alchemab/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 sequences. This was de-duplicated, leading to a total of 14,467 sequences. Non-binder sequences were randomly under-sampled using imbalanced-learn.

During stratification, we also ensure that any two sequences in the training/validation/test sets have a minimum Levenshtein distance of 1 across the CDR3 region.

In total there are

• 1310 sequences in training
• 163 sequences in validation
• 163 sequences in test.

vh_full	label
EVQ...	1
QVQ...	0
EVQ...	1
EVK...	0

License

The data is distributed under a CC-BY-NC 4.0 license.

References

• AVIDA-hIL6 publication
• ANARCI