PowerNovo2: A generative flow-based approach to non-autoregressive de novo peptide sequencing

Abstract

Proteomics utilizes tandem mass spectrometry (MS/MS) to determine peptide sequences, traditionally through database searches constrained by prior knowledge. De novo sequencing offers a database-free alternative but struggles with accurately modeling complex MS/MS spectra. Most current tools use autoregressive decoding, which is prone to error propagation and computationally slow. Here we present PowerNovo2, a non-autoregressive model based on generative normalizing flows. By leveraging variational inference, it effectively captures intricate token dependencies and peptide-level uncertainties. PowerNovo2 outperforms existing de novo tools in accuracy and speed, matching state-of-the-art autoregressive models like Casanovo while being 4.3 times faster. It also demonstrates competitive performance against other non-autoregressive methods such as π-PrimeNovo, particularly on long peptides and low-resolution spectra. As the first flow-based de novo sequencer, PowerNovo2 provides a scalable, accurate solution for large-scale proteomic applications.