smiles canonicalization

Large language models (LLMs) exhibit low efficacy in molecular generation under few-shot settings, frequently producing syntactically invalid SMILES strings. Method: We propose SmiSelf, a cross-chemical-language framework that leverages formal grammar rules to automatically map invalid SMILES to syntactically valid SELFIES representations while preserving molecular semantics. Contribution/Results: SmiSelf achieves the first LLM-based generation of 100% syntactically valid and chemically reasonable molecules. Experiments demonstrate that it maintains 100% molecular validity while fully preserving key physicochemical properties (e.g., logP, synthetic accessibility, QED) of the original molecules; it matches or exceeds baseline models in diversity, novelty, and drug-likeness. Crucially, SmiSelf integrates seamlessly into existing SMILES-based generative models without requiring retraining. This work establishes a reliable, scalable paradigm for LLM-driven molecular design in drug discovery and materials science.

This work addresses a critical limitation in current molecular large language models—their inadequate adherence to the principle that “structure determines function,” resulting in weak foundational structural understanding. To overcome this, the authors propose MolBasic, a novel framework that introduces a structure-first paradigm centered on bidirectional translation between SMILES strings and molecular graphs. By aligning topological and sequential representations through a multi-level structure-aware benchmark, and integrating progressive learning with standardized chain-of-thought prompting, MolBasic enhances the model’s capacity to evolve from basic structural perception to advanced reasoning. The approach achieves substantial improvements in structural comprehension and consistently delivers robust gains across downstream tasks, including molecular property prediction and target-oriented molecular optimization.

Large language models (LLMs) exhibit weak comprehension of molecular structures encoded in SMILES, struggling even with fundamental tasks such as ring counting—severely limiting their utility in molecular science. To address this, we propose CLEANMOL, the first framework that establishes a hierarchical SMILES parsing task suite explicitly grounded in molecular graph structure. It decomposes molecular understanding into well-defined, graph-level subtasks: subgraph identification, ring detection, atomic connectivity verification, and global graph matching. CLEANMOL innovatively introduces an adaptive difficulty-scoring mechanism to curate a molecule-centric pretraining dataset and incorporates graph-structural priors to design chemistry-aware supervision signals. After fine-tuning on open-source LLMs, CLEANMOL achieves substantial improvements on foundational tasks—including ring counting—and attains state-of-the-art performance on the Mol-Instructions benchmark.

This work addresses the challenge of cross-modal heterogeneity and semantic misalignment between SMILES strings (1D sequences) and molecular graphs (2D structures) in molecular language modeling. To this end, we propose a fine-grained cross-modal alignment framework built upon the T5 architecture. Our method jointly encodes SMILES tokens and graph-structured molecular representations—generated via a GNN—and introduces a learnable cross-token cross-attention mechanism that explicitly models semantic correspondences between SMILES tokens and graph nodes/edges. Unlike conventional approaches relying on simple concatenation or unimodal modeling, our mechanism enables deep integration of structural knowledge with sequence-based language generation. Empirically, the model achieves significant improvements over existing state-of-the-art methods on two key tasks: molecular caption generation and IUPAC name prediction. These results demonstrate that structure-aware multimodal co-modeling effectively enhances both molecular understanding and generative capabilities.

Existing SMILES pretraining models rely solely on single-token supervision, neglecting substructural semantics, and are trained only on corrupted SMILES strings—leading to weak supervisory signals and train-inference mismatch. To address these limitations, we propose SMI-Editor, an edit-based pretraining paradigm that randomly perturbs molecular substructures (rather than individual atoms or bonds) and reconstructs the original valid SMILES, thereby enabling fragment-level supervision and joint modeling of chemical validity. Built upon a Transformer architecture, SMI-Editor explicitly incorporates SMILES syntactic constraints and chemical substructure priors. This work is the first to introduce edit operations into molecular language modeling. Evaluated across multiple downstream tasks, SMI-Editor achieves state-of-the-art performance—outperforming several 3D-aware representation models—and significantly enhances molecular semantic understanding and generation capabilities.

This work addresses the limitation of small language models (SLMs) in perceiving critical graph topological structures when predicting molecular properties from SMILES strings. To overcome this, the authors propose a context-augmented prompting framework that dynamically integrates, during inference, prediction prompts generated by graph neural networks (GNNs) with interpretable subgraphs, thereby enabling structure-aware zero-shot molecular property prediction for the first time. The approach synergistically combines GNNs, subgraph extraction, confidence estimation, and edge-ablation intervention analysis. Evaluated on the MUTAG and Tox21 datasets, the method achieves up to a 74% relative improvement in accuracy, demonstrating conclusively that incorporating graph-based contextual information significantly enhances the molecular understanding capabilities of small language models.

Accurately converting two-dimensional molecular structure diagrams from printed documents into SMILES strings remains a challenging task, with existing vision-language models exhibiting suboptimal performance. This work proposes MolSeek-OCR, which formulates the problem as an image-conditioned sequence generation task. Built upon the DeepSeek-OCR-2 architecture, the model employs a two-stage progressive fine-tuning strategy: first applying parameter-efficient fine-tuning via LoRA, followed by selective full-parameter fine-tuning with a hierarchical learning rate schedule. Training leverages a hybrid dataset combining synthetic and real patent images, effectively mitigating training instability and substantially enhancing generalization. The proposed approach achieves state-of-the-art performance among image-to-sequence models in exact-match accuracy for molecular structure recognition.

This study addresses the limitations of conventional molecular representations—such as SMILES and IUPAC—in large language models (LLMs), which suffer from parsing difficulties, low generation accuracy, and poor robustness with complex structures. The work presents the first systematic evaluation of how different molecular representations affect LLM performance and introduces MolJSON, a novel representation grounded in explicit molecular graph structure. Experimental results across 78,045 samples using state-of-the-art models including GPT-5 and Claude Haiku 4.5 demonstrate that MolJSON substantially outperforms traditional formats in translation, constrained generation, and shortest-path reasoning tasks. Specifically, IUPAC-to-MolJSON translation achieves 71.0% accuracy (versus 43.7% for SMILES), constrained generation reaches 95.3% (compared to 64.0% for SMILES), and path reasoning attains 98.5% accuracy—all while requiring fewer inference tokens.

This study addresses the lack of a standardized textual representation for molecules in large language models (LLMs) and the frequent oversight of how representation choice critically impacts model performance. The authors systematically evaluate nine molecular representations—including SMILES, InChI, IUPAC, and CML—across eight chemical tasks using sixteen diverse LLMs, encompassing general-purpose, reasoning-enhanced, and chemistry-specific models. Performance is assessed through generation quality (via LLM-as-a-judge), alongside mechanistic analyses such as tokenization audits, linear probing, and attention mapping. The work reveals, for the first time, a strong dependence of representation efficacy on task type: IUPAC excels in semantic and generative correctness, structured formats are better suited for structural tasks, and CML demonstrates the strongest overall performance. Based on these findings, the authors propose a task-aware representation routing strategy, challenging the prevailing “representation-agnostic” evaluation paradigm and uncovering fundamental differences in how representations are encoded mechanistically.

This study addresses the lack of systematic comparison among subword tokenization algorithms in SMILES-based chemical language modeling, where unvalidated default schemes are commonly adopted. Under a fixed 165-token base vocabulary, controlled experiments across diverse chemical corpora and pre-tokenization strategies compare Byte Pair Encoding (BPE) and Unigram Language Modeling (Unigram-LM). The work reveals, for the first time, that the subword vocabularies generated by these two methods exhibit minimal overlap—Jaccard coefficients ≤0.161 overall and ≤0.05 among high-frequency tokens—and display systematic differences in segmentation granularity: Unigram-LM produces 29–41% more tokens on average, while BPE acts effectively as a coarsened variant of Unigram-LM for 80–99% of molecules. These findings demonstrate that the choice of tokenization algorithm is a critical design decision rather than a negligible default setting in molecular sequence modeling.