AmelieSchreiber's picture
Update README.md
d940ece
metadata
license: mit
language:
  - en
tags:
  - biology
  - protein sequences
  - binding sites
  - active sites
size_categories:
  - 100K<n<1M

This dataset is obtained from a UniProt search for protein sequences with family and binding site annotations. The dataset includes unreviewed (TrEMBL) protein sequences as well as reviewed sequences. We refined the dataset by only including sequences with an annotation score of 4. We sorted and split by family, where random families were selected for the test dataset until approximately 20% of the protein sequences were separated out for test data. We excluded any sequences with <, >, or ? in the binding site annotations. We furthermore included any active sites that were not listed as binding sites in the labels (seen in the merged "Binding-Active Sites" column). We split any sequence longer than 1000 residues into non-overlapping sections of 1000 amino acids or less after the train test split. This results in subsequences of the original protein sequence that may be too short for consideration, and filtration of the dataset to exclude such subsequences or segment the longer sequences in a more intelligent way may improve performance. Pickle files containing only the train/test sequences and their binary labels are also available and can be downloaded for training or validation of the train/test metrics.