SCoV: a frame-gated cross-modal model for identifying viral sequences in short metagenomic fragments

Technical Summary: SCoV – A Frame-Gated Cross-Modal Model for Viral Sequence Identification

Problem Statement
Accurate identification of short viral reads within metagenomic data is a critical bottleneck in viromics and pathogen surveillance. Existing approaches face a persistent trade-off between accuracy and efficiency due to three inherent challenges: the extreme fragmentation of metagenomic data where viral sequences are a vanishingly small fraction; the limited discriminative signals contained within short fragments (e.g., 300–500 bp); and the ambiguity in predicting Open Reading Frames (ORFs) within these short sequences, which often leads to truncation or unreliable gene annotation. Furthermore, many existing methods rely heavily on reference databases or large pretrained models (e.g., DNABERT-2), which increase computational and memory costs, or they fail to explicitly model reading-frame uncertainty during the encoding process.

Methodology
The authors propose SCoV (Short-fragment Cross-modal viral sequence identifier), a compact, dual-branch neural architecture designed for end-to-end joint encoding of nucleotide sequences and six-frame codon embeddings without requiring explicit ORF annotation.

1. Dual-Branch Architecture:

   • Nucleotide Branch: Utilizes dilated convolutions with parallel kernels (sizes 5 and 15) and residual blocks to capture local sequence motifs and patterns at different scales.
   • Six-Frame Codon Branch: Encodes the input sequence into six candidate reading frames (three forward, three reverse complement) without explicit ORF prediction. These frames are processed by a shared Bidirectional Gated Recurrent Unit (BiGRU) to extract potential coding semantics.

2. Frame-Gated Cross-Modal Interaction (CIA):

   • A novel attention module dynamically fuses representations from the nucleotide and codon branches.
   • Frame Priors: The model calculates a "frame-gating weight" based on the proportion of stop codons in each reading frame. Frames with fewer stop codons are assigned higher plausibility.
   • Attention Mechanism: These weights are incorporated into the attention logits ($\log g$), allowing the model to adaptively favor candidate reading frames that are biologically more plausible while maintaining position-level interaction between nucleotide motifs and codon semantics.

3. Classification and Training:

   • The fused representation is passed through a classifier to output binary viral-versus-host probabilities.
   • The model is trained using a composite loss function combining label-smoothed cross-entropy and supervised contrastive loss to enhance discriminative representation.
   • SCoV is designed to be memory-efficient, utilizing approximately 0.436 million parameters.

Key Contributions

• Explicit Handling of Frame Uncertainty: Unlike prior methods that merge codon features after global pooling or rely on single overlapping codon sequences, SCoV retains six independent frame-specific representations and uses a frame-gated mechanism to dynamically weight them based on stop-codon proportions.
• Reference-Free and Compact: The model operates without explicit ORF prediction or large pretrained backbones, making it suitable for resource-constrained environments.
• Cross-Modal Fusion: It introduces a specific mechanism to inject frame priors into the cross-modal attention space, enabling dynamic fusion of nucleotide and codon features under reading-frame uncertainty.

Results
SCoV was evaluated on internal test sets (500 bp and 300 bp fragments) and three independent external metagenomic datasets (Antarctic seawater, agricultural soil, and human gut).

• Internal Performance:

  • On the 500 bp test set, SCoV achieved an F1-score of 0.9184 and an AUROC of 0.9734, outperforming the strongest baseline (PPR-Meta) by 0.0513 in F1-score.
  • On the 300 bp test set, SCoV achieved an F1-score of 0.8836 and an AUROC of 0.9526, improving upon PPR-Meta by 0.0487 in F1-score.
  • All improvements were statistically significant ($p < 0.001$ after Holm–Bonferroni correction).

• External Generalization:

  • SCoV achieved the highest F1-scores across all three external habitats (seawater, soil, gut) for both fragment lengths, demonstrating robust cross-habitat generalization.

• Ablation Studies:

  • Removing the cross-modal interaction (w/o CIA) reduced the 500 bp F1-score by ~2.1 percentage points.
  • Removing the codon branch entirely (w/o Codon) reduced the 500 bp F1-score by ~3.98 percentage points, confirming the complementary value of six-frame codon representations.

• Resource Efficiency:

  • With 0.436M parameters and a peak memory footprint of 35.9 MB, SCoV is the most compact among the top-performing methods evaluated, though its inference throughput is lower than convolution-based baselines due to the BiGRU encoding.

Significance and Claims
The paper claims that SCoV offers a favorable performance–compactness trade-off for short viral-fragment classification. By explicitly modeling reading-frame uncertainty through a frame-gated cross-modal attention mechanism, it addresses the limitations of existing tools that either ignore frame ambiguity or rely on computationally expensive pretrained models. The authors position SCoV as a practical, memory-efficient option for large-scale metagenomic prescreening, particularly in scenarios where ORF prediction is unreliable, discriminative signals are scarce, or GPU memory is limited. The study concludes that the integration of six-frame codon semantics with nucleotide motifs via dynamic, frame-aware fusion is essential for improving viral identification in short metagenomic fragments.