MS²-SMILES AlignNet

A contrastive objective is indifferent to how wrong a wrong answer is. Trained to pull a spectrum toward its true structure and push it away from everything else, it treats a positional isomer of the correct metabolite exactly as it treats an unrelated lipid — both are negatives, both get shoved to the far side of the embedding space. That is the wrong geometry for compound identification, where the candidates competing for the top rank were all drawn from the same narrow precursor-mass window and are therefore close chemical relatives of one another.

MS²-SMILES AlignNet, built at Heilongjiang University of Chinese Medicine, answers this with a second training term. Alongside the usual InfoNCE loss, a Tanimoto-MSE term requires the cosine similarity between two embeddings to reproduce the Tanimoto coefficient between the corresponding Morgan fingerprints. Alignment and chemical neighbourhood are optimised together, so the shared space is not only correct on the diagonal but ordered away from it — which is what decides a ranking among near-isomers.

The surrounding task is cross-modal retrieval: instead of matching a query spectrum against a library of measured reference spectra, the model embeds spectra and molecular structures into one 256-dimensional space and ranks candidate SMILES by similarity, so a compound can be named without anyone having previously run it on an instrument. CSU-MS² is the closest published neighbour and the reference point the paper measures against.

#Key Features

• Hybrid alignment loss: the total objective is β·InfoNCE + α·Tanimoto-MSE, with α = 0.5, β = 1.0 and contrastive temperature τ = 0.07. The second term is the paper's central design claim.
• Dual-branch spectral encoder: one branch consumes a sparse 0.05 Da binned vector spanning 20–2000 m/z plus neutral-loss bins; the other consumes up to 100 peaks as a sequence with sinusoidal m/z position embeddings. Their two 256-dimensional outputs are concatenated and projected back to 256.
• Acquisition conditions as model input: a 135-dimensional global feature vector encodes instrument, adduct, collision energy, charge and ionisation mode, diagnostic ions and neutral losses, which is how the model holds up across heterogeneous acquisition settings.
• Graph and fingerprint fusion on the molecular side: a three-layer graph convolutional encoder with multi-head global attention readout is fused with a 2048-bit Morgan fingerprint, rather than relying on either representation alone.
• Precomputed structure embeddings: because the molecular branch is independent of the query, a whole structure library is embedded once and each new spectrum costs a single dot product against the mass-filtered candidates.

#Technical Details

Both spectral branches use CLIP-style residual attention blocks — six layers, eight heads, QuickGELU activations — while the molecular branch is graph-convolutional rather than attentional. Atoms are featurised into 74 dimensions covering element, valence, hybridisation, aromaticity, ring membership, chirality and CIP code, with bonds carried in a six-channel adjacency tensor. Embeddings from both sides are L2-normalised before the loss.

Training used more than 1.6 million high-quality spectrum–structure pairs. On the unified test subset of the CASMI 2022 benchmark, the model reports Recall@1 of 54.23% in positive ion mode and 45.37% in negative ion mode, the latter 14.81 percentage points above CSU-MS². Generalisation and structural-isomer discrimination were re-examined on CASMI 2016. The paper reports no total parameter count.

#Applications

The model is aimed at the annotation bottleneck in untargeted metabolomics: an LC-MS/MS run routinely yields thousands of features, and only a small fraction match a reference spectrum in any library. Ranking a precursor-mass-filtered candidate set turns those unmatched features into short, prioritised shortlists that an analyst can pursue with authentic standards or orthogonal evidence. Natural-product and traditional-medicine chemistry — the group's own setting — is a natural fit, since those extracts are dominated by scaffolds that spectral libraries cover thinly and by isomeric series that generic similarity scoring cannot separate.

#Impact

The Tanimoto-MSE term is a compact and portable idea: any contrastive spectrum–structure model can add a similarity-preserving regulariser without changing its encoders, and the negative-mode gain suggests the benefit is largest exactly where fragmentation is sparsest and ranking hardest. Two caveats bound how far the reported numbers should be read. The headline comparison is against a single method rather than a panel including the established fingerprint-prediction tools, and the public repository ships the complete training and inference source — dataset construction, architecture, losses, retrieval — but no checkpoint file, so reproducing the CASMI results means retraining from a corpus the paper describes only by size. As a retrieval model it also ranks structures that are already in the candidate library; it does not propose a structure that is absent from it.