A DNA-Local, Constraint-Aware Dual-Head Transformer for Pseudorandom Stream Generation

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Pseudorandom number generators (PRNGs) in DNA computing often map bits to DNA symbols indirectly.
• This indirect method may overlook DNA-specific constraints like GC balance and homopolymer limits.

Purpose of the Study:

• To propose a novel constraint-aware Transformer framework for direct DNA-local PRNG generation.
• To generate DNA sequences and bitstreams simultaneously within the A/C/G/T alphabet.

Main Methods:

• Developed a dual-head decoder-only Transformer framework for DNA-local PRNG.
• Model generates DNA bases and derives bitstreams via dynamic DNA-to-bit coding rules.
• Evaluated using real genomic data, synthetic data, and independent validation.

Main Results:

• Satisfied NIST SP 800-22 acceptance criteria and achieved high min-entropy (0.9955-0.9964 bit/bit).
• Preserved core DNA-compatibility constraints including GC balance and homopolymer control.
• Leakage analyses showed no systematic bit-level dependence or long-fragment copying.

Conclusions:

• The framework enables reproducible, constraint-aware DNA-local PRNG generation.
• Supports multilayer validation for both bitstreams and DNA sequences.
• Offers an improved approach for DNA-oriented computational workflows.