The paper shows that a deep learning model can predict chemically informative Raman spectra from standard brightfield microscope pictures of single cells. Raman spectroscopy gives label‑free molecular “fingerprints” of cells but needs slow, expensive instruments. Brightfield imaging is fast and common but normally only shows shape and contrast. Pic2Spec aims to get the chemical detail of Raman without the hardware, by computing spectra from routine images.

The authors built Pic2Spec, a physics‑informed generative neural network. They trained it on pairs of co‑registered brightfield images and measured Raman spectra from both mammalian and bacterial cells. The model learns a shared biochemical representation that links image features to vibrational spectral structure. On their test sets the generated spectra matched real measurements very closely: about 98% cosine similarity and roughly 94–95% Pearson correlation. The generated spectra also preserved key biochemical peaks and population‑level distributions.

At a high level Pic2Spec uses a latent‑variable approach inspired by variational autoencoders. The model compresses image information into a low‑dimensional biochemical code and then decodes that code into a Raman spectrum. Convolutional decoders were tailored to the spectral output so the network learns which image patterns correspond to which spectral bands. In bacterial tests, the inferred spectra could discriminate mutation‑driven states and predict expression of green fluorescent protein (GFP) with about 88% accuracy, outperforming standard image‑only analysis by roughly 20%.

If robust and generalizable, this approach could let ordinary microscopes act as “virtual spectrometers.” That would remove the need for costly Raman hardware and long acquisition times for many applications. Possible uses include higher‑throughput molecular phenotyping, drug and strain screening, and monitoring cells over time without labels or destructive assays. The authors present this as a first step toward making label‑free biochemical profiling more accessible.