A generative diffusion model (in the style of RFdiffusion / AlphaFold-adjacent design tools) starts from random Cα coordinates and iteratively removes noise, pulling the chain toward a physically valid fold. Here the target fold is a canonical right-handed α-helix, and a single "denoising time" t (1 = pure noise, 0 = fully generated) linearly blends noisy and target coordinates each step, with residual thermal jitter set by the sampling temperature.
Reverse diffusion: x(t) = t·x_noise + (1-t)·x_target
Helix geometry (Cα): x_i = r·cos(iθ), y_i = r·sin(iθ), z_i = i·h
r = 2.3 Å θ = 100°/residue h = 1.5 Å (Pauling-Corey-Branson α-helix)
Confidence: pLDDT ≈ 100·exp(-RMSD_bond / 2 Å)
Energy: E = Σ_bonds k(dᵢ-3.8Å)² + Σ_{|i-j|≥3, d<4Å} 8·(4-d)²
- Sequence length — number of residues folded (N), rebuilds the noise cloud and target helix.
- Sampling temperature — thermal jitter added on top of the denoised structure, degrading confidence and raising energy (like high-temperature sampling in generative models).
- Denoising speed — how fast t falls from 1 to 0 per second.
- Regenerate — resamples the noise seed and restarts t at 1.
- Confidence / Rainbow N→C — recolors residues by per-residue local geometry quality or by chain position, mirroring real structure viewers (AlphaFold pLDDT heatmaps vs. PyMOL rainbow).
Real pipelines (RFdiffusion, Chroma, ESM-based design) chain this kind of structure generation to an inverse-folding sequence model and a structure-prediction scorer, then validate top candidates with wet-lab binding or stability assays before synthesis.