GDSBM-AMP: Discrete Schrödinger Bridge for Antibacterial Peptide Design

GDSBM-AMP generates antibacterial peptides with a discrete Schrödinger bridge over the Hamming graph on the 20 standard amino acids, following the MadSBM formulation (Minimal-Action Discrete Schrödinger Bridge Matching for Peptide Sequence Design, 2026) with two modifications. First, an action penalty weighting the learned control against the frozen ESM-2 logits, active at lambda = 0.1 (losses.py, action_penalty). Second, a gradient parametrisation of the control, u(x,y,t) = v(y,t) - v(x,t), which constrains the control to derive from a scalar potential and makes it conservative on the Hamming graph: its circulation vanishes on every cycle. Biological plausibility is therefore carried by the pretrained protein language model rather than by a hand-specified substitution matrix. The submitted library is produced by unguided ancestral sampling; the conditional guidance and SMC machinery in condition.py is not on the generation path.