From Unknowns to Insights: Bioinformatics For Discovery in Non-Targeted Screening

Non-targeted screening (NTS) using high-resolution tandem mass spectrometry (MS/MS) has become a cornerstone of analytical science across domains such as environmental monitoring, food safety, forensics, and exposomics. Yet despite its broad utility, NTS is still fundamentally limited by low annotation rates: the vast majority of MS/MS spectra remain unidentified. In this talk, I will present a collection of bioinformatics approaches that overcome this bottleneck and enable large-scale, context-rich spectral interpretation.

First, I will introduce the nearest neighbor suspect spectral library, an open resource built from repository-scale molecular networking of over 500 million MS/MS spectra. This library contains 87,916 high-quality spectra of molecular analogs structurally related to known compounds. Leveraging the suspect library routinely doubles annotation rates across a wide range of NTS applications, from environmental and food samples to human biofluids, greatly expanding the accessible chemical space.

Next, I will highlight SIMBA, a transformer-based deep learning framework for predicting structural similarity directly from MS/MS spectra. Unlike traditional similarity metrics such as modified cosine, SIMBA captures complex spectral relationships involving multiple chemical modifications. Its predictions are both quantitatively powerful and interpretable, enabling the discovery of molecular families and structural motifs that are invisible to conventional search strategies.

Finally, I will present a conceptual and computational framework for reference data-driven metabolomics, which reframes spectral interpretation by asking not just “what is this molecule?” but rather “where did this signal likely come from?” By linking reference spectral data to curated metadata, this approach enables pseudo-annotation of unknown spectra based on their occurrence patterns across reference samples. Applied to a comprehensive food reference dataset, it resulted in a ≥500% increase in the number of interpretable MS/MS spectra, allowing the reconstruction of detailed individual dietary profiles from biological samples, even when most compounds remained unidentified.

Together, these innovations represent a shift toward scalable, context-aware analysis of NTS data. They empower researchers to move beyond isolated identifications toward a more holistic understanding of the origin, structure, and significance of the myriad chemical signals detected in real-world samples—where every spectrum becomes a stepping stone toward discovery.