Learning Fragmentation Physics or Exploiting Sequence Priors? Benchmarking Bias in Deep Learning Models for De Novo Peptide Sequencing

Deep learning models have advanced de novo peptide sequencing, but their predictions may reflect both physics-based spectral evidence and learned peptide-sequence priors. Systematically measuring such prior-associated behavior is important for benchmarking model robustness beyond conventional proteomics data. Here, we introduce the Prior Bias Index (PBI), a general framework for measuring the extent to which model behavior shifts toward prior-associated reference patterns under controlled conditions, and implement it as DeNovo-PBI, a benchmark for quantifying prior bias in de novo peptide sequencing models. DeNovo-PBI combines benchmark dataset construction, in silico sequence and spectral perturbation workflows, PBI-based metrics, and analysis algorithms to evaluate three forms of prior-associated behavior: sequence-distribution dependence, database amino-acid-pair order preference, and mutation-group prediction consistency under shared sequence context. In addition to experimentally acquired peptide spectra, we generated in silico spectra from random, natural, and mutated peptide sequences and selectively removed fragment ions that distinguish N-terminal residue orders. Across these assays, deep learning models showed peptide-sequence-distribution-dependent performance and strong directional amino-acid-pair order preferences even when order-diagnostic spectral evidence was removed. DeNovo-PBI provides a quantitative benchmark for measuring, comparing, and interpreting learned bias in de novo peptide sequencing models.