πDIA-CLIP: efficient identification of highly heterogeneous proteomics data via a generalized zero-shot framework

Data-independent acquisition mass spectrometry has increasingly emerged as a cornerstone for characterizing highly heterogeneous biological systems, such as single-cell proteomics, metaproteomics, and spatial proteomics, offering unparalleled identification depth and quantification reproducibility. Current DIA analysis frameworks, however, require semi-supervised training within each run for peptide-spectrum match (PSM) re-scoring, which is prone to overfitting and lacks generalizability across diverse species and experimental conditions. Here, we present {pi}DIA-CLIP, a generalized framework shifting the DIA analysis strategy from semi-supervised training to zero-shot cross-modal representation learning through integrating dual-encoder contrastive learning and encoder-decoder architectures to establish a unified, high-precision representation for spectral features and peptides. Notably, the generalized zero-shot nature of {pi}DIA-CLIP facilitates an inference-only architecture, streamlining the analysis to achieve exceptional computational efficiency. Extensive evaluations across five distinct benchmarks demonstrate that {pi}DIA-CLIP consistently outperforms existing tools, yielding an up to 44.6% increase in protein identification alongside a reduction in entrapment identifications reaching a maximal 52.5%. Furthermore, the enhanced identification depth facilitates the discovery of novel biomarkers and the elucidation of intricate cellular mechanisms.