Mapping the Missing Metabolome 

AIMe expands MS2-based structural annotation far beyond experimental libraries while retaining interpretable fragment assignments. 

A neuro-symbolic AI framework has expanded the searchable chemical space of tandem mass spectrometry to more than 100 million known small organic molecules, providing structural leads for metabolites that lack experimental library matches.  

Experimental tandem mass spectrometry (MS2) libraries contain spectra for fewer than one percent of known compounds, leaving most features in untargeted metabolomics without a reference match. AI Molecule Explorer, or AIMe, organizes more than 800 million predicted spectra into a searchable resource called MS2KOSMOS.  

“At the core of AIMe is DeepMS2Reasoner, a model that simulates how molecules fragment inside a mass spectrometer,” said Carla Gomes in a press release. “The result is a predicted spectrum and an annotated map of how a molecule came apart – a feature that makes AIMe's outputs interpretable in chemical terms, not just computationally useful.” 

Across three benchmark datasets, DeepMS2Reasoner outperformed the other spectrum-prediction models tested. On the National Institute of Standards and Technology (NIST) dataset, predicted and experimental spectra had an average cosine similarity of 0.83, compared with 0.61–0.75 for other approaches. 

Applied to 111 abundant but unidentified metabolites that differed between germ-free mice and animals with a gut microbiota, AIMe found close predicted matches for roughly one-third. Its fragment assignments helped researchers construct and rank candidates for spectra without close matches. Synthesis and direct MS2 comparison subsequently confirmed two previously undescribed polyamine derivatives, including an unusual macrocyclic compound also detected in human samples. 

AIMe returned putative annotations for roughly 2.69 million of 7.1 million GNPS spectral clusters, compared with around 416,000 previously annotated at the same similarity threshold. Without orthogonal evidence, however, these remain putative structural annotations rather than confirmed identifications. 

Future versions could strengthen those assignments by incorporating orthogonal evidence, including chromatographic retention times and NMR spectra, alongside the MS2 predictions. 

Imaging the Drug–Protein Complex  

Isotopically resolved MALDI FT-ICR imaging distinguishes drug-bound and unmodified protein to map covalent target engagement in tissue. 

An isotopically resolved matrix-assisted laser desorption/ionization (MALDI) imaging method has enabled direct measurement of covalent drug target engagement in tissue by distinguishing intact drug-bound and unmodified protein. 

Because covalent drugs bind irreversibly, their target engagement cannot be inferred from free-drug concentration as it can for reversible drugs. To measure it directly, the researchers optimized MALDI Fourier-transform ion cyclotron resonance mass spectrometry (FT-ICR MS) for intact proteins in mouse brain sections. 

The high resolving power of FT-ICR MS was central to the approach. By resolving individual isotopologues and averaging the 10 most abundant, the team improved image signal-to-noise while excluding interfering ions. Ratiometric analysis then calculated percent target engagement from the relative signals of bound and unbound protein at each 100 µm pixel. 

The method was tested with S-XL6, a covalent drug candidate designed to stabilize mutant superoxide dismutase 1 (SOD1) in familial amyotrophic lateral sclerosis. Imaging showed an average target engagement of 78 percent across brain sections, including the motor cortex and other disease-relevant regions. Measurements from brain homogenates reached 91 percent by MALDI MS and 88 percent by LC-MS, with the authors attributing the lower imaging value partly to detection bias against the larger drug-linked SOD1 complex. 

Intact hemoglobin provided a spatial marker for blood vessels. Its localized distribution contrasted with the widespread S-XL6–SOD1 signal, supporting penetration of the drug beyond the vasculature and across the blood-brain barrier. 

The current FT-ICR method is optimized for targets up to around 17 kDa, with larger proteins requiring further instrument or sample-preparation changes. By mapping the drug-bound protein itself, the approach could help distinguish drug exposure from engagement of the intended target in therapeutically relevant tissue. 

The First Denisovan Radius  

ZooMS and deeper proteomics identify five Denisovan remains, including the first confirmed radius from the archaic human group.  

A study of five Denisovan remains from Bianfu Cave in southwest China has identified the first confirmed radius from the archaic human group, offering a rare glimpse of Denisovan postcranial anatomy. 

More than 60,000 fragmentary bones were recovered from the site, making indiscriminate molecular screening impractical. The researchers used morphology to narrow the assemblage to 22 possible hominin fragments before applying zooarchaeology by mass spectrometry (ZooMS). Collagen fingerprints identified two parietal fragments and a partial radius as hominin; attempts to recover ancient DNA were unsuccessful. 

Liquid chromatography–tandem mass spectrometry (LC–MS/MS) provided finer taxonomic resolution. Analysis of the three bones and two previously excavated teeth recovered between five and 16 endogenous proteins per specimen. All five carried COL1A2 R996K, an amino acid variant found consistently in molecularly identified Denisovans but absent from Neanderthals and modern humans. Phylogenetic analysis independently grouped each specimen with Denisova 3. 

The remains came from layers dated to around 167,000–134,000 years ago. Proteins from the two teeth also contained abundant Y-linked amelogenin peptides, indicating that both belonged to male individuals. 

Most notably, the partial radius provides the first opportunity to examine Denisovan radial morphology. Its large proximal dimensions and probable medially oriented radial tuberosity resembled Neanderthals, whereas its overall external shape and mid-neck geometry were closer to modern humans. The authors interpret this mosaic as evidence for biomechanical demands distinct from those inferred for either group. 

Combining morphological pre-screening with ZooMS and deeper proteomics enabled Denisovan assignments where ancient DNA could not be recovered. Applying the strategy to other fragmented assemblages could expand the small pool of molecularly identified postcranial remains available for reconstructing Denisovan anatomy. 

Finding Palladium Without Fine Structure  

A machine-learning workflow recognizes elemental isotope patterns without relying on fully resolved fine structure.  

A machine-learning workflow named MEDUSA-HR has identified palladium (Pd)-containing ions in complex high-resolution mass spectra at concentrations down to 10⁻⁸ M, without requiring the fine isotope structure normally used to establish elemental composition. 

Palladium has six naturally occurring isotopes, but its characteristic isotope patterns can become difficult to recognize when low-intensity signals overlap or fine structure is unresolved. That complicates detection of trace Pd in catalytic mixtures and contaminated samples. 

MEDUSA-HR combines graph-based deisotoping with machine-learning classification to identify elements from their broader isotope distributions. The graph algorithm first groups peaks belonging to individual ions, including multiply charged species, before classifiers assess the resulting patterns for specific elements. On simulated spectra, graph-based deisotoping substantially outperformed earlier linear and gradient-boosting methods, while classifiers trained largely on synthetic spectra were validated against experimental electrospray ionization high-resolution mass spectrometry (ESI-HRMS) data. 

For a Pd complex, the model reliably identified Pd-containing ions from 10⁻³ to 10⁻⁷ M and still detected them at 10⁻⁸ M, although false-positive assignments began to appear at the lowest concentration. The workflow also tracked Pd-containing intermediates during reactions. In one operando experiment, adding potassium iodide caused Pd–Cl signals to disappear as Pd–I species emerged. 

Beyond palladium, the same architecture was extended to nickel, copper, silver, chlorine, and bromine, and tested with time-of-flight, quadrupole time-of-flight, and Fourier-transform ion cyclotron resonance mass spectra. The authors also applied it to laboratory materials, reagents, and a vehicle exhaust sample to screen for Pd contamination. 

The method cannot directly recognize monoisotopic elements, while lower resolving power can reduce performance and different instruments may require adjustment of the deisotoping threshold. The authors suggest that automated elemental screening could make routinely acquired high-resolution mass spectra more useful for tracking catalytic intermediates and detecting trace metal contamination. 

Newsletters

Receive the latest analytical science news, personalities, education, and career development – weekly to your inbox.

I have read and understand the Privacy Notice*

Subscribe