Hierarchical Phase Transitions in Bayesian Flow Networks Enable Training-Free Molecular Design

Abstract

Bayesian flow networks (BFNs) jointly model atomic coordinates and atom types for structure-based drug design, but their generative dynamics remain poorly understood. We show that the two channels resolve on different terms: atom types pass through a sharp commitment transition at t_c1, whereas coordinate precision improves smoothly with no critical point of its own, reaching sub-Ångstrom resolution only at a stated threshold t_c2. Extreme-value theory predicts the type channel’s posterior half-maximum in closed form, and the predicted $1/\sqrt{{\beta }_1}$ scaling holds across seven β₁ values (R² = 0.996) and over 13 points spanning three BFN architectures and two domains, though measured times run 11%–20% below the predicted absolute values. Exploiting that structure, phase-aware iterative refinement aligns three rounds to these boundaries and trains no auxiliary regressor, reward function, or property predictor. On 100 CrossDocked2020 test pockets at K = 200, it reaches a mean Vina Dock of −8.72 kcal/mol. Gradient-guided methods that need such training stay ahead by a small margin: over the 100-pocket intersection, the paired gap to MolJO (−8.98) is 0.25 kcal/mol (95% CI [−0.13, +0.60], paired Wilcoxon p = 0.022); equivalence testing at a preregistered ±0.5 kcal/mol margin does not establish equivalence within it. CByG (−9.16) leads by 0.44 kcal/mol on unpaired means. Phase-aware iteration improves the Round-1 to Round-3 mean by 1.34 kcal/mol (paired Wilcoxon p = 4.80 × 10^–18, Cohen’s d_z = −1.59) and shifts the median by −1.44 kcal/mol, a shift Best-of-K resampling cannot produce. Paired on the pocket, an audited rerun passes every PoseBusters check 4.7 percentage points more often than MolJO.