Deep Learning Challenges in DIA Proteomics Data Analysis

Georg Wallman, co-founder and managing director of Aplusia, sat down with LCGC International at the ASMS conference to discuss the cutting edge of AI-driven mass spectrometry. Wallman's company operates as a biotech consultancy, specializing in challenging drug discovery problems for pharma and biotech clients, and his presentation at ASMS gave a vivid picture of just how rapidly the field is moving. The centerpiece of Wallman's talk was PeptDeepKontext, a universal deep learning model designed to predict peptide properties—including spectra, retention times, and charge states—across diverse instruments, laboratory conditions, and post-translational modifications.¹ The model addresses a long-standing problem: existing tools tend to fail when taken out of the specific setup in which they were trained. PeptDeepKontext instead treats inter-lab variability not as a nuisance to eliminate, but as a dimension to learn from, drawing on data from more than 50 studies spanning a wide range of LC instruments, mass spectrometers, solvents, and gradient profiles.

In this interview clip, Wallman discusses:

• What are the main challenges of applying deep learning to DIA proteomics data, such as data volume, missing values, or model generalizability across different instruments and sample types? 

Wallman also introduced fragDETR, a graph-level fragmentation approach borrowed from metabolomics and adapted for peptides. The technology unlocks richer spectral predictions —accounting for internal fragments, neutral losses, and side-chain fragmentation—yielding up to three times more identified peaks and five times greater intensity gains in challenging cases such as heavily modified peptides and non-tryptic sequences. Looking beyond instrumentation, Wallman offered a compelling perspective on where proteomics sits within pharmaceutical drug discovery. Once a niche core-facility tool, proteomics is now capable of processing hundreds of samples per day and is increasingly positioned at the centre of high-throughput drug discovery platforms, particularly for covalent drug programs and targeted protein degradation. As instrument capabilities continue to outpace what biology alone demands, Wallman sees a transformative period ahead for the field.

Reference

1. Wallman, G.; Kotb, M.; Lebedev, M.; Müller-Reif, J. B.; Mann, M. PeptDeepKontext: A Universal Model for Predicting Peptide Properties Across Instruments, Conditions, and Post-Translational Modifications Trained on a New Foundational Dataset. Presented at ASMS 2026, in San Diego, California, USA. https://asms.org/docs/default-source/conference/74th-asms-final-program_as-of-may-8-2026.pdf?sfvrsn=1234fc3_0 (accessed 2026-06-24).