NomadFlow lets protein motif groups move on their own: 19 of 21 multi-group tasks yield unique designs

Multi-fragment protein design no longer needs the position of each fragment along the chain or their mutual placement to be specified in advance: the model receives only each fragment's own geometry and decides how they combine into one chain. Earlier motif scaffolding inserted the motif as a fixed structure, with its three-dimensional placement and sequence position both given and the mutual pose of several motifs commonly fixed as well; conditional generation removes the sequence position but usually treats the motif as a single rigid object. Across 21 multi-group tasks with 200 designs sampled per task and checked by ProteinMPNN and ESMFold, 19 tasks yield at least one unique successful design; the misses are 4JHW, where the mobile regime produced no successful design, and 7K4V_ext2x10, where both regimes give zero.

Each motif group moves as a rigid body during generation, so the relative geometry of atoms inside a group cannot drift and no intra-group geometric penalty is needed. Per-residue updates applied independently would pull a group's internal geometry apart; earlier routes either inserted the motif wholesale or relied on a penalty. The denoiser predicts an SE(3) increment for each motif frame separately; these increments are averaged into one common rigid transformation and applied to all frames of the group at once, with the same projection used in training and sampling.

A group's pose is carried by an auxiliary frame at its center of mass, the flow-matching loss is placed on it, and gradients return through the averaging step to the predictions of the motif frames. A group pose has six degrees of freedom, and learning it as the independent motion of every motif residue would learn something other than how the group actually moves. The auxiliary frame is not a motif residue but expresses the group pose in the frame interface the network already uses; at sampling, per-residue noise on motif frames is switched off and replaced by one common SE(3) perturbation per group.

Choosing the mutual arrangement of the groups and completing the scaffold are separated in time: the groups are mobile in the first half, then fixed while only the scaffold is denoised. Placing the groups and completing the scaffold along one global time axis mixes two different subproblems. All reported runs use tlock = 0.5, the same value as in finetuning, so half the trajectory goes to placing the motif groups and half to completing the scaffold.

The demonstrated scope is 21 multi-group tasks built from a single PDB chain, with 200 designs sampled per task and success judged by whether ESMFold folding after ProteinMPNN redesign reproduces each group's geometry; these tasks guarantee at least one compatible layout, do not test assembly of motifs drawn from different proteins, and computational success does not imply expression or function.

The authors describe the benchmark as a step toward analyzing partition granularity and structural context separately, with group counts and added-segment lengths chosen from intuition rather than calibrated in advance; the shape of successful layouts varies strongly with the condition geometry, and the low-dimensional family seen on 5AOU_ext20 and 5AOU_QUAD_ext20 reflects this model's successes under the PMPNN+ESMFold filter rather than proof that the task is one-dimensional; because MoDiff and FADiff did not release code and trained weights together, no cross-model numerical comparison was run, and ablations of individual choices such as tlock remain future work.