Artificial intelligence for molecular discovery to illuminate the dark proteome

Mass spectrometry (MS)-based proteomics has generated vast amounts of data available in public repositories, yet much of this remains underexplored, leaving valuable biological insights obscured within uncharted data. This presentation will delve into pioneering AI-driven approaches designed to unlock the full potential of these datasets, uncovering insights that transcend the limitations of the original data analyses. First, I will introduce a deep neural network approach, called "GLEAMS," which learns meaningful embeddings from MS data across large repositories. By embedding spectra into a low-dimensional space where those generated by the same peptide cluster together, GLEAMS enables the efficient clustering and identification of vast datasets. Using GLEAMS, we analyzed 31TB of human proteomics data from the MassIVE Knowledge Base, encompassing 666 million spectra from 220 publicly available experiments. This approach allowed us to probe the "dark matter" of the human proteome, propagating peptide labels within high-quality clusters and using open modification searches to annotate 73% more spectra than conventional methods. This transformative capability redefines how we can annotate and understand uncharacterized portions of the proteome. Second, I will discuss Casanovo, an advanced AI tool for de novo peptide sequencing, emphasizing its utility for discovery-driven proteomics. Casanovo harnesses the latest advances in large language models by framing de novo peptide sequencing as a translation task from peaks in an MS/MS spectrum to sequences of amino acids. Casanovo has been trained using tryptic and non-enzymatic data, allowing it to achieve state-of-the-art de novo peptide sequencing performance across various settings. Furthermore, Casanovo has yielded new insights into immunopeptidomics and metaproteomics data, as well as elucidating the dark human proteome in ways that sequence database searching alone cannot achieve. These insights pave the way for more comprehensive views of complex proteomes. Together, GLEAMS and Casanovo demonstrate the power of combining open-access data with cutting-edge AI to push the boundaries of MS-based proteomics. I will showcase how these approaches unlock previously hidden layers of biological insight, offering a glimpse into the future of proteomics powered by AI.