Inter-isolate similarity networks reveal task-dependent graph signal for antimicrobial resistance prediction from MALDI-TOF mass spectra.

Antimicrobial resistance (AMR) prediction from matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectra is typically formulated as instance-wise classification, even though clinical isolates are naturally related by spectral similarity. We ask whether isolate-level similarity carries AMR information beyond per-isolate spectral classifiers, and what breaks first when we move such a model to another clinical site. We construct k-nearest-neighbor graphs from train-normalized spectra and introduce Hybrid-GNN, which combines a per-isolate multilayer perceptron encoder with GraphSAGE aggregation through a learnable gate initialized at zero, so that any graph contribution must be learned rather than imposed. Across 13 DRIAMS-A bacterium-antibiotic combinations spanning Escherichia coli, Klebsiella pneumoniae, and Staphylococcus aureus, Hybrid-GNN achieved statistically significant aggregate AUPRC-R gains over logistic regression, random forest, support vector machine, LightGBM, and multilayer perceptron baselines, although the margin over the matched multilayer perceptron control was modest. Component ablations, simple-neighbor controls, and graph neural network baselines showed that graph aggregation is useful but task-dependent. Zero-shot evaluation on DRIAMS-B, DRIAMS-C, and DRIAMS-D showed that cross-site transfer fails through a specific, diagnosable mechanism: site-level covariate shift inverts neighborhood attachment, so resistant external isolates attach to source-site neighborhoods less enriched for resistance than those of susceptible isolates, and the graph branch then propagates the wrong prior. The inversion was most pronounced for Klebsiella beta-lactam combinations. Inter-isolate similarity networks are therefore both a source of task-dependent predictive signal and a diagnostic for when a MALDI-TOF AMR model can be transported between sites and when domain adaptation is required.