ChARGe: A Chemistry Tool-Augmented Reasoning Framework to Accelerate AI-Driven Molecular Design and Reaction Prediction

ChARGe (Chemistry Augmented Reasoning for Generating molecules and Reactions) is an open-source framework co-developed by Lawrence Livermore National Laboratory (LLNL) and Binghamton University. It adopts a tool-augmented reasoning paradigm, integrating large language models (LLMs) with professional chemical computing tools to enable augmented reasoning for molecular generation and reaction prediction. It supports iterative optimization and validation, addressing challenges faced by pure LLMs in chemistry—including the need for specialized chemical knowledge, molecular validity, and synthetic feasibility—and provides an interpretable AI-assisted tool for fields like drug discovery.

Artificial intelligence is developing rapidly in the field of chemistry (especially molecular generation and reaction prediction), but pure language model-based methods face key challenges: requirements for specialized chemical knowledge, constraints on molecular structure validity, and practical considerations of synthetic feasibility. Traditional molecular generation methods often produce a large number of candidates but lack real-time verification of chemical validity (e.g., SMILES with correct syntax but non-existent chemically, or high synthetic difficulty); scenarios like drug discovery require simultaneous optimization of mutually constrained properties such as activity, toxicity, and synthetic difficulty.

Core Design Philosophy

ChARGe adopts the 'tool-augmented reasoning' paradigm: LLMs handle high-level reasoning and hypothesis generation, while professional chemical tools are responsible for validation and computation, balancing the generative capabilities of LLMs with the professionalism and accuracy of chemical computing.

Core Architecture: Hypothesis-Validation-Optimization Cycle

1. Hypothesis Generation: LLMs generate candidate molecules/reaction schemes based on prompts;
2. Validation: Validate whether candidates meet constraints via built-in tools (SMILES validity check, Synthetic Accessibility Score (SAScore), molecular density calculation, etc.);
3. Optimization: Candidates that fail validation enter iterative optimization and are improved based on user feedback;
4. Task Abstraction: Provide a unified interface through the Task base class to support extension to specific chemical scenarios.

Technical Details

• SMILES Validation: The verifySMILES function filters invalid structures;
• SAScore: Evaluates synthetic difficulty (1-10, lower score means easier synthesis);
• Multi-Objective Optimization: Combines multiple constraints (e.g., valid SMILES, density ≥0.8, SAScore ≤1.2);
• Iterative Interface: The refine method supports continuous optimization based on user feedback.

In the lead compound optimization task:

• System Role: "You are a helpful chemistry assistant";
• User Goal: "Generate a drug-like molecule";
• Validation Constraints: Valid SMILES, density ≥0.8, SAScore ≤1.2.

Operation Flow:

1. LLM generates initial candidate SMILES;
2. Validate SMILES validity;
3. Calculate density and SAScore;
4. Check if all constraints are met;
5. Return if satisfied, otherwise enter the refine cycle.

This process ensures that the generated molecules are chemically feasible and valuable.

ChARGe provides a scalable and verifiable engineering foundation for chemical AI:

• Interpretability: Validation steps are clear, and failure reasons are traceable;
• Expert Collaboration: Chemical experts can focus on defining validation logic without deep diving into LLM mechanisms;
• Iterative Optimization: Supports human-machine collaborative progressive optimization, aligning with the actual drug discovery process;
• Multi-Scenario Extension: Can be extended to scenarios like reaction prediction and material design by inheriting the Task base class.

Limitations

• The validation toolset is relatively basic (only a few indicators like SAScore and density);
• Currently mainly supports the Gemini model; extending to other models requires additional development;
• Lacks consideration of 3D molecular conformations and molecular dynamics properties.

Future Directions

• Integrate more diverse chemical tools (e.g., docking scores, ADMET prediction);
• Support multi-modal inputs (e.g., protein structure images);
• Implement distributed parallel optimization to accelerate large-scale screening;
• Integrate with experimental automation platforms.