Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Flexibility-conditioned protein structure design with flow matching
Authors: Vsevolod Viliuga, Leif Seute, Nicolas Wolf, Simon Wagner, Arne Elofsson, Jan Stühmer, Frauke Gräter
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experiments, we show that Fli PS is able to generate novel and diverse protein backbones with the desired flexibility, verified by Molecular Dynamics (MD) simulations. ... 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. ... In Tab. 1 we report the correlation and mean absolute error (MAE) of Back Flip’s predicted local RMSF compared with the value obtained from two of the three MD trajectories per protein in ATLAS. |
| Researcher Affiliation | Academia | 1Heidelberg Institute for Theoretical Studies, Heidelberg, Germany 2Dept. of Biochemistry and Biophysics at Stockholm University and Science for Life Laboratory, Stockholm, Sweden 3Max Planck Institute for Polymer Research, Mainz, Germany 4IWR, Heidelberg University, Heidelberg, Germany 5IAR, Karlsruhe Institute of Technology, Karlsruhe, Germany. |
| Pseudocode | No | The paper describes methods using mathematical formulations and descriptive text, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps in a code-like format. |
| Open Source Code | Yes | Fli PS and Back Flip are available at https: //github.com/graeter-group/flips. |
| Open Datasets | Yes | We train Back Flip with a mean squared error loss on per-residue flexibilities of backbones from the ATLAS dataset (Vander Meersche et al., 2024). ... We train Fli PS on the PDB dataset (Berman et al., 2000) introduced in Frame Diff (Yim et al., 2023b) annotated with Back Flip-predicted residue-level flexibilities. ... We also apply the same screening procedure to 3673 natural proteins from the SCOPe dataset (Fox et al., 2014; Chandonia et al., 2022). |
| Dataset Splits | Yes | We filter ATLAS for proteins up to the length of 512 residues, resulting in a total of 1294 proteins, which we split in 1035 backbones for training, 130 for validation and 129 for testing (further details in Appendix A.2). ... We split the dataset into train, valid, and test set according to the splitting scheme 80:10:10 and limit protein size to 512 residues. |
| Hardware Specification | Yes | We train the model on a single NVIDIA A100 GPU and use the value of RMSE between the ground truth and predicted local flexibility scores on the validation set as the early stopping criterion. ... We train Fli PS for a total of 21 GPU days on eight NVIDIA A100 GPUs (details in Appendix A.9.) |
| Software Dependencies | Yes | We conduct MD simulation using GROMACS v2023 (Abraham et al., 2015) with the all-atom force field CHARMM27. |
| Experiment Setup | Yes | The hyperparameters are chosen as in GAFL (Wagner et al., 2024) and λflex is set to 100. ... We train the model for a total of 21 GPU days on eight NVIDIA A100 GPUs (details in Appendix A.9.). ... We batch proteins together during training, where each batch contains at most 32 proteins. ... During training, 10% of the time we fully mask one-hot encoded per-residue flexibility and perform training of Fli PS as an unconditional model... For remaining 90%, we randomly unmask flexibility as a window of size U([0.2, 0.4]) and with the window center positions U([1, N]) where N stands for the total protein length. |