FliPS

Deep generative models for protein backbones have largely been conditioned on static targets: a motif to scaffold, a symmetry group, a binding partner. The structures they produce tend to be highly thermostable and correspondingly rigid. Many protein functions depend on the opposite property — enzymes open and close loops through a catalytic cycle, and molecule-specific binders need local mobility to engage a partner. FliPS, from the Gräter group at the Heidelberg Institute for Theoretical Studies with collaborators in Mainz, Stockholm, and Karlsruhe, makes flexibility itself the design objective.

The work introduces two models. BackFlip is an SE(3)-equivariant network that predicts a per-residue flexibility profile directly from a backbone, using no sequence and no protein language model. FliPS solves the inverse problem: an SE(3)-equivariant conditional flow matching model that samples backbones displaying a supplied target profile. Together they form a pipeline in which candidates are generated conditionally, ranked by BackFlip, and the top design is verified in all-atom molecular dynamics. The paper was presented at ICML 2025.

#Key Features

• Flexibility as a conditioning signal: A target per-residue profile is embedded as a node feature in eight bins up to 3 Å, letting a designer draw the mobility they want along the chain and sample backbones that display it.
• Sequence-free flexibility prediction: Because BackFlip reads only geometry, it is differentiable with respect to the backbone and can be used as an auxiliary training loss inside the generative model — something a language-model-based predictor cannot do.
• Local RMSF: The authors define flexibility as fluctuation measured against a residue's local 12-residue neighbourhood rather than a global superposition, removing artifacts that global RMSF introduces for multi-domain proteins.
• Screening and training-free guidance: BackFlip screening ranks any pool of backbones against a target profile, and BackFlip guidance adds a gradient term that steers an existing unconditional model at inference, though it underperforms the conditional model.

#Technical Details

BackFlip adapts the Clifford Frame Attention blocks of Geometric Algebra Flow Matching (GAFL), using four attention blocks with node and edge embeddings of 64 and 32 dimensions for 0.68 M parameters. It trains on 1,294 proteins from the ATLAS molecular dynamics database (1,035 train / 130 validation / 129 test) and reaches a Pearson correlation of 0.80 and mean absolute error of 0.17 Å against MD-derived profiles, close to the 0.84 upper bound set by disagreement between independent MD replicas, and far above pLDDT (0.54) or crystallographic B-factors (0.16). Inference takes 0.6 s for a 300-residue protein against 118 s for an AlphaFold 2 pLDDT proxy.

FliPS extends the 16.7 M-parameter GAFL model with the flexibility embedding, a differentiable BackFlip auxiliary loss, and profile masking during training. It trains on 22,977 PDB structures of 60–512 residues annotated by BackFlip, for 21 GPU-days on eight A100s. Across ten natural and ten hand-drawn target profiles, MD simulations of the top-ranked designs recover the targets at Pearson r of 0.70 and 0.78, against 0.58 and 0.56 for screened RFdiffusion samples and 0.45 and 0.50 for FoldFlow2. Novelty is comparable to those baselines and the designs carry more β-strand content, but self-consistency RMSD rises to 1.5–2.0 Å because flexible backbones are intrinsically harder to refold consistently.

#Applications

The pipeline suits design campaigns where motion is part of the specification: enzyme active sites that require loop opening and closure, binders for DNA, RNA, or small molecules that need local induced fit, and switchable or allosteric scaffolds. Because flexibility conditioning is an additional input rather than a replacement for existing constraints, it composes naturally with motif and symmetry conditioning. BackFlip is independently useful as a fast filter, ranking large pools of candidate backbones before the expensive inverse-folding and refolding steps.

#Impact

FliPS is the first generative model to condition protein backbone generation on a continuous dynamic property, extending controllable design beyond the static targets that dominate the field. Code, pretrained weights, and the flexibility-annotated PDB and SCOPe datasets are released under the MIT license, with a Colab notebook for inference. Two limitations are stated plainly by the authors: flexibility is defined by 300 ns simulations, so slower functional motions are outside the model's ground truth, and conditioning on high-mobility profiles trades away some designability as conventionally measured.