International Conference on Artificial Intelligence, Internet of Things and Cloud Computing Technology (AIoTC)
Quantum Boltzmann Machine Enhanced Variational Autoencoder for Cds Sequence Modeling
Junya Wang, Qi Gao, Yin Ma, Pengcheng Zhu, Shiguang Feng
Abstract
Coding sequences (CDSs) are high-dimensional discrete biological sequences with recognizable sequence-level statistics such as GC content and codon usage. In this paper, CDS data are used as a structured sequence-modeling benchmark to study whether a quantum Boltzmann-machine-enhanced VAE (QVAE) can learn latent dependencies in discrete biological sequences. The QBM prior uses a diagonal Ising-type Hamiltonian to model dependencies among binary latent variables, and converts the negative-phase update into an Ising-type sampling task suitable for quantum or coherent-Ising samplers. Compared with Gaussian VAE and Bernoulli DVAE baselines, QVAE+Mamba achieves high decoded-sequence diversity with a mean normalized pairwise edit distance of 0.691, while preserving selected dataset statistics including GC-content and codon-frequency trends. The reported QVAE+Mamba result is obtained using a physical coherent Ising machine, while classical simulation gives similar behavior in our additional checks. These results suggest that a quantum structured prior can support high-dimensional discrete sequence modeling. The generated sequences should not be interpreted as protein-preserving or therapeutically optimized mRNA designs without additional biological constraints and functional validation.