LitScrape

This is a protein binder interaction dataset built from multiple design papers and supplementary information source files.

Each row is a target-binder pair with:

• the target name and target sequence
• the binder identifier and binder sequence
• a binary interaction label
• the paper/source of origin

Columns

• source_publication: paper or source string associated with the interaction
• target: target name
• target_sequence: target amino-acid sequence
• binder_id: binder identifier from the source dataset
• binder_sequence: binder amino-acid sequence
• label: binary interaction label (1 = positive, 0 = negative)
• binding_affinity_nm: binding affinity in nanomolar units when reported; blank otherwise

Construction

We did not use a threshold to determine whether an interaction is binding or non-binding. Instead, we used the author-designated labels. Some methods only include binder sequences for positives, while others include the binder sequences for negatives in the supplementary information. We include the author reported binding affinities whe possible, noting that each method may use a different technology to measure this affinity.

Sources

• [Overath et al.](Predicting experimental success in de novo binder design: a meta-analysis of 3,766 experimentally characterised binders)
• Latent-X
• Dickinson PPI Challenge
• Bindcraft Supplementary Information
• AlphaProteo Supplementary Information
• PPIFlow
• RFDiffusion Supplementary Information