|
Description
|
This repository includes the trained transformer-based model for the small subset Cryo2StructData dataset, as well as the training and validation split files. These split files categorize density map EMD-IDs into low, medium, and high resolutions. The training and validation sets contain 1680 and 187 density maps, respectively, with a split ratio of 90:10. (2023-10-24)
|
|
Notes
| The trained model checkpoints are for predicting amino acid types, secondary structure types and backbone atom types. The backbone_atom_prediction.ckpt is used during the inference phase to classify each voxel into one of four different classes representing three backbone atoms (Cα,C and N) and no presence of any backbone atoms. Similarly, the amino_acid_type_prediction.ckpt is used to classify each voxel into one of twenty-one different amino acid classes representing twenty different amino acids and unknown or absence of amino acid. Finally, the secondary_structure_prediction.ckpt is used to classify each voxel into one of four different classes representing three secondary structure atoms (coils, α-helices, and β-strands) and no presence of any secondary structure atoms.
The inference code that utilizes these checkpoints is available in the Cryo2StructData GitHub repository : github.com/BioinfoMachineLearning/cryo2struct |
