Physics-Informed Modeling

E_PHYS

Differentiable Cα energy terms enforcing physical protein geometry during diffusion-based structure generation.

Total Physics Loss

L_phys = w_bond·E_bond + w_angle·E_angle + w_clash·E_clash

λ(t)·L_phys added to score-matching loss with linear warmup

Physics loss warmed up gradually from 0 over warmup_epochs

Bond Length Energy

E_bond = Σ_bonds k_b (d − d₀)²

Harmonic penalty on Cα–Cα virtual bonds. d₀ = 3.8 Å (ideal Cα distance). Differentiable via PyTorch autograd.

Bond Angle Energy

E_angle = Σ_angles k_a (θ − θ₀)²

Harmonic penalty on consecutive Cα–Cα–Cα angles. θ₀ = 2.1 rad. High initial values indicate unconstrained diffusion in early epochs.

Steric Clash Energy

E_clash = Σ_{i<j} max(0, d_min − d_ij)²

Repulsive term for atoms closer than d_min = 3.5 Å. Prevents non-physical atom overlap during structure generation.

Training Log — Epoch 1 Physics Values

loss_phys = 1147.84
  e_bond  = 144.4
  e_angle = 2000.0 ← high early (unconstrained diffusion)
  e_clash = 3.44
# Physics terms decrease as model learns physical geometry

Physics Config

d₀ (ideal Cα bond)

3.8 Å

θ₀ (ideal Cα angle)

2.1 rad

d_min (clash threshold)

3.5 Å

w_bond

0.1

w_angle

0.1

w_clash

0.5