Abstract
Large language models succeed by combining large-scale pretraining with meaningful discrete tokens. In molecular machine learning, SMILES is widely used as a token representation, but it is primarily a linearization format for molecular graphs rather than a semantic decomposition of chemistry. We propose VQ-Atom, a semantic tokenization framework that assigns discrete atom-level tokens based on local chemical environments via vector quantization. Unlike SMILES tokens, VQ-Atom tokens encode graph-local chemical context and are aligned with molecular structure. On protein-cold drug--target interaction prediction using the KIBA dataset, VQ-Atom substantially improves global ranking performance, achieving AUROC of 0.79 while substantially outperforming both SMILES-based and continuous molecular representations under an identical downstream architecture. Furthermore, VQ-Atom enables approximately 3 times faster downstream training than continuous atom-level representations by replacing per-atom continuous features with reusable discrete tokens. These results suggest that molecular tokenization is not merely a preprocessing step, but a central design choice. In particular, well-structured tokens can encode substantial chemical semantics, reducing the burden on downstream learning. VQ-Atom can be interpreted as defining a molecular language, where tokens correspond to chemically meaningful atomic environments, suggesting that token design may constitute an additional axis of machine learning research alongside architecture, objectives, and optimization.
Explore similar work
Jun 22, 2026cs.LG
Chemical language models (cLMs) are widely assumed to learn surface-level syntactic patterns rather than learning meaningful molecular semantics. Here, we apply sparse autoencoders (SAEs) to MolFormer, an encoder-only cLM, to mechanistically examine how molecular representations are built across layers. We discover that early layers rely on position-tracking latents to parse molecular grammar, while later layers encode atom-in-substructure and pharmacologically relevant features. Additionally, we show that non-canonical SMILES produce more disruptive representation shifts than invalid SMILES, driven by position-latent disruption propagating across layers. To support further exploration, we develop InterMol, an interactive visualizer for SAE activations on molecular strings and structures.
Christian Kenneth, Etowah Adams, Liam Bai +1
Jun 10, 2026cs.CL
Transformer-based language models for SMILES strings suffer from a locality gap: standard character-level tokenization fragments chemically meaningful motifs, forcing models to repeatedly learn local syntax at the expense of long-range dependencies. To address this without disrupting standard tokenizers, we propose MolGram, which integrates a conditional
n-gram memory module into molecular language models. MolGram maps local string patterns to learned embeddings via scalable hash lookups and dynamically injects this regional context into hidden states. Evaluations across three tasks, including unconditional molecule generation, forward reaction prediction, and single-step retrosynthesis, show that MolGram consistently improves performance. Crucially, our analyses demonstrate that MolGram outperforms baselines with 3
× more parameters, establishing explicit local pattern memory as a highly efficient inductive bias.
Xinni Zhang, Zijing Liu, He Cao +2
Jul 6, 2026cs.CL
Every chemical language model reading SMILES begins with a tokenizer, yet the field has inherited byte-pair encoding (BPE) from natural language with little scrutiny. In natural language, BPE's principal alternative, Unigram-LM, is known to build structurally different vocabularies. Whether that contrast survives in chemistry was open. We report a controlled comparison of BPE and Unigram-LM over a fixed 165-token chemistry base, at the small vocabulary sizes where token embeddings are learnable, across three corpus typologies (diverse, drug-like, natural-products) and both pre-tokenization boundary policies. The two do not converge. In all 22 matched conditions they build near-disjoint subword vocabularies: cross-algorithm Jaccard overlap on the learned pieces never exceeds 0.161, and at most 0.05 once weighted toward the high-frequency pieces a model updates most. Unigram-LM also segments held-out molecules into 29-41% more tokens; the arms largely agree on where to cut but not how deeply, so BPE's segmentation is a strict coarsening of Unigram-LM's on 80-99% of molecules. The separation holds across corpus, boundary, and vocabulary size, persisting even at eight times that scale. The subword algorithm is therefore a modeling decision, not a free default. The study trains no language models.
Hunter Heidenreich