AI ‘molecular Editor’ Reshapes Molecules with Human-Level Precision and Control

Researchers are tackling the challenge of efficient and precise molecular manipulation with a new artificially intelligent system called El Agente Estructural. Changhyeok Choi, Yunheng Zou, and Marcel Müller, from the University of Toronto’s Departments of Chemistry and Computer Science, alongside Han Hao from the Acceleration Consortium, and Yeonghun Kang and Juan B. Pérez-Sánchez from the University of Toronto’s Department of Chemistry, have developed this multimodal agent to autonomously generate and modify molecular structures. Unlike existing generative models, El Agente Estructural operates by directly mimicking the actions of expert chemists, offering precise control over atomic arrangements and stereochemistry without requiring complete molecular reconstruction. This capability represents a significant advance, enabling chemically relevant geometry manipulation for diverse applications such as site-selective functionalisation, ligand binding, and mechanism-driven structure modification, and ultimately enhancing the capabilities of autonomous research platforms like El Agente Quntur.

Unlike existing methods reliant on databases or generative models, this agent directly mimics the precise, three-dimensional manipulation techniques employed by expert chemists.

This innovative approach bypasses the need to rebuild entire molecular frameworks, enabling targeted modifications to atomic arrangements, connectivity, and stereochemistry with unprecedented control. Unlike generative models, this approach enables precise control over atomic or functional group replacements, atomic connectivity, and stereochemistry without requiring complete molecular rebuilding.
The system operates directly on atomic indices within xyz representations, facilitating site-specific manipulation at the level of individual atoms or functional groups. Advanced operations include site-selective functionalization, symmetry-aware branch replacement, and the generation of organometallic structures and transition state geometries.

The research detailed the implementation of stereochemical transformations, such as enantiomer and cis/trans isomer interconversion, alongside fragment-level structural analysis. The system utilises natural-language processing alongside a suite of domain-informed tools and vision-language models to achieve precise control over molecular structures.
Structural analysis tools identify atomic indices and structural features within molecules, enabling site-selective manipulation and informed editing of molecular geometries. The view_xyz and zoom_xyz tools render 3D molecular views with labelled atomic indices using PyMol, allowing for visual inspection and verification of geometry.

The python_repl tool facilitates examination of geometry through five numpy operations, enabling identification of minimum distances between atoms and precise angle verification. The get_distance_angle_dihedral tool calculates interatomic distances, angles, and dihedral angles, utilising conventions applicable to single atoms or molecular fragments via centroid dummy atoms.

Atomic_neighbor_identification identifies directly connected atoms given an element or index, efficiently characterizing local coordination environments. Match_smarts_in_xyz converts xyz representations into MOL objects, identifying atomic indices corresponding to target fragments specified by SMILES or SMARTS patterns.

The get_connected_subgraph_indices tool identifies complete fragment indices using a starting and excluded atomic index, crucial for branch-replacement and geometric operations. Find_pointgroup_equivalent_atoms determines molecular point group symmetry and identifies symmetry-equivalent atoms, supporting automated selection for symmetric functionalization.

Geometric operation tools directly manipulate molecular geometry by modifying interatomic distances, bond angles, and dihedral angles, applicable to both individual atoms and connected subgroups. Fragment rotations are achieved via a tool that computes unit vectors and rotates fragments by a specified angle, facilitating isomeric conversions. Unlike methods relying on complete molecular reconstruction, this agent directly manipulates three-dimensional structures, mirroring the techniques employed by expert computational chemists.

This approach allows for precise control over atomic substitutions, connectivity, and stereochemistry while preserving the core molecular framework. Limitations acknowledged by the developers include the current focus on Cartesian coordinates and the potential for improved user interfaces through virtual or augmented reality integration. Future work will concentrate on expanding the agent’s capabilities to facilitate more autonomous exploration of chemical space, encompassing molecular, organometallic, and catalytic systems, ultimately aiming to accelerate discovery and improve understanding across diverse chemical systems.