NMA-tune: Generating Highly Designable and Dynamics Aware Protein Backbones

Authors: Urszula Julia Komorowska, Francisco Vargas, Alessandro Rondina, Pietro Lio, Mateja Jamnik

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the effectiveness of the conditioner and its impact on the sample designability using 3 proteins chosen from the literature. We found that for the dynamics conditioned samples, there exist protein sequences that will fold to protein backbones with desired normal modes, and NMA-tune outperforms existing state-of-the-art. We run MD simulations on selected samples and perform Principal Component Analysis (PCA) on their trajectories.
Researcher Affiliation Academia 1Department of Computer Science and Technology, University of Cambridge, Cambridge, United Kingdom 2Department of Molecular and Translational Medicine, University of Brescia, Brescia, Italy.
Pseudocode No The paper describes its methodology in detailed text and includes a 'Conditioning framework diagram' in Figure 6 (Appendix B), but it does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes We provide the conditioner as a ready to use plug-in to the open-source model RFdiffusion. Since RFdiffusion is already able to do motif scaffolding, this constitutes an extension to joint conditioning that can easily be used by the wider research community.
Open Datasets Yes Dataset for training ϵθ was based on the SCOPe database (Fox et al., 2013; Chandonia et al., 2021). ... We chose: triglyceride lipase (Derewenda et al., 1992) (PDB id: 4tgl), calmodulin (Khade et al., 2021) (PDB id: 1exr), and HIV-1 protease in semi-open conformation (Hornak et al., 2006) (PDB id: 1hhp).
Dataset Splits Yes 7139 proteins remained and we used train:validation:test split 0.8:0.1:0.1.
Hardware Specification Yes ϵθ is trained for 10 epochs with Adam optimiser, which took 15h on a single Nvidia A100 80GB. ... Molecular dynamics (MD) simulations were performed using GROMACS 2024 version and NVIDIA A100 80GB GPU.
Software Dependencies Yes Molecular dynamics (MD) simulations were performed using GROMACS 2024 version and NVIDIA A100 80GB GPU.
Experiment Setup Yes ϵθ is trained for 10 epochs with Adam optimiser, which took 15h on a single Nvidia A100 80GB. The learning rate 1e-4 is decreased by 0.1 after 5000 gradient updates and batch size is 32. ... We used 50 sampling steps, which is the default in RFdiffusion. ... Additionally, for targets 1hhp and 1exr we ablate the time scaling function in Table 4. ... All losses are weighted as L = 0.05 Lnoise +0.8 LNMA +0.1 Lchain +0.05 Lrg (11) where LNMA is the same as ly(x) in Equation 8.