ChemPile-MLIFT

A comprehensive multimodal dataset for chemistry property prediction using vision large language models

📋 Dataset Summary

ChemPile-MLIFT is a dataset designed for multimodal chemistry property prediction tasks, specifically focusing on the prediction of chemical properties using vision large language models (VLLMs). It is part of the ChemPile project, which aims to create a comprehensive collection of chemistry-related data for training LLMs. The dataset includes a wide range of chemical properties and is structured to facilitate the training of models that can understand and predict chemical properties based on textual descriptions, molecular representations and the image of the corresponding molecule. Thus, each example in the dataset contains the image of the molecule involved.

The origin of the dataset property data is from well-known chemistry datasets such as the QM9 dataset, which contains quantum mechanical properties of small organic molecules, and the RDKit dataset, which includes a wide range of chemical properties derived from molecular structures. Each of the subsets or Hugging Face configurations corresponds to a different source of chemical property data, allowing for diverse training scenarios:

📊 Dataset Statistics

The resulting dataset contains 61.5M examples with one image per example, making it a substantial resource for training and evaluating VLLMs in the field of chemistry.

🗂️ Dataset Configurations

The dataset includes diverse configurations covering various chemical property domains, organized by thematic groups:

💊 Drug Discovery and ADMET Properties

• BACE-multimodal
• BBBP-multimodal
• bioavailability_ma_et_al-multimodal
• blood_brain_barrier_martins_et_al-multimodal
• caco2_wang-multimodal
• clearance_astrazeneca-multimodal
• cyp2c9_substrate_carbonmangels-multimodal
• cyp2d6_substrate_carbonmangels-multimodal
• cyp3a4_substrate_carbonmangels-multimodal
• cyp_p450_1a2_inhibition_veith_et_al-multimodal
• cyp_p450_2c19_inhibition_veith_et_al-multimodal
• cyp_p450_2c9_inhibition_veith_et_al-multimodal
• cyp_p450_2d6_inhibition_veith_et_al-multimodal
• cyp_p450_3a4_inhibition_veith_et_al-multimodal
• drug_induced_liver_injury-multimodal
• freesolv-multimodal
• half_life_obach-multimodal
• human_intestinal_absorption-multimodal
• lipophilicity-multimodal
• p_glycoprotein_inhibition_broccatelli_et_al-multimodal
• pampa_ncats-multimodal
• solubility_aqsoldb-multimodal
• volume_of_distribution_at_steady_state_lombardo_et_al-multimodal

⚠️ Toxicology and Safety Assessment

• ames_mutagenicity-multimodal
• carcinogens-multimodal
• clintox-multimodal
• herg_blockers-multimodal
• herg_central_at_10uM-multimodal
• herg_central_at_1uM-multimodal
• herg_central_inhib-multimodal
• herg_karim_et_al-multimodal
• ld50_catmos-multimodal
• ld50_zhu-multimodal
• nr_ahr_tox21-multimodal
• nr_ar_lbd_tox21-multimodal
• nr_ar_tox21-multimodal
• nr_aromatase_tox21-multimodal
• nr_er_lbd_tox21-multimodal
• nr_er_tox21-multimodal
• nr_ppar_gamma_tox21-multimodal
• SIDER-multimodal
• sigma_aldrich_safety_data-multimodal
• skin_reaction-multimodal
• sr_are_tox21-multimodal
• sr_atad5_tox21-multimodal
• sr_hse_tox21-multimodal
• sr_mmp_tox21-multimodal
• sr_p53_tox21-multimodal

🎯 Bioactivity and Target Interaction

• cav3_t-type_calcium_channels_butkiewicz-multimodal
• chembl_v29-multimodal
• hiv-multimodal
• m1_muscarinic_receptor_agonists_butkiewicz-multimodal
• m1_muscarinic_receptor_antagonists_butkiewicz-multimodal
• MUV_466-multimodal
• MUV_548-multimodal
• MUV_600-multimodal
• MUV_644-multimodal
• MUV_652-multimodal
• MUV_689-multimodal
• MUV_692-multimodal
• MUV_712-multimodal
• MUV_713-multimodal
• MUV_733-multimodal
• MUV_737-multimodal
• MUV_810-multimodal
• MUV_832-multimodal
• MUV_846-multimodal
• MUV_852-multimodal
• MUV_858-multimodal
• MUV_859-multimodal
• orexin1_receptor_butkiewicz-multimodal
• sarscov2_3clpro_diamond-multimodal
• sarscov2_vitro_touret-multimodal
• serine_threonine_kinase_33_butkiewicz-multimodal
• uniprot_binding_single-multimodal
• uniprot_binding_sites_multiple-multimodal

🔬 Computational Chemistry and Quantum Properties

• flashpoint-multimodal
• mol2svg-multimodal
• opv-multimodal
• qm8-multimodal
• qm9-multimodal
• rdkit_features-multimodal
• rdkit_features-multimodal-chunk-1
• rdkit_features-multimodal-chunk-2
• rdkit_features-multimodal-chunk-3
• smiles_to_3d-multimodal
• thermosol-multimodal

📚 Chemical Knowledge Bases and Databases

• aminoacids-multimodal
• chebi_20-multimodal
• drugchat_liang_zhang_et_al-multimodal
• mona-multimodal
• moses-multimodal
• RedDB-multimodal
• zinc-multimodal

📜 License

All content is made open-source under the CC BY-NC-SA 4.0 license, allowing for:

• ✅ Non-commercial use and sharing with attribution
• ✅ Modification and derivatives
• ⚠️ Attribution required
• ⚠️ Non-commercial use only

📖 Data Fields

The dataset contains the following fields for all the configurations allowing for a consistent structure across different chemical property datasets:

• IMAGE: The image representation of the molecule, typically in PNG format. This image is crucial for vision-based tasks and is derived from the molecular structure.
• template: The filled template with the SMILES of the molecule and the corresponding chemical property.
• template_original: The template or schema used for the chemical property prediction task. This field provides a structured format for the input and output data, with different templates used for ensure diversity in the tasks.
• SMILES: The SMILES (Simplified Molecular Input Line Entry System) representation of the molecule.
• SELFIES: The SELFIES (SELF-referencIng Embedded Strings) representation of the molecule, which is a more robust alternative to SMILES for representing chemical structures.
• InChI: The International Chemical Identifier (InChI) representation of the molecule, providing a unique identifier for chemical substances.
• IUPAC: The IUPAC (International Union of Pure and Applied Chemistry) name of the molecule, which is a systematic way to name chemical compounds.
• split: The split of the dataset, indicating whether the example is part of the training, validation, or test set. This field helps in organizing the dataset for model training and evaluation.
• property or properties: The specific chemical property or properties being predicted. This field varies depending on the dataset configuration and can include properties such as solubility, bioavailability, toxicity, etc. Each configuration focuses on a different set of chemical properties.

🔬 Dataset Groups Detailed Description

💊 Drug Discovery and ADMET Properties

The Drug Discovery and ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) group encompasses datasets focused on pharmaceutical compound development and safety assessment. This collection includes bioavailability prediction (bioavailability_ma_et_al-multimodal), blood-brain barrier permeability (BBBP-multimodal, blood_brain_barrier_martins_et_al-multimodal), intestinal permeability (caco2_wang-multimodal, pampa_ncats-multimodal), hepatic clearance (clearance_astrazeneca-multimodal), drug-induced liver injury assessment (drug_induced_liver_injury-multimodal), and various CYP450 enzyme interactions (cyp2c9_substrate_carbonmangels-multimodal, cyp2d6_substrate_carbonmangels-multimodal, cyp3a4_substrate_carbonmangels-multimodal, cyp_p450_1a2_inhibition_veith_et_al-multimodal, cyp_p450_2c19_inhibition_veith_et_al-multimodal, cyp_p450_2c9_inhibition_veith_et_al-multimodal, cyp_p450_2d6_inhibition_veith_et_al-multimodal, cyp_p450_3a4_inhibition_veith_et_al-multimodal). The group also covers crucial pharmacokinetic properties such as half-life prediction (half_life_obach-multimodal), human intestinal absorption (human_intestinal_absorption-multimodal), lipophilicity (lipophilicity-multimodal), volume of distribution (volume_of_distribution_at_steady_state_lombardo_et_al-multimodal), P-glycoprotein inhibition (p_glycoprotein_inhibition_broccatelli_et_al-multimodal), and solubility (freesolv-multimodal). These datasets are essential for early-stage drug development, helping predict whether compounds will have favorable drug-like properties before expensive clinical trials.

⚠️ Toxicology and Safety Assessment

The Toxicology and Safety Assessment group focuses on predicting harmful effects of chemical compounds across various biological systems. This includes mutagenicity prediction (ames_mutagenicity-multimodal), carcinogenicity assessment (carcinogens-multimodal), acute toxicity (ld50_catmos-multimodal, ld50_zhu-multimodal), and comprehensive toxicity screening through the Tox21 initiative datasets (nr_ahr_tox21-multimodal, nr_ar_lbd_tox21-multimodal, nr_ar_tox21-multimodal, nr_aromatase_tox21-multimodal, nr_er_lbd_tox21-multimodal, nr_er_tox21-multimodal, nr_ppar_gamma_tox21-multimodal, sr_are_tox21-multimodal, sr_atad5_tox21-multimodal, sr_hse_tox21-multimodal, sr_mmp_tox21-multimodal, sr_p53_tox21-multimodal). The group also includes specialized toxicity assessments such as hERG channel blocking potential (herg_blockers-multimodal, herg_central_at_10uM-multimodal, herg_central_at_1uM-multimodal, herg_central_inhib-multimodal, herg_karim_et_al-multimodal), clinical toxicity prediction (clintox-multimodal), adverse drug reactions (SIDER-multimodal), skin reactions (skin_reaction-multimodal), and safety data assessment (sigma_aldrich_safety_data-multimodal). These datasets are crucial for environmental safety assessment and pharmaceutical safety profiling.

🎯 Bioactivity and Target Interaction

The Bioactivity and Target Interaction group contains datasets focused on molecular interactions with specific biological targets. This includes enzyme inhibition studies (BACE-multimodal for β-secretase), receptor binding and modulation data from the Butkiewicz collection (cav3_t-type_calcium_channels_butkiewicz-multimodal, m1_muscarinic_receptor_agonists_butkiewicz-multimodal, m1_muscarinic_receptor_antagonists_butkiewicz-multimodal, orexin1_receptor_butkiewicz-multimodal, serine_threonine_kinase_33_butkiewicz-multimodal), and comprehensive screening datasets like MUV (Maximum Unbiased Validation) series covering various protein targets (MUV_466-multimodal, MUV_548-multimodal, MUV_600-multimodal, MUV_644-multimodal, MUV_652-multimodal, MUV_689-multimodal, MUV_692-multimodal, MUV_712-multimodal, MUV_713-multimodal, MUV_733-multimodal, MUV_737-multimodal, MUV_810-multimodal, MUV_832-multimodal, MUV_846-multimodal, MUV_852-multimodal, MUV_858-multimodal, MUV_859-multimodal). The group also includes antiviral activity data (hiv-multimodal, sarscov2_3clpro_diamond-multimodal, sarscov2_vitro_touret-multimodal), protein binding studies (uniprot_binding_single-multimodal, uniprot_binding_sites_multiple-multimodal), and comprehensive chemical bioactivity databases (chembl_v29-multimodal). These datasets enable the development of target-specific therapeutics and understanding of molecular mechanisms of action.

🔬 Computational Chemistry and Quantum Properties

The Computational Chemistry and Quantum Properties group encompasses datasets for predicting fundamental quantum mechanical and computational chemistry properties. This includes quantum mechanical property prediction (qm8-multimodal, qm9-multimodal), molecular descriptor calculations (rdkit_features-multimodal, rdkit_features-multimodal-chunk-1, rdkit_features-multimodal-chunk-2, rdkit_features-multimodal-chunk-3), thermodynamic properties (flashpoint-multimodal, thermosol-multimodal), and molecular visualization and structure generation (mol2svg-multimodal, smiles_to_3d-multimodal). These datasets are fundamental for computational chemistry research, enabling the prediction of molecular properties from first principles and supporting the development of new theoretical models and computational methods.

📚 Chemical Knowledge Bases and Databases

The Chemical Knowledge Bases and Databases group contains datasets derived from established chemical and biological databases. This includes comprehensive chemical databases (chebi_20-multimodal, zinc-multimodal), mass spectrometry databases (mona-multimodal), molecular generation and chemical space exploration (moses-multimodal), conversational chemistry data (drugchat_liang_zhang_et_al-multimodal), and amino acid properties (aminoacids-multimodal). These datasets provide broad coverage of chemical space and enable the development of AI systems that can navigate and reason about diverse chemical information from established scientific databases and knowledge repositories.

🚀 Usage

from datasets import load_dataset, get_dataset_config_names

# Print available configs for the dataset
configs = get_dataset_config_names("jablonkagroup/chempile-mlift")
print(f"Available configs: {configs}")
# Available configs: ['BACE-multimodal', 'BBBP-multimodal', 'MUV_466-multimodal', ...

dataset = load_dataset("jablonkagroup/chempile-mlift", name=configs[0])
# Loading config: BACE-multimodal

print(dataset)
# DatasetDict({
    # train: Dataset({
        # features: ['IMAGE', 'template', 'template_original', 'SMILES', ... 
        # num_rows: 5440
    # })
    # validation: Dataset({
        # features: ['IMAGE', 'template', 'template_original', 'SMILES', ...
        # num_rows: 480
    # })
    # test: Dataset({
        # features: ['IMAGE', 'template', 'template_original', 'SMILES', ...
    # })
# })

split_name = list(dataset.keys())[0]
sample = dataset[split_name][0]
print(sample)
# {
#     'IMAGE': <PIL.PngImagePlugin.PngImageFile ..., 
#     'template': 'The chemical with the SMILES ..., 
#     'template_original': 'The {#compound|chemical!} ...', 
#     'SMILES': 'Cc1ccccc1-c1ccc2nc(N)c(C[C@@H](C)C(=O)...', 
#     'SMILES_ORIGINAL': 'Cc1ccccc1-c1ccc2nc(N)c(C[C@@H]...', 
#     'SELFIES': '[C][C][=C][C][=C][C][=C][Ring1]...', 
#     'InChI': 'InChI=1S/C27H33N3O2/c1-17-7-5-6-8-...', 
#     'IUPAC': '(2R)-3-[2-amino-6-(2-methylphenyl)...', 
#     'split': 'train', 
#     'pIC50': 9.1549015, 
#     'BACE_inhibition': 1
# }

🎯 Use Cases

• 🖼️ Multimodal Chemical Property Prediction: Training vision-language models to predict molecular properties using both molecular images and textual descriptions
• 💊 Drug Discovery: Building systems for pharmaceutical compound screening using visual molecular representations
• ⚠️ Safety Assessment: Developing multimodal models for toxicity and environmental impact prediction
• 🔬 Materials Design: Creating AI tools that leverage both visual and textual molecular information for materials science
• 📖 Scientific Multimodal Understanding: Training models to understand and reason about chemical information across multiple modalities

⚠️ Limitations & Considerations

• Scope: Focused on chemistry and materials science; domain-specific terminology and concepts
• Quality: Variable quality across sources; expert curation applied but some noise may remain
• Bias: Reflects biases present in chemical databases and scientific literature
• License: Non-commercial use only under CC BY-NC-SA 4.0
• Language: Primarily English content
• Completeness: Some datasets may have missing values or incomplete property annotations
• Image Quality: Molecular images are automatically generated and may vary in visual quality

🛠️ Data Processing Pipeline

1. Collection: Automated extraction from well-known chemistry datasets (QM9, RDKit, Tox21, etc.)
2. Standardization: Consistent formatting and SMILES representation across all configurations
3. Image Generation: Creation of molecular structure images from SMILES representations
4. Template Generation: Conversion of structured data into natural language templates
5. Quality Control: Expert curation and validation of chemical property representations
6. Deduplication: Removal of duplicate entries and data cleaning
7. Validation: Train/validation/test splits and quality checks

🏗️ ChemPile Collection

This dataset is part of the ChemPile collection, a comprehensive open dataset containing over 75 billion tokens of curated chemical data for training and evaluating general-purpose models in the chemical sciences.

Collection Overview

• 📊 Scale: 75+ billion tokens across multiple modalities
• 🧬 Modalities: Structured representations (SMILES, SELFIES, IUPAC, InChI), scientific text, executable code, and molecular images
• 🎯 Design: Integrates foundational educational knowledge with specialized scientific literature
• 🔬 Curation: Extensive expert curation and validation
• 📈 Benchmarking: Standardized train/validation/test splits for robust evaluation
• 🌐 Availability: Openly released via Hugging Face

📄 Citation

If you use this dataset in your research, please cite:

@article{mirza2025chempile0,
  title   = {ChemPile: A 250GB Diverse and Curated Dataset for Chemical Foundation Models},
  author  = {Adrian Mirza and Nawaf Alampara and Martiño Ríos-García and others},
  year    = {2025},
  journal = {arXiv preprint arXiv:2505.12534}
}

👥 Contact & Support

• Paper: arXiv:2505.12534
• Website: ChemPile Project
• Dataset: Hugging Face
• Issues: Please report data issues or questions via the Hugging Face dataset page

---

Part of the ChemPile project - Advancing AI for Chemical Sciences