Non-Destructive Prediction of Transcriptomes from Single-Cell Raman Microscopy

Abstract

Raman micro-spectroscopy is a non-destructive imaging technique that can potentially monitor whole-cell molecular compositions in vivo. However, due to the complex molecular compositions of cells, it remains unknown whether cellular Raman spectra permit comprehensive quantification of biomolecules in a useful fashion. To test this, we compared fingerprint Raman spectra in the 700-1800 cm−1 fingerprint region with concurrently measured RNA-seq transcriptomic data of Schizosaccharomyces pombe and Escherichia coli grown under a variety of culture conditions aimed at activating different portions of the transcriptome. Using partial least squares regression and machine learning algorithms, we find that most of the variance of transcriptomes can be represented by a small set of eigenvectors, and linearly linked with the changes of Raman spectra. The estimated transformation parameters, and low-dimensionality of transcriptomes, allowed us to predict reliably the expression profiles of thousands of transcripts from Raman spectra. Interestingly, in S. pombe, ncRNAs contributed to the Raman-transcriptome linearity more significantly than mRNAs, which supports their role in coordinating cellular molecular compositions. These results show that whole-cell Raman spectra can unravel cellular omics information in a non-destructive manner, and opens the possibility of conducting live-cell omics.