BackFlip & FliPS: Neural Models for Predicting and Generating Protein Flexibility

Background

Recent advances in geometric deep learning and generative modeling have enabled the design of proteins with multiple static target properties such as structural motifs and symmetry. However, most existing methods generate only static structures and ignore flexibility, which is crucial for catalytic activity and molecular recognition.

BackFlip: Predicting Per‑Residue Flexibility

We introduce BackFlip , an SE(3)-equivariant neural network that takes a protein backbone (C‑α trace) as input and predicts a per‑residue flexibility score. The model leverages equivariant convolutions to respect the rotational and translational symmetries of the protein structure.

FliPS: Conditional Flow Matching for Flexible Backbone Generation

Building on BackFlip, we propose FliPS , a conditional flow‑matching model that treats flexibility as a conditioning variable. FliPS solves the inverse problem: given a desired flexibility distribution, it generates novel backbone conformations whose predicted flexibility matches the target.

Experimental Validation

We trained the models on a curated dataset of protein structures with experimentally derived flexibility annotations. In downstream experiments, FliPS produced diverse backbone designs that satisfied the prescribed flexibility profiles. Molecular dynamics (MD) simulations confirmed that the generated structures maintained the intended flexibility over time.

Resources

The implementations of BackFlip and FliPS are publicly available at https://github.com/graeter-group/flips.