Contrastive Domain Generalization for Cross-Instrument Molecular Identification in Mass Spectrometry

Identifying molecules from mass spectrometry (MS) data remains a fundamental challenge due to the semantic gap between physical spectral peaks and underlying chemical structures. Existing deep learning approaches often treat spectral matching as a closed-set recognition task, limiting their ability to generalize to unseen molecular scaffolds. To overcome this limitation, we propose a cross-modal alignment framework that directly maps mass spectra into the chemically meaningful molecular structure embedding space of a pretrained chemical language model. On a strict scaffold-disjoint benchmark, our model achieves a Top-1 accuracy of 42.2% in fixed 256-way zero-shot retrieval and demonstrates strong generalization under a global retrieval setting. Moreover, the learned embedding space demonstrates strong chemical coherence, reaching 95.4% accuracy in 5-way 5-shot molecular re-identification. These results suggest that explicitly integrating physical spectral resolution with molecular structure embedding is key to solving the generalization bottleneck in molecular identification from MS data.

💡 Research Summary

The paper tackles the long‑standing challenge of identifying molecules from mass spectrometry (MS) data, where the physical peak patterns measured by heterogeneous instruments often diverge dramatically from the underlying chemical structures. Conventional deep‑learning approaches treat MS identification as a closed‑set classification problem, training on a fixed set of molecules under uniform acquisition conditions. Such models fail to generalize when confronted with unseen molecular scaffolds or new instrument configurations, because they implicitly learn instrument‑specific spectral fingerprints rather than a chemically meaningful representation.

To bridge this “semantic gap,” the authors propose a cross‑modal contrastive learning framework that directly aligns MS spectra with the embedding space of a pretrained chemical language model (ChemBERTa). The core idea is to treat molecular identification as a semantic alignment problem: the spectrum encoder learns to produce vectors that lie in the same space as the structure embeddings derived from SMILES strings. By doing so, the model inherits the rich chemical knowledge encoded in the language model and can perform zero‑shot retrieval on molecules it has never seen during training.

The pipeline consists of several carefully designed components. First, the raw spectra are pre‑processed to mitigate instrument‑induced heteroscedasticity. The m/z values are transformed with a logarithm (ln(m/z + 1)) to make peak width constant across the mass range, while intensities undergo a square‑root normalization followed by division by the maximum √I, compressing the heavy‑tailed intensity distribution into a stable