Leave-one-batch-out cross-validation reveals strong batch effects in Raman spectroscopy of yeast cultures

We’re interested in using spontaneous Raman spectroscopy as a readout for the chemical state of biological systems. Previous efforts have reported that Raman spectroscopy can be used to distinguish between different strains, cell types, and physiological states. In an effort to obtain analogous results relevant to our own work, we acquired spontaneous spectra of two small collections of yeast strains, then trained conventional machine learning models to predict both strain and species identity from the Raman spectra alone. We found that, in this dataset, experimental batch effects dominated strain-level biological signals but not species-level signals. As similar batch effects may be present in other studies, careful experimental design is warranted, especially when machine learning is used for analysis. This work is primarily for researchers using Raman spectroscopy to detect meaningful differences between biological variables of interest (e.g., species, strain, growth condition, cell type, cell state, etc). We're sharing these results to call attention to the existence of experimental batch effects in spontaneous Raman spectra of biological samples, and of the consequent importance of rigorous experimental design and cross-validated analysis when using this technique.