binding-prediction
il6-binding-prediction
Human IL-6 binding dataset
Nanobodies binding IL-6 were obtained from the Github repo for Tsuruta et al. (2023). Labels for antibody sequences were provided from the Github repo as-is.
Briefly, we first removed any nanobody sequence having lower than 75%; human germline sequence identity was determined using ANARCI.
Among the remaining 232084 sequences, we only use antibodies that have confirmed binding to one IL-6 variant or has no binding to any IL-6 variant, leading to 211920… See the full description on the dataset page: https://huggingface.co/datasets/alchemab/il6-binding-prediction.sarscov2-binding-prediction
SARS-CoV-2 binding dataset
Dataset of 104972 antibodies screened for binding the SARS-CoV-2 HR peptide, described in Engelhart et al. (2022), were obtained from Zenodo.
Average predicted logKD values were used for classifying sequences as binders and non-binders:
logKD<3 = binders
logKD>=4 = non-binders
logKD>=3 and logKD = ambiguous; removed.
Using these criteria, we have 51590 sequences remaining; these were stratified into an 80:10:10 ratio for training, test, validation… See the full description on the dataset page: https://huggingface.co/datasets/alchemab/sarscov2-binding-prediction.her2-binding-prediction
HER2 binding dataset
HER2 binding antibodies have been obtained from the Github repo for Mason et al. (2021).
Labels for antibody sequences were generated using scripts in the above Github repo. The number of negatives and positives were balanced through random undersampling using imbalanced-learn, and sequences were deduplicated.
The dataset has:
39108 antibodies in total
22779 antibodies after undersampling and deduplication
18223 in the training set, 2278 in the evaluation set… See the full description on the dataset page: https://huggingface.co/datasets/alchemab/her2-binding-prediction.