenomes; Language representation learning; Virus identification.

Fig. 1

Schematic overview of the VirRep framework. a Workflow of VirRep for predicting viruses from metagenomes. b The detailed model architecture of VirRep. c The multi-step training strategy based on the pre-train-fine-tune paradigm to train VirRep

Fig. 2

Performance of VirRep and other methods on multiple human gut virome datasets. a–f The MCC values of VirRep and the eight popular methods on the GGCM-test dataset (a), IMGVR-gut dataset (b), DEVoC dataset (c), GPIC dataset (d), crAss-like phage dataset (e), and Lak-phage dataset (f) at various sequence length intervals. g The runtime of each method across the five sequence length intervals, where the average runtime is represented by the bar height and the error bars depict the 95% confidence intervals

Fig. 3

The ablation experiments of the two encoders, pre-training, and the first-stage fine-tuning. a Radar plots showing the MCC, precision, recall, and specificity achieved by the full implementation of VirRep, the semantic-encoder-based classifier, and the alignment-encoder-based predictor on the GGCM-test dataset. b The distribution of MCC values for VirRep with complete training through all stages (full training) compared to the version without pre-training, and the version with pre-training but without first-stage fine-tuning. Comparisons are shown across five sequence length intervals on the GGCM-test, IMGVR-gut, DEVoC, GPIC, crAss-phage and Lak-phage datasets. Significance levels are denoted as ****: $P\le 0.0001$, ***: $P\le 0.001$, **: $P\le 0.01$, *: $P\le 0.05$, based on the paired t-test. Each point represents a test dataset

Fig. 4

Comparing VirRep’s performance with that of other methods and method combinations on simulated metagenomic samples with varying viral proportions. a Precision-recall curves for VirRep, geNomad, and the six alignment-free methods at viral proportions of 5, 10, 50, and 90%. Numbers show the AUPRC (area under the precision-recall curve) values. b Average F1 score, precision and recall for VirRep, geNomad, and the five method combinations composed of VirSorter2 and one alignment-free method at viral proportions of 5, 10, 50, and 90%. c Average F1 score, precision and recall for VirRep, geNomad, and method combinations composed of VIBRANT and one alignment-free method at viral proportions of 5, 10, 50, and 90%. Error bar shows the 95% confidence intervals over 5 replicates

Fig. 5

Application of VirRep to 128 real human gut metagenomes from 74 colorectal cancer patients and 54 healthy controls. a The number of viral populations of each category (x-axis) obtained by each method (y-axis). The maximum value in each column is highlighted in bold red font. b The significance (log₁₀ transformed q-values) of viral populations (VPs) is given by the bar height. Horizontal line shows FDR at the level of 0.05. VPs with P < 0.001 and q < 0.05 are colored in dark gray, while others are colored in light gray. Shown are the top 90 significant VPs. c The average accuracy of tenfold cross-validation (repeated 10 times) of the logistic regression models versus the size of the marker set for each method. d Genome maps for viruses VP1279 and VP2811. e The phylogenetic tree of viruses VP1279 and VP2811