Distance Metrics and Cross-Instrument MS/MS Transfer
Benchmarking Distance Metrics and Spectral Embeddings for Cross-Instrument MS/MS Transportability
Six domain-adaptation distance metrics were computed on nine spectral embeddings, from fingerprints to a 116M-parameter pretrained transformer and two contrastive models, across 25 MassBank instrument types (101,359 spectra). No pair-level distance metric predicts cross-instrument spectral similarity: under the Quadratic Assignment Procedure, five of 54 embedding × metric combinations reach uncorrected p < 0.05 and none survives Westfall–Young familywise correction (global p = 0.268). A synthetic positive control detects m/z shifts as small as 0.3 Da, so the null is not an insensitive pipeline. Up to 51% of spectral similarity variance sits in the compound × instrument-pair interaction: the same compound transfers well across one instrument pair and poorly across another, which a single scalar per pair cannot carry. Resolution class match between two instruments’ mass analyzers predicts transfer better than any embedding (ρ = +0.682 across 31 pairs).