Conjecture: Limited Genome Vocabulary Hampers Discriminative Codebook Learning

Establish whether the four-nucleotide genome vocabulary (A, T, C, G) fundamentally limits discriminative codebook learning in vector-quantized tokenizers for genomic sequences, thereby causing loss of fine-grained details in learned representations within the VQDNA framework.

Background

The paper proposes VQDNA, a framework that learns a vector-quantized (VQ) codebook as a genome vocabulary for tokenization and downstream sequence modeling. To enhance representational granularity, the authors introduce Hierarchical Residual Quantization (HRQ) to expand the effective vocabulary in a coarse-to-fine manner. The conjecture motivates HRQ by positing that the inherent four-letter nucleotide alphabet may constrain VQ codebook learning, potentially suppressing fine-grained genomic patterns in the quantized embeddings.

While the authors provide empirical evidence that HRQ improves performance and captures biologically meaningful patterns (e.g., SARS-CoV-2 lineage distinctions), they explicitly frame the limitation of the original vocabulary as a conjecture rather than a proven result, leaving its formal establishment open.