QuantumChem-200K: A Large-Scale Open-Source Organic Molecule Dataset for Quantum Chemical Property Screening and Language Model Evaluation

QuantumChem-200K is a large-scale open-source dataset containing 200,000 organic molecules, designed specifically for quantum chemical property calculation and language model benchmarking. It fills the gap in public large-scale quantum chemistry data, supports AI-assisted molecular discovery, and provides a key data foundation for scenarios such as drug discovery and material design.

In recent years, AI has made breakthroughs in drug discovery and materials science, with large language models demonstrating the ability to understand and generate chemical structures. However, high-quality large-scale chemical datasets are a limiting bottleneck. Traditional datasets have issues such as limited scale, incomplete annotations, or restricted access. Researchers urgently need open and comprehensive annotated data resources, leading to the emergence of QuantumChem-200K.

Building the dataset involves technical decisions across multiple stages:

• Molecular Screening: Ensure chemical diversity, covering different sizes, functional groups, and structures, while excluding difficult-to-process or unstable structures;
• Computational Methods: Balance accuracy and efficiency, selecting appropriate theoretical levels (e.g., DFT);
• Quality Control: Check computational convergence, verify result rationality, and detect outliers;
• Metadata Annotation: Supplement computational parameters, molecular sources, and confidence scores to improve usability.

Applications in Quantum Chemical Property Screening

• Drug Discovery: Quickly screen electronic properties, reaction activities, etc., of potential drug molecules to accelerate lead compound optimization;
• Material Design: Explore optimal structures for organic optoelectronic materials and establish structure-property correlation models;
• Reaction Prediction: Infer reaction activity and selectivity to support synthetic route planning.

Benchmarks for Language Model Evaluation

• Molecular Representation Understanding: Test the ability to parse and generate chemical representations such as SMILES;
• Property Prediction and Reasoning: Evaluate the ability to infer physicochemical properties based on structure;
• Scientific Text Generation: Verify the ability to generate accurate chemical descriptions.

QuantumChem-200K advances the data infrastructure for AI in chemistry:

• Lowering Barriers: Open data allows more teams to participate in research without expensive computational resources;
• Promoting Innovation: Standardized benchmarks stimulate algorithmic innovation and accelerate domain progress;
• Interdisciplinary Collaboration: Facilitates cooperation between chemistry, computer science, and other disciplines based on shared data;
• Industrial Translation: Provides resources for building industrial-grade molecular screening systems, shortening the path to application.

Future challenges to address:

• Expand dataset scale, improve computational accuracy, and cover more molecular types;
• Conduct regular maintenance and updates, incorporating new molecules and precise property values;
• Encourage community participation, collect user feedback, supplement data, and share experiences to form a positive cycle.