SpecMol

Spectroscopy is how chemists actually establish what a molecule is: NMR, infrared absorption, and mass spectrometry are the routine evidence behind structure assignment in a synthesis lab, a natural-product isolation, or a metabolite identification. Molecular language models have largely worked on SMILES strings and molecular graphs instead, partly because there is no agreed way to feed a spectrum to a text model, and partly because published spectral results use incompatible tolerances and binning conventions that make cross-study comparison impossible.

SpecMol addresses both problems at once. It is a 7-billion-parameter molecular foundation model that represents spectra, SMILES strings, IUPAC names, and 3D atomic coordinates in one serialized text format, so a single decoder-only model handles spectra-to-structure elucidation, structure-to-spectra simulation, name conversion, molecular question answering, and SMILES-to-3D generation without modality-specific encoders. Alongside it the authors release SpecMol-Bench, an evaluation protocol that fixes splits, tolerances, and metrics across those tasks. The work comes from Shanghai AI Laboratory with collaborators at Carnegie Mellon University, Brown University, and several other institutions, posted to arXiv in September 2025 and revised through March 2026. The code repository distributes the project under the name MolSpectLLM.

SpecMol sits between spectra-to-structure elucidation systems such as SpectraLLM and NMR-specific representation models such as UltraNMR, differing in that it both reads and writes spectra and emits 3D coordinates directly.

#Key Features

• Tagged textual serialization of spectra: Each modality is wrapped in XML-like tags carrying nucleus, frequency, solvent, multiplicity, coupling constants, and integration, with numeric values in a fixed-point format that preserves magnitude for a tokenizer.
• Bidirectional spectra–structure reasoning: One model elucidates SMILES from a set of spectra and simulates 13C NMR, 1H NMR, IR, and MS/MS spectra from a SMILES string, rather than requiring a separate model per direction.
• 3D coordinates as a generation target: Conformers are serialized as fixed-point coordinate sequences aligned to canonical SMILES atom order, producing parsable SDF geometries directly from a SMILES string or an IUPAC name.
• Multi-modal evidence fusion: Combining modalities sharply reduces ambiguity — Tanimoto similarity to the true structure rises from 0.329 (13C NMR alone) and 0.383 (1H NMR alone) to 0.615 and 0.644 once IR and MS are supplied.
• SpecMol-Bench evaluation protocol: A chemistry-aware metric suite with standardized spectral preprocessing, peak-matching tolerances, and geometry-aware validity checks, covering 1,834 test instances under a Bemis–Murcko scaffold split.

#Technical Details

SpecMol is initialized from the Qwen2.5-7B base model and trained in three phases: continual pre-training on roughly 10 million chemistry documents plus a multimodal corpus, multi-task supervised fine-tuning, and a final LoRA instruction-tuning pass that restores strict output formatting. The corpus draws molecular descriptions from PubChem (11.9M molecules), simulated IR/MS/NMR spectra from QM9S (~130,000), experimental NMR from NMRBank (225,809), and 790,000 records from the Multimodal Spectroscopic Dataset; after deduplication and removal of any overlap with the evaluation sets, about 5M molecules and 0.2M spectrum–molecule pairs remain.

On spectra-to-SMILES elucidation, SpecMol reaches 27.0% exact sequence accuracy and 55.55 RDKit fingerprint similarity, against 2.01% for the best generalist baseline (Gemini-3-flash) and 0.00% for the chemistry specialists ChemDFM-v2.0-14B and Intern-S1. On the forward problem it raises 13C NMR peak F1 from 0.254 to 0.479 (MAE 0.149 ppm), attains 0.449 weighted Jaccard on 1H NMR, and reaches cosine similarities of 0.554 (IR) and 0.423 (MS); under NMR-Solver-style set scores it exceeds the commercial predictor MestReNova on 1H NMR (0.827 vs. 0.741). SMILES-to-3D validity is 89.68% versus 69.5% for GPT-5, and molecular QA accuracy is 68.6% versus 32.0% for the untuned Qwen2.5-7B base.

#Applications

The immediate audience is analytical and synthetic chemistry: proposing candidate structures from a freshly acquired NMR/IR/MS panel, sanity-checking a synthesized compound by simulating the spectra its assigned structure should produce, converting between IUPAC names and SMILES, and generating a starting conformer for docking or property calculation. Because the model consumes and emits plain text, it slots into scripted pipelines alongside existing cheminformatics tooling, and its scaffold-split evaluation makes it a reference point for groups building spectroscopy-aware assistants for natural product dereplication or metabolite identification.

#Impact

The contribution is as much methodological as it is a model: by pinning down a serialization format and a metric suite, SpecMol-Bench gives the spectroscopy-plus-LLM literature a common yardstick it previously lacked, and the controlled comparison shows how far frontier general-purpose models still are from laboratory-grade structure elucidation. The caveats are real. The work is an arXiv preprint that has not been peer reviewed; the Apache-2.0 repository ships code, metrics, and partial data, but no model checkpoint has been released, so the reported numbers cannot yet be reproduced end to end. The authors also report that the model trails proprietary frontier systems on general chemical question answering, misses low-intensity diagnostic peaks in simulated IR and MS spectra, and can be beaten by mass-spectrometry-specific elucidation tools on chemotypes where explicit fragmentation rules matter.