ConfSeq

Chemical language models built on SMILES and SELFIES have become the default tool for two-dimensional molecular tasks, but three-dimensional molecular modeling has remained the territory of graph-based diffusion and equivariant networks. The obstacle is representational: there has been no compact one-dimensional string that captures a molecule's 3D geometry while respecting SE(3) invariance, so sequence models had nothing to read or write.

ConfSeq, developed at the Shanghai Institute of Materia Medica of the Chinese Academy of Sciences and published in Nature Machine Intelligence in 2026 after a May 2025 preprint, supplies that missing string. It extends SMILES with discretized internal coordinates — torsion angles, bond angles, and a newly defined pseudo-chirality symbol — inserted as angle-bracketed tokens (for example <113>) at the positions of the corresponding atoms and bonds. Because internal coordinates are invariant to rotation and translation, the encoding is SE(3)-invariant by construction, and because the angular values fall in a bounded range they discretize cleanly into a finite vocabulary. The result stays as readable and as short as SMILES, and it supports enumeration-style data augmentation.

With a molecule's geometry expressed as a token string, conformer prediction, 3D molecular generation, and 3D representation learning all become ordinary sequence-modeling problems solvable with off-the-shelf Transformers. That reframing is the paper's central claim, and it is what separates ConfSeq from approaches that bolt spatial-coordinate modules onto a language model.

#Key Features

• SE(3)-invariant string encoding: Geometry is carried by internal coordinates rather than Cartesian coordinates, so the representation is unchanged by rigid-body motion and avoids asking a language model to reason over raw numerical positions.
• One language, three task families: The same description language underpins an encoder-decoder conformer predictor, a decoder-only generator, and an encoder-only representation model, with no task-specific geometric machinery.
• Confidence scoring for free: Autoregressive token probabilities yield a per-conformer score that correlates with DFT conformational energy at an average Pearson coefficient of -0.58, letting the model rank its own outputs.
• Sampling control by temperature: Raising the sampling temperature trades precision for recall along a curve that dominates competing methods, so users tune diversity to the application rather than retraining.
• Order-of-magnitude faster inference: Generation runs at roughly 0.014 seconds per molecule, about 500 times faster than the diffusion baselines it is compared against.

#Technical Details

All ConfSeq models are standard Transformers. Conformer prediction uses a 6-layer encoder / 6-layer decoder with model dimension 256, trained on a GEOM-Drugs split of 39,990 training and 5,000 validation molecules with 100-fold enumeration augmentation. Unconditional generation uses a 6-layer decoder-only model with hidden dimension 768 trained on the 430,000-molecule GEOM-Drugs subset. Shape-conditioned generation feeds rotation-invariant surface descriptors from 1,024 sampled surface points through RISurConv into the same decoder, trained on 1.59 million MOSES molecules. Representation learning uses a 6-layer encoder-only model with hidden dimension 256 in a Siamese setup, trained on 95.7 million molecule pairs drawn from 2.77 million ChEMBL and BindingDB compounds.

On GEOM-Drugs conformer prediction, ConfSeq raises COV-P from 47.9% to 58.4% and lowers MAT-P from 0.86 Å to 0.77 Å against torsional diffusion. Unconditional generation reaches 99.3% validity and 82.3% PoseBusters validity, with dihedral angle MMD cut from 0.0062 to 0.0004. Shape-conditioned generation achieves 93.1% PoseBusters validity versus 85.1% for ShapeMol and 70.1% for SQUID. In ligand-based virtual screening the learned embeddings reach AUC 0.76 on DUD-E and 0.60 on PCBA, beating E3FP, MORSE, LSalign, and SHAFTS.

#Applications

ConfSeq targets small-molecule drug discovery. Conformer generation feeds docking and shape-matching pipelines; shape-conditioned generation supports scaffold hopping around a known active while diverging in 2D topology, a common route around patent barriers; and the metric-learned embeddings turn 3D similarity search into a Euclidean nearest-neighbour lookup. The authors precomputed embeddings for all 98 million 3D compounds in PubChem, so a query can be scored against the full set in about a minute on a single CPU core. Pretrained models are served through the Sciminer web platform, and the reference implementation is on GitHub under AGPL-3.0.

#Impact

ConfSeq's contribution is a representation rather than a single checkpoint, and its significance lies in showing that the tooling, scaling behaviour, and controllability of language models transfer to 3D chemistry once the geometry is tokenized properly. It sits alongside other language-model approaches to 3D molecular design such as UniLingo3DMol, but differs in encoding conformation intrinsically rather than predicting coordinates with auxiliary modules. The practical limits are real: the models are trained on drug-like organic molecules from GEOM, MOSES, ChEMBL, and BindingDB, evaluation is entirely computational, and the released code carries a copyleft licence that constrains commercial reuse.