Non-destructive prediction of transcriptomic profiles by Raman microscopy

Raman microscopy can report on whole single-cell molecular compositions in both comprehensive and non-destructive manners. However, molecular compositions of cells are diverse and compounds such as proteins have severe spectral overlaps, making them nearly intractable to interpret. Instead of pursuing the spectral decomposition, we show that transcriptomic profiles of Schizosaccharomces pombe and Escherichia coli can be computationally linked and be predicted from their single-cell Raman spectra. Our method employs the low-dimensional structure of transcriptomes, and learns a non-linear linkage between the transcriptomes and Raman spectra. Permutation tests show that the probability of accidentally finding the same prediction precision level is extremely low (p-value<0.0001), suggesting that the prediction is real. These results demonstrate that whole-cell Raman spectra could unravel cellular omics information in non-destructive manners, opening the possibility of conducting living-cell genomic analyses.