Loading…
Chemical language model that translates IR, UV-Vis and 1H NMR spectra into SMILES structures, replacing the enumerate-and-filter CASE workflow.
No providers recorded yet. Browse all providers
Computer-aided structure elucidation (CASE) systems have automated the chemist's reasoning from spectra to structure for half a century, and they do it by enumeration — derive constraints from the spectra, generate every isomer consistent with them, then filter the survivors by how well their predicted spectra match. The generation step is where the approach strains. Chemical space grows exponentially with molecular size — the GDB-17 database enumerates 166 billion molecules of up to 17 heavy atoms — so a complex compound can force a CASE system to produce and score hundreds of thousands of candidates, taking minutes to hours.
CLAMS (Chemical LAnguage Model for Structural elucidation) removes the enumeration step entirely. It treats elucidation as sequence-to-sequence translation: spectra go in, and a SMILES string comes out one token at a time, the way a language model writes text. Nothing is enumerated or filtered, so elucidation cost scales linearly with molecular size rather than combinatorially. The model was built by Xiaofeng Tan at X Scientific, posted as a preprint in October 2024 and published in the Journal of Cheminformatics in July 2025.
The design decision that carries the architecture is treating three unrelated 1D measurements as a single image. Concatenating IR, UV-Vis and 1H NMR traces into one array and reshaping it into a square lets a vision transformer's self-attention weigh evidence across modalities and across widely separated regions of one spectrum — the OH group absorbs in three distinct IR bands that a local filter would never connect.
The encoder is a 9-layer, 9-head vision transformer with hidden size 288, chosen by grid search over the functional-group classification accuracy; the full encoder-decoder carries 104,038,624 trainable parameters excluding the pretraining classifier. Training data comes from QM9S, a DFT-augmented version of QM9 holding 129,817 molecules of C, F, N, O and H with at most 9 heavy atoms and 29 total atoms. IR, UV-Vis and 1H NMR spectra were computed from its B3LYP/def-TZVP properties, leaving 127,465 usable molecules split 80/10/10. On the random split, structure elucidation accuracy is 45.2% top-1, 74.3% top-5 and 83.1% top-15 using all three modalities, against 35.6%, 64.1% and 75.5% for IR alone; the valid-SMILES ratio falls from 0.992 at k=1 to 0.920 at k=15. Functional-group F1 spans 0.92 to 1.0 across the 18 groups, and IR alone is enough for that sub-task — UV-Vis and NMR add nothing to group identification but contribute roughly nine points of elucidation accuracy, because they constrain how the groups are arranged.
Structure elucidation is routine work in pharmaceutical characterization, natural-product isolation and materials chemistry, and CLAMS is aimed at the point in that workflow where a chemist wants a short ranked list of plausible structures rather than an exhaustive candidate set. The model is a proof of concept, and the paper is explicit about what that means: spectra are DFT-simulated rather than measured, mass spectrometry and 13C NMR are absent because QM9S carries no properties from which to compute them, and molecules are capped at nine heavy atoms. Deployment would require fine-tuning on experimental spectra to absorb the discrepancy between calculated and measured data.
CLAMS is one of several recent attempts to replace CASE logic with learned sequence models, alongside the Alberts et al. transformers it was developed independently of and the catalog's SpecMol, SpectraLLM and UltraIR. Its most transferable result is the comparison the author draws against those transformers: they reach 45% top-1 on ~539k IR spectra where CLAMS reaches 33% on 102k, a gap the paper attributes to training-set size rather than architecture, which frames simulated spectral corpora as the scaling lever for this task. The codebase is released under CC BY-NC 4.0 as training and benchmarking notebooks; the trained CLAMS checkpoint itself has not been published, so reproducing the model means retraining it from the released data.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.