Biologically informative not available deconvolution excavates hidden features in proteomics datasets

Abstract

Missing values [not available (NAs)] in proteomics arise from true absence/low abundance and technical factors. Conventional imputation often obscures disease signals by ignoring the biological information inherent in protein absence. We present biologically informative NA deconvolution (BIND), an adaptive neighborhood-based workflow that classifies NAs as biological or technical. BIND binarizes matrices to encode presence/absence and uses context-aware k-nearest neighbors with Gaussian kernels to deconvolve NA origins, enabling clustering, biomarker discovery, and an NA-weighted correlation for protein–protein interaction (PPI) inference. In a simulation derived from the CCLE949 dataset, BIND achieved an Area Under the Curve (AUC) of 0.93, with k = 10 and γ = 3 selected for subsequent Cancer Cell Line Encyclopedia (CCLE) analyses, and remained robust across biological-to-technical NA ratios ranging from 3:1 to 1:3. In hepatocellular carcinoma tissues (n = 248), BIND distinguished tumor from adjacent normal, identified candidate biomarkers, and detected cancer-specific PPIs enriched in spliceosome and DNA replication pathways. A public BIND web server offers interactive visualization and supports single-cell proteomics, enhancing accessibility. Overall, BIND transforms NA patterns from a liability into biologically informative features for clustering, endotype discovery, biomarker identification, and PPI analysis.