El Agente Estructural: An Artificially Intelligent Molecular Editor

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.

💡 Research Summary

**
The paper introduces El Agente Estructural, a multimodal AI agent that enables precise, interactive generation and manipulation of three‑dimensional molecular structures through natural‑language dialogue and image inputs. Unlike conventional SMILES‑to‑3D converters, database retrieval, or deep‑generative models that produce whole‑molecule conformations, Estructural operates at the level of individual atomic indices, mirroring how a human chemist clicks, drags, and rotates atoms in a molecular viewer.

System Architecture
The core of the system is a vision‑language model (VLM) that acts as a high‑level planner. It parses user requests, extracts any visual cues from reaction schematics or molecular snapshots, and decomposes the request into a sequence of geometry‑aware actions. These actions are executed by a Python‑based tool engine that hosts a suite of domain‑specific utilities built on open‑source libraries such as ASE, RDKit, OpenBabel, spglib, and pymatgen. Constrained geometry optimizations are performed with the semi‑empirical xtb package to keep the edited structures physically realistic.

Atomic‑Index‑Centric Operations
Every manipulation is expressed as a deterministic command that references atoms by their unique indices in an xyz file. From these indices the agent constructs geometric primitives (vectors, planes, centroids) and computes distances, bond angles, and dihedral angles. By fixing selected geometric parameters during xtb optimization, the agent can enforce exact spatial constraints—e.g., “set the distance between atom 5 and atom 12 to 1.45 Å while keeping the dihedral at 120°.” This discrete command set bridges the gap between the coarse, low‑frequency reasoning of large language models and the fine‑grained, high‑frequency adjustments a human would perform manually.

Tool Categories
The toolkit is divided into four functional groups:

1. Structural analysis – identifies atomic indices, neighboring atoms, and global symmetry.
2. Geometric operations – measures and constrains distances, angles, and dihedrals.
3. Structure editing – substitutes functional groups, binds fragments, and modifies connectivity.
4. Structure generation – creates new ligands or fragments and positions them relative to a host structure.

All tools are invoked programmatically by the agent, allowing composable workflows that range from simple bond length adjustments to complex stereochemical constructions.

Case Studies
Ten representative scenarios demonstrate the breadth of the platform:

Site‑selective functionalization – the core scaffold remains untouched while a methyl or halogen group is inserted at a user‑specified carbon atom.
Ligand binding and exchange – new ligands are attached to transition‑metal centers with user‑defined coordination angles; the system automatically optimizes the metal‑ligand geometry.
Stereochemical control – Δ/Λ configurations of octahedral complexes, cis/trans isomerism, and enantiomeric inversions are constructed by fixing relevant dihedrals.
Fragment swapping for isomer interconversion – a specific bond is broken, a new fragment is grafted, and the resulting geometry is relaxed under constraints.
Image‑guided generation – a hand‑drawn reaction mechanism diagram is fed to the VLM; the agent extracts intermediate structures and generates plausible 3D geometries for transition‑state searches.

Quantitative validation shows sub‑angstrom distance errors (<0.05 Å) and angular deviations (<2°) compared to manually built reference structures, confirming that the agent’s outputs are chemically accurate.

Limitations and Future Directions
Current capabilities focus on small‑molecule and organometallic systems; extending to bulk solids, surfaces, or large nanostructures will require additional spatial reasoning and memory management. The VLM’s image interpretation, while effective for simple schematics, can misinterpret densely annotated reaction diagrams, suggesting a need for specialized vision models or iterative correction loops. Dependency on external libraries also raises maintenance concerns; the authors propose an automated tool‑generation and verification framework to mitigate this.

Future work includes (i) integrating multiple agents for cooperative solid‑state modeling, (ii) incorporating reinforcement‑learning loops for fine‑tuning geometry under physical constraints, (iii) building a self‑checking pipeline that flags inconsistent commands before execution, and (iv) scaling the system to handle high‑throughput autonomous quantum‑chemistry campaigns.

Integration with El Agente Quntur
Estructural is designed to plug into the previously released autonomous quantum‑chemistry platform El Agente Quntur. After a structure is generated or edited, the combined system can automatically set up DFT or semi‑empirical calculations, perform error recovery, and interpret results—all driven by a single natural‑language instruction such as “attach a thiophene ligand to the palladium center, optimize the geometry, and compute the HOMO‑LUMO gap.” This tight coupling eliminates the manual bottleneck that traditionally separates structure preparation from quantum‑chemical analysis.

Conclusion
El Agente Estructural represents a paradigm shift from coarse, generative‑model‑based molecular design toward fine‑grained, explainable, and controllable geometry manipulation. By marrying atomic‑index‑centric operations with multimodal language‑vision reasoning, the platform delivers human‑level precision in a fully autonomous workflow. Its ability to handle stereochemistry, ligand exchange, fragment‑level editing, and image‑guided generation opens new avenues for catalyst design, mechanistic exploration, and custom drug‑molecule tailoring. With planned extensions to solid‑state systems and more robust vision models, Estructural is poised to become a cornerstone of end‑to‑end autonomous chemistry.