El Agente Estructural: An Artificially Intelligent Molecular Editor

TL;DR

El Agente Estructural introduces a multimodal, natural-language–driven molecular editor that operates directly on 3D coordinates to enable precise geometry modifications while preserving core structures. It achieves this through an atomic-index–centric framework and a modular toolbox (structural analysis, geometric operations, editing, and generation) built on open-source chemistry libraries, with validation across diverse case studies including site-selective functionalization, fragment binding, and stereochemical organometallic construction. The work demonstrates robust integration of vision-language reasoning with domain-specific tools, enabling image-guided and mechanism-driven geometry generation beyond traditional SMILES or database-based approaches. A staged roadmap outlines integration with autonomous quantum chemistry platforms, data-driven retrieval, interactive UIs, and advanced organometallic and solid-state capabilities to accelerate discovery in chemistry and catalysis.

Abstract

We present El Agente Estructural, a multimodal, natural-language-driven geometry-generation and manipulation agent for autonomous chemistry and molecular modelling. Unlike molecular generation or editing via generative models, Estructural mimics how human experts directly manipulate molecular systems in three dimensions by integrating a comprehensive set of domain-informed tools and vision-language models. This design enables precise control over atomic or functional group replacements, atomic connectivity, and stereochemistry without the need to rebuild extensive core molecular frameworks. Through a series of representative case studies, we demonstrate that Estructural enables chemically meaningful geometry manipulation across a wide range of real-world scenarios. These include site-selective functionalization, ligand binding, ligand exchange, stereochemically controlled structure construction, isomer interconversion, fragment-level structural analysis, image-guided generation of structures from schematic reaction mechanisms, and mechanism-driven geometry generation and modification. These examples illustrate how multimodal reasoning, when combined with specialized geometry-aware tools, supports interactive and context-aware molecular modelling beyond structure generation. Looking forward, the integration of Estructural into El Agente Quntur, an autonomous multi-agent quantum chemistry platform, enhances its capabilities by adding sophisticated tools for the generation and editing of three-dimensional structures.

Figures (25)

• Figure 1: Schematic overview of the Estructural architecture and workflow. Estructural supports natural-language prompts, coordinate files, and reaction schematics as inputs, and manipulates molecular systems through four action categories: structural editing, structural generation, structural analysis, and geometric operations. The example illustrates attachment of a thiophene ligand to an organometallic complex: Estructural generates the ligand, identifies binding atoms, performs the coordination step, and validates the resulting structure. Atom color codes: pink, Co; yellow, S; red, O; blue, N; gray, C; white, H.
• Figure 2: Overview of the tools integrated into Estructural. The tools are categorized into structural analysis, geometric operation, structure editing, and structure generation, collectively enabling flexible construction, modification, and analysis of molecular geometries. Atom color codes: seashell, Pt; green, Ru; lime green, Cl; blue, N; red, O; gray, C; white, H.
• Figure 3: Schematic illustration of structure editing tools, including (a) terminal atom replacement, (b) molecule binding, and (c) branch replacement. These tools enable functionalization, fragment attachment, and subgroup substitution by modifying selected atomic sites while preserving the rest of the molecular geometry. Representative arguments are shown in each example, including the input xyz structure filename, target atomic indices for replacement or binding, fragment specifications, and the name of the newly generated output structure. Atom color codes: dark orange, Fe; red, O; blue, N; gray, C; white, H.
• Figure 4: Overview of structure generation tools for organometallic molecules. Organometallic structures are constructed using predefined coordination–symmetry templates, in which ligands are assigned to annotated atomic indices to generate stereochemically specific geometries. The representative input arguments and the corresponding generated structures for (a) square planar cis-[Pt(NH$_3$)$_2$Cl$_2$] and (b) trans-[Pt(NH$_3$)$_2$Cl$_2$], (c) $\Delta$-Ru(bpy)$_3$ (bpy = 2,2$'$-bipyridine), and (d) tetrahedral ZrCp$_2$(CH$_3$)(C$_2$H$_4$) (Cp = $\eta^5$-cyclopentadienyl). Atom color codes: seashell, Pt; green, Ru; sky blue, Zr; lime green, Cl; blue, N; gray, C; white, H.
• Figure 5: VLM Benchmark Performance by Molecule Size. Comparing the success rates of five Vision Language Models (Gemini 3 Pro, Gemini 3 Flash, Sonnet 4.5, Opus 4.5, and GPT 5.2) across varying molecule sizes, measured by the number of atoms. The results illustrate a general trend of decreasing accuracy as molecular complexity increases, with larger systems (60+ atoms) presenting a significant challenge for all tested models.