Per-Position Posterior Tempering Overtakes Trajectory Selection for Developability-Guided Antibody Lead Optimization

Generative models for antibody sequences can be steered toward favorable developability by injecting a sequence-level property scorer into the sampler that constructs each sequence residue by residue over many denoising steps. We study this guidance problem for a hybrid, germline-conditioned discrete flow matching model of paired heavy- and light-chain sequences. After benchmarking sequence-only hydrophobic-interaction chromatography (HIC) predictors on recent competition data and fixing the strongest frozen probe (Spearman ρ = 0.508) as the guidance oracle, we compare two ways of injecting it during generation: Soft Value-based Decoding (SVDD), which samples many whole-sequence rollouts and keeps the best, and per-position posterior tempering, which reweights each position's token distribution by the score of the model's predicted completion. On 20 held-out clinical leads, SVDD is more call-efficient at low budgets, but its predicted HIC saturates at roughly 1.96 as compute grows, whereas per-position tempering keeps improving with coverage and reaches 1.83, crossing over SVDD at around 400 scorer calls per lead. Per-position tempering also preserves full population diversity (8/8 unique designs per lead versus 1/1 for SVDD's greedy selection), and both methods' gains are corroborated by an independent developability scorer (SSH2.0) that was never used for guidance. Our results characterize a compute regime where per-position guidance is preferable to trajectory selection for antibody lead optimization.