Adapting censored regression methods to adjust for the limit of detection in the calibration of diagnostic rules for clinical mass spectrometry proteomic data

TL;DR

This paper introduces a censored data approach to handle detection limits in mass spectrometry proteomic data, improving calibration of diagnostic rules for pancreatic cancer detection.

Contribution

It adapts censored regression techniques to account for detection limits, enhancing prediction accuracy in proteomic mass spectrometry analysis.

Findings

Censored regression methods outperform traditional approaches in predictive accuracy.

Replacing missing spectral data with estimated averages improves model calibration.

The approach effectively handles low-abundance protein measurements in mass spectrometry.

Abstract

Despite the recent advances in mass spectrometry (MS), summarizing and analyzing high-throughput mass-spectrometry data remains a challenging task. This is, on the one hand, due to the complexity of the spectral signal which is measured, and on the other, due to the limit of detection (LOD). The LOD is related to the limitation of instruments in measuring markers at a relatively low level. As a consequence, the outcome data set from the quantification step of proteomic analysis often consists of a reduced list of peaks where any peak intensities below the detection limit threshold are reported as missings. In this work, we propose the use of censored data methodology to handle spectral measurements within the presence of LOD, recognizing that those have been censored due to left-censoring mechanisms on low-abundance proteins. We apply this approach to the particular problem of calibrating prediction rules through prior estimation of the average isotope expression in MALDI-FTICR mass-spectrometry data, collected in the context of a pancreatic cancer case-control study. Our idea is to replace the set of incomplete spectral measurements with the average intensity estimates and use those as new input to a prediction model. We evaluate the proposed methods, with respect to their predictive ability, by comparing their performance with the one achieved using the complete information as well as alternative/competitive methods to deal with the LOD.