Font Size: 

This paper presents MetaVIS, an alignment-free taxonomic identification system for 16S ribosomal RNA (rRNA) sequences, developed to address the computational bottleneck of BLASTn-based taxonomy assignment, which scales as O(N·L) per query and becomes impractical at metagenomic scale. MetaVIS reframes taxonomic classification as an image-recognition problem: variable-length DNA sequences are converted into fixed-size 64×64×3 RGB images via three-dimensional Tetrahedron Chaos Game Representation (CGR), then classified through a pipeline combining a MobileNetV3-Large backbone with spatial attention, taxonomy-aware ArcFace and NT-Xent contrastive fine-tuning, a centroid-indexed hierarchical beam search, and a MinHash/edit-distance leaf-fusion reranker. On a controlled Bacillaceae subset (18,766 sequences) the system attains 98% genus-level and 66% complex-aware species-level accuracy; on the full SILVA NR99 138.1 database (510,508 sequences) it attains 78% genus and 52% species accuracy at a mean query latency of 2.0 ms, a 126.6× speedup over BLASTn. A biologically principled complex-aware evaluation metric that accounts for the resolution limits of the 16S gene is proposed and validated.