Iterative Self-Supervised Signal Curation Enables High-Fidelity Peptide Profiling in Parallel Nanopore Sensing

The recent repurposing of nanopore arrays for proteomics generates extensive high-throughput datasets, but stochastic molecular translocations and sensor heterogeneity inevitably introduce massive non-informative signal artifacts. Therefore, an automated and unbiased data curation method is essential. Here, we present NanoCurator, an iterative, self-supervised CNN-LSTM autoencoder that automatically extracts high-fidelity translocation fingerprints from complex raw data. By evaluating reconstruction error and Lempel-Ziv complexity via adaptive thresholding, NanoCurator effectively segregates genuine signals from diverse noise profiles in both simulated and empirical datasets. Crucially, this curation significantly elevates classification accuracy by 1.8% for a 15-peptide panel, 2.3% for post-translational modifications, and 1.5% for heterogeneous mixtures, while reducing data volume requirements, thereby establishing a robust, highly generalizable paradigm for automated signal quality control. Ultimately, this versatile framework accelerates precise peptide identification and enables the reliable, large-scale application of massively parallel nanopore sensing in proteomics.