This tool evaluates the humanness (human-likeness) of antibody sequences. We use antibody language models trained on large-scale antibody sequence data to compute perplexity (PPL) for input sequences — lower PPL indicates closer proximity to the human antibody sequence distribution.
The output includes each sequence's PPL, percentile within the corresponding human reference distribution, and humanness score.
Training data scale (approximate): human VH ~15M, human VL ~17M, VHH ~18M; plus ~3.7M human paired (VH-VL) records.
In addition, we have compiled PPL values for 107 marketed antibody drugs (tmp/antibody_ppl.csv), which will be displayed alongside the human reference distribution for intuitive comparison of input sequence humanness.
1. Antibody Amino Acid Sequences (supports up to 10 FASTA entries):
Two input modes are supported: ① single-chain sequence (VH or VL, no |); ② paired sequence (VH|VL, exactly one |, with heavy and light chains on either side), e.g.: QVQL...|DIQM....
Parsed sequences: 0, Total residues: 0
Model Performance (Current Version)
Antibody MLM Model Performance: PPL = 1.39 Acc = 0.9162 Antibody Autoregressive (GPT) Model Performance: PPL = 1.47 Acc = 0.8974
Human Reference PPL Distribution (random sample of 50,000)
[Human Paired (VH-VL)] count = 50000 mean = 1.3932574 std = 0.2717338 min = 1.0743669 max = 4.3343801 percentiles: p1 = 1.1153903 p5 = 1.1390824 p10 = 1.1522373 p25 = 1.1818369 p50 = 1.2772889 p75 = 1.5558073 p90 = 1.7858048 p95 = 1.9270052 p99 = 2.2393836 [Human VH] count = 50000 mean = 1.5204114 std = 0.3366472 min = 1.1229296 max = 4.9564614 percentiles: p1 = 1.1613247 p5 = 1.1914894 p10 = 1.2172106 p25 = 1.2833952 p50 = 1.4067207 p75 = 1.6604064 p90 = 1.9894504 p95 = 2.2036375 p99 = 2.6674972 [Human VL] count = 50000 mean = 1.4069984 std = 0.2291513 min = 1.1429344 max = 4.5616646 percentiles: p1 = 1.1683489 p5 = 1.1841683 p10 = 1.1988634 p25 = 1.2449281 p50 = 1.3433586 p75 = 1.4984691 p90 = 1.6906679 p95 = 1.8413126 p99 = 2.2237684
Lower PPL indicates closer proximity to the human antibody sequence distribution. Humanness scores and recommendations will be provided based on percentile ranks.
Last updated: 2026-05-26