Structural Prediction
SCORE-BASED DIFFUSIONCα coordinate prediction via score-based diffusion (T=1000) conditioned on Pairformer representations. Interactive 3D viewer loads experimental PDB structures from RCSB PDB.
Score Matching Objective
L_score = 𝔼[‖s_θ(x_t, t, c) + ε/σ(t)‖²]
x_t = √ᾱ_t · x_0 + √(1−ᾱ_t) · ε · β linear: β_0=0.0001, β_T=0.02, T=1000
🌊
T = 1,000
Diffusion Steps
linear β schedule
🧮
3-layer MLP
Score Network
→512→256→3 (Cα xyz)
🔬
841
Proteins w/ PDB
supervised structural loss
⚗
0.01
Physics Loss λ
bond+angle+clash
3D Structure Viewer
PDB ID (4 characters)
HIV-1 / MTB Examples
Representation
PaTz Structural Output: The score diffusion head predicts Cα coordinates for all 2,620 proteins. Experimental structures shown here are from RCSB PDB — predicted structures (when exported) will be overlaid alongside. Go to Molecular Visualization for IG-colored residue views.
Diffusion Architecture
T (timesteps)1000
Noise scheduleLinear (β₀=0.0001 → β_T=0.02)
Forward processx_t = √ᾱ_t·x_0 + √(1−ᾱ_t)·ε
Score inputx_t ∈ ℝ^(L×3), t ∈ [1,T], c=Pairformer emb
Score network(d_single + d_time_embed + 3) → 512 → 256 → 3
d_time_embed64 (sinusoidal positional)
n_score_layers3
OutputCα displacement ∈ ℝ^(L×3)
Proteins w/ PDB (train)841 (supervised Cα loss)
Physics Energy Terms (E_phys)
E_phys = w_bond·E_bond + w_angle·E_angle + w_clash·E_clash
E_bond = Σᵢ (|rᵢ₊₁−rᵢ|−3.8Å)²
E_angle = Σᵢ (cos θᵢ − cos θ₀)²
E_clash = Σᵢ₍ᵢ₊₃₎ max(0, 3.0−|rᵢ−rⱼ|)²
E_angle = Σᵢ (cos θᵢ − cos θ₀)²
E_clash = Σᵢ₍ᵢ₊₃₎ max(0, 3.0−|rᵢ−rⱼ|)²
λ_phys0.01 (warm up over 5 epochs)
clash_weight1.0
distance_weight0.5
angle_weight0.3
torsion_weight0.2
Target Cα bond3.8 Å