An end-to-end differentiable deep learning framework for translating binary digital payloads into biochemical-compliant DNA sequences ({A, C, G, T}).
- Differentiable Channel: Gumbel-Softmax continuous relaxation with simulated substitution noise.
- Biochemical Loss Constraints: Differentiable penalties for GC-content window adherence (40%–60%) and homopolymer run suppression (max run < 3).
- Architecture: Joint Source-Channel Coding (JSCC) autoencoder mapping 128-bit blocks to 80-nucleotide sequences.
- Effective Code Rate: 1.62 bits/nucleotide
- Bit Error Rate (BER): < 0.005
- Biological Policy Violation Rate: Reduced from 34.2% to 2.1%
data_generator.py: Generates uniform binary and structured ASCII payloads.model.py: Neural autoencoder encoder-decoder architecture.noise_channel.py: Continuous Gumbel-Softmax biological channel.loss.py: Multi-objective loss with GC and homopolymer penalty terms.train.py: Training pipeline.evaluate.py: Evaluation benchmarks across BER, policy violations, and code rate.