DiffNovo

A tandem mass spectrum of an unknown metabolite is a list of fragment masses and intensities, and the standard way to read it is to look it up: score the spectrum against a reference library and return the best-matching compound. That works only for compounds somebody has already measured. In an untargeted metabolomics run most detected features are not in any library, and a lookup can only ever return a molecule that is already known, never propose one that isn't.

DiffNovo treats the problem as generation rather than retrieval. It writes a SMILES string from scratch, conditioned on the spectrum, using a conditional diffusion model: a forward process adds Gaussian noise to a continuous embedding of the SMILES tokens, and a learned reverse process denoises it back to a valid string while the fragmentation pattern steers each denoising step. Because the condition is applied throughout the reverse trajectory rather than only at the start, the emerging structure stays tied to the observed peaks instead of drifting toward whatever the decoder finds fluent.

The model was developed by researchers at the Zhengzhou Tobacco Research Institute of the China National Tobacco Corporation with Jinling Institute of Technology, Monash University and China Pharmaceutical University, and published in Analytical Chemistry in November 2025. It sits alongside the catalog's other mass-spectrometry models at a different point in the pipeline: FIDDLE assigns a molecular formula to a spectrum, AIMe predicts spectra and their fragmentation pathways in the forward direction, and SpecMol reads NMR, IR and MS as text. DiffNovo shares its name with an unrelated transformer-diffusion model for de novo peptide sequencing; the two have no authors, code or benchmarks in common.

#Key Features

• Spectrum-conditioned reverse diffusion: Fragmentation features guide every denoising step, so the generated structure is constrained by the spectrum rather than merely initialized from it.
• BART SMILES encoder-decoder: A sequence-to-sequence backbone over SMILES tokens supplies the continuous embedding space the diffusion operates in and captures long-range dependencies such as ring closures and branch nesting.
• High-resolution spectrum encoder: Peaks are encoded at fine mass granularity rather than coarsely binned, preserving the small mass differences that distinguish one neutral loss from another.
• Generation over retrieval: Output is a synthesized SMILES string, so the model can propose structures absent from any reference library.
• Code released under MIT: Training, sampling and evaluation scripts are public, though the repository is not runnable as published — several modules the scripts import, including the SMILES tokenizer and the dataloader layer, are absent from it.

#Technical Details

The system has three parts. A BART encoder-decoder over SMILES runs at a model width of 768 with six encoder and three decoder layers, feed-forward width 3,072, and a regex-based SMILES tokenizer whose vocabulary is derived from the training corpus; sequences are handled up to 128 tokens. The diffusion denoiser is a 12-layer transformer with hidden size 1,024, 16 attention heads and a 512-dimensional latent channel, trained with 6,000 timesteps on a square-root noise schedule, predicting the clean signal directly under an end-to-end mean squared error objective. The implementation builds on the Diffusion-LM, MolT5 and TGM-DLM codebases.

Training and evaluation are both on NIST20. The authors report gains over Spec2Mol, MS2Mol and Spec2Smiles on SMILES-level accuracy, fingerprint-based similarity and alignment of physicochemical properties. No pretrained checkpoint accompanies the code, and the repository is not a runnable end-to-end pipeline: NIST20 is a commercial library, and the model, training and evaluation scripts import modules the repository does not contain — the SMILES tokenizers (mytokenizers, mytokenizers_BPE) and the whole dataloader layer (mydatasetsCsv, improved_diffusion.text_datasets and improved_diffusion.image_datasets).

#Applications

The target users are analytical chemists and metabolomics groups facing spectra that return no confident library hit — natural product discovery, drug and xenobiotic metabolite identification, plant and food chemistry, and exposure analysis. In practice a generative model of this kind produces a shortlist of candidate structures for a human to triage against retention time, isotope pattern and orthogonal spectra, rather than a single answer to be trusted outright.

#Impact

DiffNovo is part of a shift in mass spectrometry from scoring a spectrum against a library to generating the structure that would explain it, a direction also pursued by graph-diffusion and multimodal approaches. Its specific contribution is showing that continuous diffusion in a SMILES embedding space, rather than discrete diffusion over molecular graphs, is a workable route to the same goal. The evidence base is narrow: results are confined to NIST20, with no evaluation on an external spectral library and no released weights, so how well the approach transfers across instruments and compound classes is not yet established by published benchmarks.