Machine Learning on Signal-to-Noise Ratios Improves Peptide Array Design in SAMDI Mass Spectrometry

Emerging peptide array technologies are able to profile molecular activities within cell lysates. However, the structural diversity of peptides leads to inherent differences in peptide signal-to-noise ratios (S/N). These complex effects can lead to potentially unrepresentative signal intensities and can bias subsequent analyses. Within mass spectrometry-based peptide technologies, the relation between a peptide's amino acid sequence and S/N remains largely nonquantitative. To address this challenge, we present a method to quantify and analyze mass spectrometry S/N of two peptide arrays, and we use this analysis to portray quality of data and to design future arrays for SAMDI mass spectrometry. Our study demonstrates that S/N varies significantly across peptides within peptide arrays, and variation in S/N is attributable to differences of single amino acids. We apply supervised machine learning to predict peptide S/N based on amino acid sequence, and identify specific physical properties of the amino acids that govern variation of this metric. We find low peptide-S/N concordance between arrays, demonstrating that different arrays require individual characterization and that global peptide-S/N relationships are difficult to identify. However, with proper peptide sampling, this study illustrates how machine learning can accurately predict the S/N of a peptide in an array, allowing for the efficient design of arrays through selection of high S/N peptides.