Diffusion for Molecular Design

Module: Generative AI | Difficulty: Advanced

Equivariant Diffusion

where includes both atomic positions and types.

EDM (Equivariant Diffusion Model)

SE(3) Equivariance

class EquivariantDiffusion(nn.Module):
    def __init__(self, net):
        super().__init__()
        self.net = net
    def forward(self, x_t, t, edge_index):
        h = self.net(x_t, t, edge_index)
        return h  # predicted noise
    def loss(self, x0, edge_index):
        t = torch.randint(0, 1000, (x0.size(0),))
        noise = torch.randn_like(x0)
        xt = self.q_sample(x0, t, noise)
        return ((noise - self(xt, t, edge_index))**2).mean()

Research Insight: Equivariant diffusion models achieve state-of-the-art on molecule generation benchmarks because they respect the physical symmetries of 3D molecular structures.