[Paper] El Agente Estructural: An Artificially Intelligent Molecular Editor

Source: arXiv - 2602.04849v1

Overview

The paper introduces El Agente Estructural, a multimodal AI assistant that lets users edit and generate 3‑D molecular structures through natural‑language commands. By combining domain‑specific chemistry tools with vision‑language models, the system mimics how a human chemist would “grab” atoms or functional groups and reposition them, opening a new way to interact with molecular modelling software.

Key Contributions

• Natural‑language‑driven molecular editing – Users can specify atomic replacements, stereochemistry changes, or ligand swaps with plain English (or other supported languages).
• Multimodal reasoning – The agent fuses text, 2‑D sketches, and 3‑D visual cues, enabling image‑guided generation from reaction schematics or microscopy snapshots.
• Geometry‑aware toolset – A library of chemistry‑specific operations (bond formation/breakage, conformer optimization, stereocenter enforcement) runs under the hood, guaranteeing chemically valid outcomes.
• Integration with autonomous quantum‑chemistry pipelines – The editor is designed to plug into the larger El Agente Quntur multi‑agent platform for end‑to‑end property prediction and reaction planning.
• Extensive case‑study validation – Demonstrated on tasks such as site‑selective functionalization, ligand exchange, isomer interconversion, and fragment‑level analysis, showcasing real‑world relevance.

Methodology

1. Input Parsing – A large language model (LLM) processes the user’s textual instruction, extracting intent (e.g., “replace the para‑hydrogen with a nitro group”).
2. Vision‑Language Fusion – When a 2‑D sketch or 3‑D snapshot is supplied, a vision transformer aligns visual elements with the parsed intent, locating the target atoms or bonds.
3. Tool Invocation – The system selects from a curated toolbox of geometry‑aware operations (bond edit, conformer generation, stereochemistry enforcement). Each tool is wrapped as a micro‑service with a clear API, allowing seamless orchestration.
4. Constraint Checking – Before committing changes, a rule engine validates chemical feasibility (valence, aromaticity, steric clashes) and, if needed, triggers a short quantum‑chemical relaxation (e.g., semi‑empirical geometry optimization).
5. Feedback Loop – The edited structure is rendered back to the user, who can issue follow‑up commands, enabling an interactive “conversation” with the molecule.

The architecture is deliberately modular: the LLM, vision model, and chemistry tools can be swapped out or upgraded without redesigning the whole system.

Results & Findings

Across all case studies, the system produced chemically valid structures with minimal need for manual post‑processing, demonstrating that multimodal reasoning can replace many repetitive, script‑based editing steps.

Practical Implications

• Accelerated prototyping – Chemists and material scientists can quickly iterate on molecular designs by typing “add a methyl group to the ortho position” instead of scripting geometry edits.
• Lower barrier to entry – Developers building cheminformatics platforms can embed the editor as a plug‑and‑play component, exposing powerful editing capabilities to users without deep domain expertise.
• Enhanced automation pipelines – When coupled with El Agente Quntur, the editor can automatically generate candidate structures for high‑throughput quantum‑chemical screening, closing the loop between hypothesis generation and property evaluation.
• Educational tools – Interactive, language‑driven manipulation can serve as a teaching aid in organic chemistry courses, allowing students to explore stereochemistry and reaction mechanisms in real time.
• Cross‑disciplinary workflows – The multimodal interface makes it easier for data scientists, AI engineers, and chemists to collaborate, as the same natural‑language commands can be understood by both humans and machines.

Limitations & Future Work

• Dependence on LLM quality – Ambiguous or poorly phrased instructions can lead to unintended edits; robust prompt engineering or clarification dialogs are needed.
• Scalability of geometry optimization – The current workflow uses semi‑empirical methods for quick relaxation; integrating faster GPU‑accelerated quantum‑chemical engines could improve throughput for large systems.
• Domain coverage – While the toolbox handles many organic and coordination chemistries, exotic functional groups (e.g., organometallic clusters) are not yet supported.
• User‑feedback integration – Future versions aim to learn from correction loops, allowing the agent to refine its tool‑selection policy based on user acceptance/rejection of edits.

Overall, El Agente Estructural showcases how AI‑driven multimodal interfaces can transform molecular modelling from a code‑heavy activity into a conversational, interactive experience—an advancement that could reshape workflows across drug discovery, materials design, and chemical education.

Authors

• Changhyeok Choi
• Yunheng Zou
• Marcel Müller
• Han Hao
• Yeonghun Kang
• Juan B. Pérez‑Sánchez
• Ignacio Gustin
• Hanyong Xu
• Mohammad Ghazi Vakili
• Chris Crebolder
• Alán Aspuru‑Guzik
• Varinia Bernales

Paper Information

• arXiv ID: 2602.04849v1
• Categories: physics.chem-ph, cs.AI, cs.MA
• Published: February 4, 2026
• PDF: Download PDF