Count your peaks! Bucket counts and co-occurrences from as many spectra as possible.

Standard spectral similarity measures such as cosine similarity treat all peaks equally, ignoring the fact that some fragment ions (e.g. low-mass hydrocarbon cations) appear in over half of all spectra while others are highly specific. [Count your bits](https://doi.org/10.1101/2025.06.16.659994) showed that count-based molecular fingerprint variants substantially improve specificity for molecular similarity; analogous considerations apply to spectral peak matching, where frequency information is currently discarded. To enable frequency-aware alternatives, we compute corpus-level statistics from large-scale tandem mass spectrometry datasets: per-bucket spectrum counts (how many spectra contain a peak in each m/z bin) and the full pairwise co-occurrence matrix (how many spectra contain peaks in both bin *i* and bin *j*).