Correcting replicate variation in spectroscopic data by machine learning and model-based pre-processing

In this study we present a pre-processing and an augmentation approach based on Extended Multiplicative Signal Correction (EMSC) for removing and modelling replicate variation in spectroscopic data. The EMSC replicate correction method estimates replicate variation from replicate samples and integrates the estimated variation into the EMSC model. In the field of deep learning, augmentation is a frequently applied approach to deal with variability in images. In this study, we suggest augmentation of vibrational spectroscopic data with replicate variation. Replicate correction and replicate augmentation can be considered as two inverse procedures which are compared in our study. Three data sets of Fourier Transform Infrared spectra of yeasts, filamentous fungi and bacteria were used in this study to compare classification performance on genus and species level by (1) Random Forest using replicate corrected spectra vs. (2) Deep Learning using augmented spectra. Both technical and biological replicate correction/augmentation were considered.

Automated summary

This study introduces a pre-processing and augmentation approach based on EMSC to eliminate replicate variation in spectroscopic data. It compares replicate correction and replicate augmentation as two inverse procedures in terms of classification performance, using different types of spectral data sets for experimentation.