MSACL 2025 Abstract

Unlocking Latent Phenotypic Predictive Power in Unstructured Molecular Data via Self-Supervised Machine Intelligence

Gabriel Asher (1), Léa Amar (1), Baljit Ubhi (1), Jennifer M. Campbell (1), Niall O’Connor (1), Jack Geremia (1)
(1)Matterworks, Inc., Somerville, MA 02143, USA

Baljit Ubhi, Ph.D. (Presenter) Matterworks

Presenter Bio: Baljit is a seasoned life sciences executive with over 20 years of experience driving innovation at the nexus of science, technology, and business. She has held leadership roles across commercial strategy, product management, and marketing in the biotech and analytical instrumentation sectors.

Currently serving as Chief Commercial Officer at Matterworks, Baljit leads the commercialization of AI-driven technology for the application of predictive biology. Her prior roles include executive positions at Panome Bio and MOBILion Systems, where she spearheaded go-to-market strategies for next-generation technologies.

Beyond her corporate leadership, Baljit is a strategic advisor, and board member collaborating with early- to growth-stage companies to scale their commercial operations and product-market fit. She is the founder and president of South Asians in Mass Spectrometry (SAMS), advocating for diversity and inclusion in STEM fields.

Relevant Financial Disclosures (within past 24 months, reported on Aug 07, 2025) No relevant financial relationship(s) to disclose. Conflict mitigation NOT REQUIRED.

INTRODUCTION:
Traditional mass spectrometry (MS)-based omics workflows rely on human-curated features such as peaks and MS/MS identifications. While effective, these approaches force structure onto inherently unstructured data, discarding large volumes of raw signals that may contain valuable predictive information.

OBJECTIVES:
To overcome these limitations, we developed Large Spectral Models (LSMs)—self-supervised foundation models trained on diverse, unlabeled MS spectra. LSMs learn a semantic representation of raw mass spectral data, enabling phenotype prediction directly from instrument output without prior peak picking or molecular annotation.

METHODS:
In this study, we fine-tuned LSMs on small, labeled phenotype datasets to demonstrate predictive capabilities in real-world biological contexts. Using public LC-MS data from Sah et al. (2024) on ovarian and cervical cancer patients, we trained a predictive model using raw MS spectra alone.

RESULTS:
The LSM-based approach significantly outperformed a conventional supervised method based on manually selected lipid biomarkers. Our model achieved mean accuracy of 0.96, sensitivity of 0.98, and specificity of 0.95, with a tenfold reduction in false negative error rates. Importantly, LSM embeddings captured latent biochemical signals not detected through traditional workflows. Visualization of the reduced LSM representations revealed clear phenotypic clustering, and model benchmarking across 1,000 test/train splits confirmed reproducibility and robustness. By removing the need for feature engineering or annotation, LSMs reduce time-to-insight and expand access to predictive biology for users with limited data science expertise.

CONCLUSION:
This work illustrates a scalable, application-agnostic path for phenotype prediction from raw MS data. As a first-of-its-kind foundation model in mass spectrometry, the LSM approach unlocks unstructured molecular information previously discarded, redefining how spectral data can drive diagnostics, translational research, and precision medicine.