fastDNA

Getting an equilibrated 3D conformation for a DNA or RNA sequence normally means running a simulation — even with oxDNA's speed, that's still seconds to minutes per structure, which adds up fast when you want to screen thousands of candidate designs or generate training data at scale.

fastDNA is a flow-matching generative model built around an SE(3)-equivariant transformer with Clifford geometric algebra attention, trained to map directly from sequence to an equilibrium 3D conformation — no simulation trajectory required at inference time. Model sizes range from 2M to 70M parameters, trained on HPC infrastructure with PyTorch Lightning and tracked in Weights & Biases.

The equivariance built into the architecture matters here: a DNA helix's physical properties shouldn't depend on how you happened to orient it in space, and baking that symmetry into the model rather than hoping the network learns it from data makes training far more sample-efficient.