AbBFN2 Training Data

The dataset is a filtered, cleaned and split subset of the Observed Antibody Space (OAS) used during training of AbBFN2. The dataset is constructed by downloading all paired sequences found in the OAS on May 24, 2024. Sequences with more than eight residues missing from either terminus, unproductive sequences based on OAS labelling, and sequences missing conserved cysteine residues were removed. These sequences were then clustered first based on CDR3 loops followed by the whole sequence, as described in the AbBFN2 manuscript, and split into the train, test, and hold-out sets (99%, 0.5%, 0.5%) based on clusters, such that no cluster can appear in multiple splits. Here, VH and VL amino acid sequences (derived from OAS), species (derived from OAS), light chain loci (inferred based on gene labels), V(D)J gene calls (derived from OAS), and Therapeutic Antibody Profiler (TAP) scores (derived using an in-house implementation) are provided. The dataset contains three CSV files with the following columns: h_sequence (string): The VH sequence l_sequence (string): The VL sequence species (string): The species that the sequence originates from l_locus (string): The light chain locus derived from the LV gene label h_v_call (string): The HV gene label assigned in OAS h_d_call (string): The HD gene label assigned in OAS h_j_call (string): The HJ gene label assigned in OAS l_v_call (string): The LV gene label assigned in OAS l_j_call (string): The LJ gene label assigned in OAS psh (float): The PSH score assigned using an in-house implementation of TAP ppc (float): The PPC score assigned using an in-house implementation of TAP pnc (float): The PNC score assigned using an in-house implementation of TAP sfvcsp (float): The SFvCSP score assigned using an in-house implementation of TAP