The metabolic readout is a set of ratios rather than a list of identified molecules. Two-photon autofluorescence from reduced NAD(P)H and from oxidized flavins produces an optical redox ratio that the authors describe as a proxy for redox balance [6]. A higher value marks a more reduced cell, which the paper says is consistent with increased glycolysis and/or impaired oxidative phosphorylation [7]. That "and/or" is the part to hold onto: a cell running more glycolysis and a cell with damaged respiration can land on the same number, and the image will not separate them.
The Raman channels are class-level too. Pump tuning to 2,930 and 2,850 cm-1 gives protein and lipid, 3,015 and 2,880 cm-1 give unsaturated and saturated lipid, and what comes out are lipid/protein and lipid-saturation ratios [8]. All four points sit inside a 165 cm-1 window [21]. The contrast is background-free and proportional to concentration [14], which is what makes those ratios worth quoting, and hyperspectral SRS can yield fuller Raman-like spectra for molecular discrimination [22], so four-point sampling is a design decision rather than the limit of the technique.
Against transcript-based work the gain is that nothing is being predicted: spatial omics infer metabolic programs from transcripts or proteins and do not directly report dynamic metabolic state or robustly link activity to cell type [11]. Against mass spectrometry imaging the trade is explicit. MALDI, DESI and SIMS map identified molecular species and, paired with multiplex protein markers, reach high sensitivity [13]; they are also mostly destructive and often trade molecular coverage for spatial resolution, which the paper says limits longitudinal and multimodal measurement [12]. REDCAT reverses the priority, offering class-level chemistry at submicron resolution with cell identity attached [3].
Registration is where the claims are won or lost. The label-free imaging runs first, CODEX typing follows, and a pipeline built on MaxFuse matches the two and links single-cell protein markers to metabolic features [3]. MaxFuse was published for integrating modalities with weakly linked features [15], which names the difficulty rather than settling it. The label-free arm alone contributes seven contrast channels that all have to land on the same cell as the antibody panel [19], and the abstract and briefing available here carry no registration accuracy figure, so every per-cell metabolic assignment rests on a step whose error rate is unstated.
The liver demonstration recovers gradients running from central vein to portal vein [10], a geometry already known, which is what a method paper should show. The Nature Methods briefing accompanying the 2026 paper calls functional histopathology a suggested use [16][20], and suggested is the right tense. That briefing also cites DBiTplus, a separate 2026 effort to put imaging-based and sequencing-based spatial omics on the same section for clinical pathology [17]. One section carrying several assays is now the thing being competed over.
