Stable De Novo Protein Design via Joint Conformational Landscape and Sequence Optimization

K50dG_dmsv2_full.csv includes the complete generated and experimentally tested dataset, which contains a total of 39,210 designs.

K50dG_dmsv2_all_4.csv includes a filtered dataset of 20,668 designs, containing one sequence for each of the four different models. In this dataset, a single ID corresponds to sequences generated by all four methods.

K50dG_dmsv2_low_ipae.csv includes a further filtered dataset of 13,442 designs, used in Figures 2 and 3. This dataset contains proteins with low AF2 IPAE values, selected to avoid potential issues with generating homo-oligomers and to ensure that the proteins exist as monomers.

For Figure 2, which presents a one-to-one comparison of the four methods, we selected IDs that include sequences from all four methods within the 13,442 low-IPAE set. This results in a subset of 5,708 sequences (1,427 per method).

For Figure 3, we used the full 13,442-sequence low-IPAE dataset, since this figure does not involve one-to-one comparisons between methods.

This dataset also includes the pre-trained design-model weights used by the Joint TrMRF / TrROS pipeline (https://github.com/yehlincho/Joint_Model_Stability). Download the two bundles below and unzip them into `design_models/`.

Model weight bundles

TrMRF_weights.zip (~19 MB)
- model_TrMRF_A.npy - 5 residual blocks, 100-channel input (TrMRF.ipynb)
- model_TrMRF_seqid_retrain_3blocks.npy - 3-block seqid retrain, 101-channel input (TrMRF_v2)
- model_TrMRF_seqid_retrain_5blocks.npy - 5-block seqid retrain (models.py TrMRF())

TrROS_weights.zip (~162 MB) - trRosetta ensemble + backgrounds (standard public trDesign weights)
- models/model_xaa.npy … models/model_xae.npy - 5-model trRosetta ensemble
- bkgr_models/bkgr01.npy … bkgr_models/bkgr05.npy - 5 background models