Fig. 1: Study design and peak-calling performance of uMAIA and other algorithms.

a, Experimental design. Zebrafish embryos at selected developmental stages were sectioned, and MALDI-MSI was performed for alternating sections. b, Four-dimensional metabolic atlas analysis of regional and developmental trends. c, Illustration of uMAIA modules: (1) adaptive peak calling: image extraction from MSI acquisitions, (2) featurization: identifying ions across acquisitions and (3) normalization: reducing experimental variability in intensity distributions. d, Peak positions (Da) and intensities across three spectra (indicated with colors) for three molecules of comparable m/z (estimated values shown with dashed lines). Px., pixel. e, Scatterplot between mass shifts of two molecules of similar mass. Top right: Pearson’s R indicated. f, Heatmaps of extent and direction of mass shifts in space for molecules in d and two other molecules of similar mass. g, Illustration of the peak-calling approach. Spectra are reduced to a frequency (across pixels) histogram (left). Plots of 1,000 spectra stacked above corresponding histograms for 3 representative cases (right). h, Overview of uMAIA’s adaptive peak-calling algorithm (left). Intervals are iteratively expanded in the frequency histogram, decreasing a threshold (t₀, t₁, t₂) as new intervals are considered. Peak intensity is integrated for each interval across spectra, generating images (right). i, Representative example of bins called using MALDIquant, Mirion and uMAIA over the frequency histogram (top) and corresponding images (bottom). Binning intervals shown by lines (color coded by method). j, Frequency histograms (top) and images (bottom) featuring a case of different peak-calling behavior among MALDIquant, Mirion and uMAIA. k, Sensitivity analysis on real data comparing uMAIA, MALDIquant and binning for Orbitrap and Fourier-transform ion cyclotron resonance (FTICR) MS acquisitions. Venn diagrams depict set sizes of molecules passing three assessment criteria to estimate true positive (TP) counts. Numbers within Venn diagrams indicate quantity of high-confidence TPs satisfying all criteria. FP estimate indicated (gray number within outer circle). l, Left: line plot tracking performance (average mutual information score) of uMAIA, binning and MALDIquant on simulated spectra (n = 50; shaded area depicts s.d.). Right: heatmaps displaying average mutual information score over simulations for different SNRs and peak densities (n = 50).