A research team has introduced I2M, a box-based detection framework engineered for the accurate recognition and reconstruction of chemical structures from molecular images. Published in the Journal of Cheminformatics, the system establishes a detection-based paradigm that directly identifies atoms, bonds, and charges through spatial localization. By predicting both the locations and categories of these constituent chemical elements, the framework preserves two-dimensional molecular layouts while converting visual depictions of molecules into structured data.
Accurate molecular image recognition represents a critical capability across computational chemistry, supporting core applications in drug discovery, materials science, and chemical patent analysis. However, substantial volumes of chemical matter currently exist only in implicit visual forms that are not directly machine-readable. This lack of digital representation creates a major bottleneck, limiting the systematic curation and integration of valuable chemical data into structured databases.
Conventional methodologies have faced persistent obstacles when attempting to parse molecular images automatically. Traditional rule-based techniques and existing machine learning-based approaches often perform poorly when processing complex graphics, particularly images exhibiting overlapping atoms, ambiguous chemical bonds, or degraded visual quality.
To resolve these technical difficulties, the researchers designed I2M using Vision Transformer (ViT) encoders enhanced with adapter modules, combined with pre-trained Detection Transformer (DETR) weights for improved accuracy and efficiency. The architecture incorporates robust query selection mechanisms adapted from DETR to heighten decoding accuracy across visual inputs. After detecting individual elements, the framework applies advanced validation routines to handle intricate cases and assemble the detected atoms, charges, and bonds into chemically interpretable molecular graphs.
The authors evaluated I2M against several prominent molecular recognition systems, including MolVec, DECIMER-Image-to-SMILES, DECIMER Image Transformer, MolGrapher, and MolScribe. Across diverse chemical image benchmarks, I2M demonstrated favorable performance in both recognition accuracy and computational efficiency.
To support ongoing research and standardize future evaluations in the field, the team also developed and released a chemical structure image benchmark along with accompanying datasets. By making these benchmark datasets publicly available, the authors aim to facilitate convenient, automatic testing and reproducible evaluation of emerging molecular image recognition models, helping bridge the gap between chemical imagery and digital molecular representations.
Rather than relying solely on sequential string generation, the framework constructs fully interpretable molecular graphs directly from spatially resolved parts.
Subject of Research: Molecular structure recognition from images using detection transformers
Article Title: I2M: a box-based detection framework for molecular structure recognition from images
Article References: Tang, B., Peng, J., Wu, T., Wu, W., Lv, W., Ge, R., Ng, H. L., Hu, H., Zhou, S., & Qian, Y. (2026). I2M: a box-based detection framework for molecular structure recognition from images. Journal of Cheminformatics. https://doi.org/10.1186/s13321-026-01290-7
Image Credits: AI Generated
DOI: 10.1186/s13321-026-01290-7
Keywords: cheminformatics, molecular structure recognition, molecular images, I2M, Detection Transformer, Vision Transformer, DETR, molecular graphs, drug discovery, materials science, patent analysis, chemical structure benchmark, adapter modules, atom detection, bond recognition, MolScribe, DECIMER, MolGrapher
Cite Scienmag News
Jason Bradley. (November 20, 2026). Box-Based Detection Framework Improves Molecular Structure Recognition from Images. Scienmag. https://scienmag.com/box-based-detection-framework-improves-molecular-structure-recognition-from-images/
Jason Bradley. "Box-Based Detection Framework Improves Molecular Structure Recognition from Images." Scienmag, 20 November 2026, https://scienmag.com/box-based-detection-framework-improves-molecular-structure-recognition-from-images/. Accessed 20 November 2026.
Jason Bradley. "Box-Based Detection Framework Improves Molecular Structure Recognition from Images." Scienmag. November 20, 2026. https://scienmag.com/box-based-detection-framework-improves-molecular-structure-recognition-from-images/








