Label-Free Targeted Proteomics Data Analysis Workflow Selection - Benchmarking AI-based and Data-Driven Approaches

As global proteomics continues to advance, the number of identifiable proteins has increased substantially. However, this does not inherently ensure optimal quantitative performance. While targeted assays using isotope-labeled peptides can be developed, label-free strategies remain an attractive and cost-efficient option for methodological validation. Yet, systematic evaluations of data analysis workflows for label-free targeted proteomics, particularly those incorporating artificial intelligence (AI)-based tools, are still limited. Therefore, this study aimed to benchmark multiple data analysis approaches for label-free targeted proteomics as a validation framework for results obtained from global analyses. Missing-data imputation (MDI) strategies, including no MDI, k-nearest neighbors, data-driven, MSstats and AI-based methods, were evaluated, alongside consolidation and testing frameworks such as mathematical summation, best-peak selection, AI-based scaling with univariate statistics, p-value integration, MSstats Tukey's median polish or linear models and multivariate testing. Data-driven MDI combined with p-value integration consistently showed the strongest performance across accuracy, precision, specificity, and false-discovery rate, outperforming all other strategies. These findings demonstrate that careful selection of data analysis workflows can yield substantially improved quantitative outcomes compared with commonly used approaches such as simple mathematical summation. Although our conclusions are based on a controlled benchmarking dataset comprising three yeast proteins spiked into a constant human background using only one targeted approach, they can be generally applied as a default workflow for biomarker validation using the label-free approach.