Single cell RNA-Seq (scRNA-seq) and other profiling assays have opened new windows into understanding the properties, regulation, dynamics, and function of cells at unprecedented resolution and scale. However, these assays are inherently destructive, precluding us from tracking the temporal dynamics of live cells, in cell culture or whole organisms. Raman microscopy offers a unique opportunity to comprehensively report on the vibrational energy levels of molecules in a label-free and non-destructive manner at a subcellular spatial resolution, but it lacks in genetic and molecular interpretability. Here, we developed Raman2RNA (R2R), an experimental and computational framework to infer single-cell expression profiles in live cells through label-free hyperspectral Raman microscopy images and multi-modal data integration and domain translation. We used spatially resolved single-molecule RNA-FISH (smFISH) data as anchors to link scRNA-seq profiles to the paired spatial hyperspectral Raman images, and trained machine learning models to infer expression profiles from Raman spectra at the single-cell level. In reprogramming of mouse fibroblasts into induced pluripotent stem cells (iPSCs), R2R accurately (r>0.96) inferred from Raman images the expression profiles of various cell states and fates, including iPSCs, mesenchymal-epithelial transition (MET) cells, stromal cells, epithelial cells, and fibroblasts. R2R outperformed inference from brightfield images, showing the importance of spectroscopic content afforded by Raman microscopy. Raman2RNA lays a foundation for future investigations into exploring single-cell genome-wide molecular dynamics through imaging data, in vitro and in vivo .