---
dataset_info:
  features:
  - name: vh
    dtype: string
  - name: vl
    dtype: string
  - name: label
    dtype:
      class_label:
        names:
          '0': non-binder
          '1': binder
  splits:
  - name: train
    num_bytes: 10272366
    num_examples: 41787
  - name: eval
    num_bytes: 1141623
    num_examples: 4644
  - name: test
    num_bytes: 1268252
    num_examples: 5159
  download_size: 1194761
  dataset_size: 12682241
---
# SARS-CoV-2 binding dataset

Dataset of 104972 antibodies screened for binding the SARS-CoV-2 HR peptide, described in [Engelhart et al. (2022)](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9606274/), were obtained from [Zenodo](https://zenodo.org/record/5095284).

Average predicted logK<sub>D</sub> values were used for classifying sequences as binders and non-binders:
* logK<sub>D</sub><3 = binders
* logK<sub>D</sub>>=4 = non-binders
* logK<sub>D</sub>>=3 and logK<sub>D</sub> = ambiguous; removed.

Using these criteria, we have 51590 sequences remaining; these were stratified into an 80:10:10 ratio for training, test, validation, leading to:
* 41787 sequences in training
* 5159 sequences in validation
* 4644 sequences in test

Example

| vh | vl | label |
| ------- | ------- | ----- |
| EVQ...  | DIQ...  | 1     |
| EVQ...  | DIQ...  | 0     |
| EVQ...  | DIQ...  | 0     |
| EVQ...  | DIQ...  | 1     |


References
* [Engelhart et al. (2022) paper](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9606274/)
* [Zenodo link for dataset](https://zenodo.org/record/5095284)