Flexibility-conditioned protein structure design with flow matching
Vsevolod Viliuga, Leif Seute, Nicolas Wolf, Simon Wagner, Arne Elofsson, Jan Stühmer, Frauke Gräter · 2025
Read Flexibility-conditioned protein structure design with flow matching by Vsevolod Viliuga, Leif Seute, Nicolas Wolf, Simon Wagner, Arne Elofsson, Jan Stühmer, Frauke Gräter (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-18211.txt.
Introduction
Vsevolod Viliuga and co-authors (2025) teach a protein design machine to aim not just at a shape, but at how much the shape should wiggle. Proteins work by moving: enzymes flex to catalyze, receptors flex to recognize. Yet most design tools target static structures. The authors first build BackFlip, a neural network that reads a protein backbone and predicts, residue by residue, how flexible each part is. On top of it they build FliPS, a flow-matching generator that works backward: hand it a desired flexibility profile and it produces backbones that display it.
Both networks respect the geometry of three-dimensional rotations, so their predictions do not depend on how a structure is oriented. In experiments the generator produces novel, varied proteins whose measured flexibility tracks the requested profile. Conditioning design on motion, not just form, opens a route to proteins engineered for the movements their jobs demand.
Catalog record
| CREATOR | Vsevolod Viliuga, Leif Seute, Nicolas Wolf, Simon Wagner, Arne Elofsson, Jan Stühmer, Frauke Gräter |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-18211
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Flexibility-conditioned protein structure design with flow matching Vsevolod Viliuga, Leif Seute, Nicolas Wolf, Simon Wagner, Arne Elofsson, Jan Stühmer, Frauke Gräter Abstract Recent advances in geometric deep learning and generative modeling have enabled the design of novel proteins with a wide range of desired properties. However, current state-of-the-art approaches are typically restricted to generating proteins with only static target properties, such as motifs and symmetries. In this work, we take a step towards overcoming this limitation by proposing a framework to condition structure generation on flexibility, which is crucial for key functionalities such as catalysis or molecular recognition. We first introduce BackFlip, an equivariant neural network for predicting per-residue flexibility from an input backbone structure. Relying on BackFlip, we propose FliPS, an SE(3)-equivariant conditional flow matching model that solves the inverse problem, that is, generating backbones that display a target flexibility profile. In our experiments, we show that FliPS is able to generate novel and diverse protein backbones with the desired flexibility, verified by Molecular Dynamics (MD) simulations. FliPS and BackFlip are available at https://github.com/graeter-group/flips . † † affiliationnotice: Equal contribution Figure 1: Flexibility-conditioned protein backbone design. Given a target flexibility profile (1), the conditional generative model FliPS is used to sample several candidate backbones (2). Flexibility profiles of backbones are predicted with BackFlip (3) and the most promising candidate is selected (4). The flexibility of the protein can then be validated in a Molecular Dynamics simulation (5). 1 Introduction The past few years have arguably marked the golden age of de novo protein design. Rapid advancement of deep learning-based approaches and their application to the protein structure prediction problem have enabled an unprecedented level of modeling accuracy ( Jumper et al. 2021 ; Dauparas et al. 2021 ; Lin et al. 2022 ) . This, in turn, paved the way for more controllable and precise design of protein structures with specific functional properties. For example, proteins with on-demand chemical reactivities or binding propensities are of particular interest for biotechnological applications ( Lovelock et al. 2022 ) , therapeutics development ( Bennett et al. 2024 ) and even sustainability problems, such as plastic degradation ( Chen et al. 2022 ) . Only recently, several deep generative models for protein structure design based on diffusion ( Song et al. 2021 ) or flow matching ( Lipman et al. 2023 ; Albergo & Vanden-Eijnden 2022 ) have successfully demonstrated high biophysical consistency of generated samples ( Watson et al. 2023 ; Bose et al. 2024 ; Yim et al. 2023a ) . The molecular context that such state-of-the-art models can be conditioned on is typically restricted to static properties like the presence of a certain motif in the final structure, adherence to specific symmetry groups or binding to target proteins or ligands ( Ingraham et al. 2023 ; Krishna et al. 2024 ; Yim et al. 2024 ; Lin et al. 2024 ; Huguet et al. 2024 ) . Protein structures generated by deep learning methods tend to be highly thermostable, ( Watson et al. 2023 ; Zambaldi et al. 2024 ) , which suggests high degree of structural rigidity ( Scandurra et al. 1998 ) , and incorporating flexibility into specific parts of a backbone remains a great challenge. Proteins, however, must possess dynamic behavior in order to encode function such as chemical reactivities or specific recognition of binding partners. For example, enzymes have complicated free-energy landscapes ( Benkovic et al. 2008 ) and traverse from one conformation to another in their catalytic cycle, which includes loop opening and closure, concurrent domain movements and local folding and unfolding events ( Miller & Benkovic 1998 ; Liao et al. 2018 ; Zinovjev et al. 2024 ) . Also molecule-specific protein binders require local structural flexibility in order to engage in a complex with DNA, RNA or ligand molecules ( Glasscock et al. 2023 ; Pacesa et al. 2024 ; Guo et al. 2024 ) . Since protein structure defines protein function, the limited structural diversity and high rigidity of proteins designed with state-of-the-art generative models impede the exploration of a vast potential functional space. Main contributions In this work, we propose a framework for generating protein structures conditioned on structural flexibility relying on two key innovations. We first introduce (1) BackFlip – an equivariant neural network trained to predict per-residue flexibilities of a given protein backbone entirely independent of sequence information. Utilizing BackFlip for large-scale flexibility annotation and an auxiliary loss, we propose (2) FliPS – a generative model for protein structure conditioned on per-residue flexibilities, defined as magnitude of local fluctuations in Molecular Dynamics (MD) simulation. FliPS is an extension of the recent Geometric Algebra Flow Matching (GAFL), an SE(3)-equivariant flow matching model for protein structures with diverse secondary structure composition of the generated backbones ( Wagner et al. 2024 ) . We combine the two proposed models and introduce a flexibility conditioning pipeline, in which backbones are generated with the conditional model FliPS and ranked with BackFlip (Fig. 1 ). In our experiments we demonstrate that the proposed flexibility-conditioning pipeline generates novel and diverse protein structures that display a given flexibility profile, verified by MD simulations on the timescale of 300 ns. 1.1 Related work Generative models for protein structure have been conditioned on static structural features, such as symmetry groups or motifs ( Watson et al. 2023 ; Yim et al. 2024 ) and on protein or ligand targets ( Watson et al. 2023 ; Ingraham et al. 2023 ; Krishna et al. 2024 ) . Deep learning-based tools have been successfully combined with classical modeling tools in the past to design dynamic protein structures, such as allosterically switchable assemblies ( Pillai et al. 2024 ) or pH-responsive filaments ( Shen et al. 2024 ) , without explicitly harnessing intrinsic flexibility. Only few works attempted incorporating dynamics into generative models for protein design. ( Komorowska et al. 2024 ) propose conditioning a pre-trained diffusion model on normal modes that approximate local harmonic movements around an equilibrium state. Instead of learning a conditional model, gradients of an analytical loss on normal modes are used for guidance upon inference. ( Kouba et al. 2024 ) introduce the backbone flexibility prediction model FlexPert-3D that relies on embeddings from a large protein language model (pLM), which we evaluate in Tab. A.1 . They propose a framework to condition sequence generation models to better align with a target flexibility, leaving the structure invariant. We discuss relevant works in more details in Appendix A.1 . 2 Background 2.1 Flow Matching for protein structure generation Riemannian Flow Matching The goal of flow matching ( Lipman et al. 2023 ; Tong et al. 2024 ) is to sample from an unknown distribution p_{1}:\mathcal{M}\rightarrow[0,1] on the data domain \mathcal{M} by learning a flow \phi_{t}:[0,1]\times\mathcal{M}\rightarrow\mathcal{M} that transforms a simple prior distribution p_{0} to the target p_{1} via the pushforward [\phi_{1}]_{*}p_{0}=p_{1} . The flow is parametrized by a learnable, time dependent vector field v(x,t):\mathcal{M}\times[0,1]\rightarrow\mathcal{T}_{x}\mathcal{M} on the tangent space \mathcal{T}_{x}\mathcal{M} according to the flow ODE, \frac{\mathrm{d}}{\mathrm{d}t}\phi_{t}(x)=v(\phi_{t}(x),t),\quad\phi_{0}(x)=x\,, (1) which is numerically integrated during inference. Lipman et al. 2023 have shown that the vector field v can be learned by defining trajectories between samples x_{0}\sim p_{0} and x_{1}\sim p_{1} and regressing on the