We report the results of a VAMAS (Versailles Project on Advanced Materials and Standards) interlaboratory study on the identification of peptide sample TOF-SIMS spectra by machine learning. More than 1000 time-of-flight secondary ion mass spectrometry (TOF-SIMS) spectra of six peptide model samples (one of them was a test sample) were collected using 27 TOF-SIMS instruments from 25 institutes of six countries, the U. S., the U. K., Germany, China, South Korea, and Japan. Because peptides have systematic and simple chemical structures, they were selected as model samples. The intensity of peaks in every TOF-SIMS spectrum was extracted using the same peak list and normalized to the total ion count. The spectra of the test peptide sample were predicted by Random Forest with 20 amino acid labels. The accuracy of the prediction for the test spectra was 0.88. Although the prediction of an unknown peptide was not perfect, it was shown that all of the amino acids in an unknown peptide can be determined by Random Forest prediction and the TOF-SIMS spectra. Moreover, the prediction of peptides, which are included in the training spectra, was almost perfect. Random Forest also suggests specific fragment ions from an amino acid residue Q, whose fragment ions detected by TOF-SIMS have not been reported, in the important features. This study indicated that the analysis using Random Forest, which enables translation of the mathematical relationships to chemical relationships, and the multi labels representing monomer chemical structures, is useful to predict the TOF-SIMS spectra of an unknown peptide.

Time-of-flight secondary ion mass spectrometry (TOF-SIMS) is extensively employed for the structural analysis of the outermost surfaces of organic materials, including biological materials, because it provides detailed compositional information and enables high-spatial-resolution chemical mapping. In this study, a combination of TOF-SIMS and data analysis was employed to evaluate biological materials composed of numerous proteins, including unknown ones. To interpret complicated TOF-SIMS data of human hair, an autoencoder, a dimensionality reduction method based on artificial neural networks, was applied. Autoencoders can be used to perform nonlinear analysis; therefore, they are more suitable than principal component analysis (PCA) for analyzing TOF-SIMS data, which are influenced by the matrix effect. As a model sample data, the TOF-SIMS depth profile of human hair, acquired via argon gas cluster ion beam sputtering and Bi32+ primary ion beam, was employed. Useful information, including the characteristic distributions of amino acids and permeated surfactants on the outermost surface of the hair, was extracted from the results obtained from the autoencoder. Furthermore, the autoencoder extracted more detailed features than did PCA. Therefore, autoencoders can become a powerful tool for TOF-SIMS data analysis.

In this study, the authors evaluated the distribution of low-abundance beauty ingredients in human skin tissues. The distribution of collagen tripeptide, a beauty ingredient, in the human skin was evaluated by applying multivariate curve resolution (MCR) to the time-of-flight secondary ion mass spectrometry mapping data, including reference information. The intensity of the secondary ion peaks was increased by the accumulation of secondary ion intensity in the depth direction obtained by argon cluster sputtering. Consequently, the collagen tripeptide distribution in the skin was evaluated by separating it from collagen peptide, although the conventional analysis was difficult because of the dilution of the collagen tripeptide in the skin. Additionally, the distribution of the collagen tripeptide in the skin may be determined with an autoencoder. In this study, the distribution of small amounts of beauty ingredients (the collagen tripeptide) in the skin was observed by integrating the secondary ion intensity in the depth direction with Ar cluster sputtering and applying MCR or autoencoder.