Supplementary information

Library size and structure, not synthetic labels, explain apparent
benchmark-experiment discrepancies in antibody machine learning

Kim D., Bae M.

------------------------------------------------------------------------

STable1_reported_quantities.csv
    52 rows.  every quantity reported in the article, with its source file

STable2_multiple_comparisons.csv
    18 rows.  Holm-Bonferroni and Benjamini-Hochberg corrected p-values by family

SData1_matched_library_2DD8_S.csv
    6,000 rows.  6,000 CDRH3 variants with Absolut! energies, SARS-CoV spike RBD

SData2_exp1_alphaseq_baselines.csv
    22 rows.  random split vs prospective evaluation

SData3_exp2_search.csv
    270 rows.  search on the unmatched 6.9M lattice

SData4_exp3_enrichment.csv
    12 rows.  panning re-analysis, AUROC by task

SData5_panning_diversity_rarefied.csv
    9 rows.  panning diversity after rarefaction

SData6_exp9b_esm_perpos.csv
    216 rows.  one-hot vs ESM-2, unmatched libraries

SData7_exp12_featurisation.csv
    9 rows.  AlphaBind third-party prospective campaign

SData8_exp14_crossover.csv
    63 rows.  ridge vs CNN crossover, 3 splits

SData9_exp16_model_zoo.csv
    504 rows.  seven models across four measured targets

SData10_exp18_noise_recheck.csv
    150 rows.  noise injection on the matched library

SData11_exp20_matched_search.csv
    225 rows.  search at matched size and structure

SData12_exp21_seed_expansion.csv
    100 rows.  one-hot vs ESM over 20 splits

SData13_exp21_per_size_tests.csv
    5 rows.  paired tests per training size

SData14_exp22_crossover_seeds.csv
    252 rows.  ridge vs CNN crossover, 12 splits

------------------------------------------------------------------------

The matched library is the data generated in this work. Each row is one
CDRH3 sequence with the lattice binding energy computed by the Absolut!
binary and its margin against the parent: all 286 single mutants of the
parent plus random double and triple mutants, 6,000 sequences in total.

A second library, for influenza N9 neuraminidase, was built and then
withheld when its parent energy disagreed with the released library.

STable1 maps every number in the article to the result table it came
from, so any quantity in the text can be traced without rerunning the
analysis.

All files are UTF-8 CSV with a header row.
