Learning to engineer protein flexibility
Authors: Petr Kouba, Joan Planas-Iglesias, Jiri Damborsky, Jiri Sedlar, Stanislav Mazurenko, Josef Sivic
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the Pearson correlation coefficient (PCC) between the predicted flexibility and the ground truth flexibility (RMSF) for each protein in the ATLAS test set. We report the average PCC across the test set. Our predictors Flexpert-Seq and Flexpert-3D outperform the baselines when using only the sequence on the input and when also using the structure, respectively. See Table 2 for the sequence case and Table 3 for the case when using sequence and structure. Table 4 shows that our method Flexpert-Design successfully managed to engineer increased flexibility in the engineered region. |
| Researcher Affiliation | Academia | 1Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University 2Loschmidt Laboratories, Department of Experimental Biology and RECETOX, Masaryk University 3International Clinical Research Centre, St. Anne s University Hospital Brno |
| Pseudocode | No | The paper includes diagrams and describes methodologies in detail, for instance, in Section 4.2 titled "FLEXPERT-DESIGN: FLEXIBILITY ENGINEERING VIA FLEXIBILITY-AWARE INVERSE FOLDING", which refers to Figure 2 for an overview. However, there are no explicitly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | The code, together with the instructions on how to download the data and trained weights for the reproduction of this work, can be found at https://github.com/Kouba Petr/Flexpert. |
| Open Datasets | Yes | For working with data on MD simulation, we select the recently published ATLAS dataset (Vander Meersche et al., 2023), particularly due to its excellent level of data curation. We further tested our predictors Flexpert-Seq and Flexpert-3D on a separate dataset called md CATH (Mirarchi et al., 2024), which features MD trajectories obtained at various simulation temperatures ranging from 320 K to 450 K. The evaluation is done using the CATH4.3 dataset. |
| Dataset Splits | Yes | We split the dataset using topology splitting, with the additional requirement that no topologies present in the CATH4.3 test set are present in our ATLAS training set. We evaluate the Pearson correlation coefficient (PCC) between the predicted flexibility and the ground truth flexibility (RMSF) for each protein in the ATLAS test set. The recently introduced md CATH dataset presents a challenging test of generalizability for our predictors trained on the ATLAS dataset. Both predictors were evaluated on the full dataset (5398 proteins, columns Full data ) as well as on its subset excluding topologies present in the ATLAS training set (4013 proteins, columns Topo. filtered ). |
| Hardware Specification | No | The paper mentions "We finetuned the model for 12 hours on 1 GPU" and acknowledges access to "the LUMI supercomputer". However, it does not provide specific models or types for the GPU or the supercomputer's architecture (e.g., NVIDIA A100, specific CPU details, etc.). |
| Software Dependencies | No | The paper mentions several software tools like "Protein MPNN", "Alpha Fold2", "ESMFold", "Prot Trans", "Lo RA", "Pro Dy", "GROMACS software", and "Colabfold", but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | We fine-tune them using the combined LF lexpert loss: LF lexpert = θ LF lex + (1 θ) LSeq, where θ [0, 1] is the parameter mixing the two losses LF lex and LSeq into a single loss. We observed a drop of only one percentage point in terms of sequence recovery, when we trained with θ = 0.8. We decided to first train Protein MPNN in a standard way for 100 epochs on topology split CATH4.3 dataset. Second, we finetuned the model for 12 hours on 1 GPU using the LF lexpert, which resulted in additional 8-11 epochs. The set of flexibility-increasing instructions F τ,S is constructed in the following way: (i) ... (iii) For each residue i in each selected segment S, its flexibility instruction fi is incremented by τ > 0. The presented median enrichment ratios Med(rij) and the proportion of flexibility-increasing mutations were obtained using the CATH4.3 test set with engineered segments of length |S| = 50, and flexibility engineering instructions corresponding to native flexibility incremented by τ = 5 in the engineered segment. |