Deposit optics, film thickness and acquisition date confound Raman spectra of dried droplets

Laboratory mixtures of known composition serve as ground truth for explainable AI (XAI) in vibrational spectroscopy, on the assumption that the class difference is the spectrum of the added molecule. We tested this assumption on a public Raman dataset (785 nm) of 16 dried droplets in a 2 × 2 design: a phosphatidylcholine–phenylalanine–lactate base with or without collagen–mucin and with or without DNA. Besides the 16,107 spectra, we used three records stored with the data but not analysed before: white-light images, LiveTrack focus heights and acquisition timestamps. All inference was at the droplet level. DNA raised the base bands more than any DNA band, and the two added components were not additive (p = 0.032). DNA doubled the Raman yield of the deposit (×1.97, p = 2 × 10-4). Deposit optical density in the white-light images predicted yield across droplets (r = 0.93) and within most droplets, and with optical density in the model the DNA effect fell to ×1.05 (95% CI 0.77–1.44). Focus heights agreed with a thicker film: yield rose with local height in DNA-free deposits but much less with DNA (p = 3 × 10-4). Band ratios to phenylalanine still changed with DNA after adjustment for optical density. Five maps had been acquired 8–15 days after the rest; they showed a 3.2-fold higher background, and acquisition date was predictable from held-out droplets. The DNA class therefore combines deposit optics, a modest compositional change and a date-dependent background. Attributions should be judged against this decomposed effect rather than against reference bands of DNA.