Abstract
Metabolomics, using non-targeted tandem mass spectrometry, generates rich biological data, but novel metabolites with structures absent from databases are challenging to analyze. Existing algorithms predict isolated chemical features such as molecular fingerprints or structural classes but fail to integrate them into reliable structure-level predictions, particularly for complex metabolites or spectra with high noise and sparse fragments. Here we present ComFaceID, a foundation model that revolutionizes de novo structure profiling through generating 500-dimensional embeddings from MS² spectra, enabling parallel prediction of diverse structural descriptors. ComFaceID consistently outperforms state-of-the-art tools across key tasks, including library search, classification, and fingerprint prediction, even under challenging conditions such as complex metabolomic backgrounds with noisy and information-poor spectra. By integrating outputs from these multi-task predictions, we further developed a multi-parameter framework which significantly enhances prioritization accuracy over single-parameter approaches. When applied to 3,334 actinomycete extracts, ComFaceID uncovered 6 novel compounds across 3 structural classes, including a unique hexahydroindolizine scaffold. Two compounds showed potent, broad-spectrum cytotoxicity superior to doxorubicin, with lower off-target toxicity. In an H1N1 infection model, the ComFaceID pipeline identified over 40 unannotated likely biomarkers, revealing pulmonary inflammation-induced gut metabolic remodeling marked by increased saturated fatty acyl phospholipids and reduced bile acids. By bridging spectral interpretation and novel metabolite discovery, ComFaceID establishes a new workflow for structure-informed metabolomics.
Introduction
Metabolomics—the study of small molecules in biological systems—is often hindered by the "dark matter" of unknown compounds. While tandem mass spectrometry (MS²) is the gold standard for identifying these molecules, existing algorithms struggle when a compound isn't already in a database. Current tools usually predict isolated features (like a molecular fingerprint or a chemical class) but fail to integrate them into a holistic "identity." Furthermore, their accuracy plummets when dealing with "noisy" data from crude biological extracts where signals are weak or fragmented. This creates a bottleneck: researchers often have to spend months purifying a compound before they even know if it is actually new or biologically active.
The ComFaceID Framework
To address these challenges, researchers developed ComFaceID (Compound FaceID), a deep learning foundation model. Much like facial recognition software identifies a person despite different lighting or angles, ComFaceID identifies the "structural essence" of a molecule even when the mass spectrum is noisy or incomplete.
The model uses multi-scale convolutional layers to capture local spectral features and a global attention mechanism to integrate them into a 500-dimensional embedding. It was trained using a Siamese neural network with metric learning, ensuring that the "distance" between two spectral embeddings in its latent space matches the structural similarity (Tanimoto coefficient) of the actual molecules.
To quantify how "novel" a compound is, the authors developed the Metabolite Structural Novelty Score (MSNs). This framework maps three orthogonal descriptors—structural class, fingerprint similarity, and mass deviation—into a 3D coordinate system. The further a point is from a "perfect match" (1,1,1), the higher its novelty score. This allows for prioritization of compounds at the crude extract stage, enabling labs to ignore "known" molecules and focus resources only on truly unique chemical scaffolds.
Benchmarking and Performance
The study utilized several key datasets for training and validation, including a training set of 63,339 known compounds and benchmarking sets like CASMI 2022 and an "AntitumorDataset" designed to mimic high-noise environments. ComFaceID consistently outperformed state-of-the-art tools like SIRIUS (for fingerprints) and CANOPUS (for classification), particularly for larger molecules (>300 Da) and low-quality spectra. The system is also compatible with Large Language Models (specifically DeepSeek-R1-Distill-Qwen-32B), allowing researchers to use AI agents to automate data analysis and visualization.
Real-World Applications
Discovery of Novel Anticancer Agents
When applied to 3,334 actinomycete fermentation extracts, the ComFaceID pipeline uncovered 6 novel compounds across 3 structural classes. One notable discovery was Speibonmycin A, a novel polycyclic polyketide identified using the MSNs score. In bioactivity tests, Speibonmycin A proved to be 450 times more potent than the standard chemotherapy drug doxorubicin against certain leukemia cells, while showing significantly lower toxicity to healthy cells. The model also uncovered a unique hexahydroindolizine core, a chemical architecture previously unreported in natural products.
H1N1 Biomarker Identification
In a murine H1N1 infection model, the pipeline identified 42 consistently altered features in the fecal metabolome, revealing that 56% of the dysregulated biomarkers were lipid-like molecules. This provided rare molecular evidence of how a respiratory infection triggers systemic remodeling of the gut metabolome. Specifically, the study revealed an accumulation of pro-inflammatory long-chain fatty acids and a depletion of anti-inflammatory bile acids, highlighting a clear gut-lung connection during viral infection.
Conclusion and Future Outlook
ComFaceID represents a major leap forward in structure-informed metabolomics by bridging spectral interpretation and novel metabolite discovery. While the authors note that extremely low-intensity spectra still pose challenges and the MSNs thresholds require further refinement, the integration of multi-task predictions and LLM-compatible automation points toward a future where AI agents handle the entire pipeline from raw data to hypothesis generation.
Summary updated