AiChemist Data Management Plan Version 1

AiChemist Data Management Plan Version 1 Description Data Management Plan (DMP) for the Horizon Europe MSCA "Explainable AI for Molecules - AiChemist". This plan defines which kind of data will be generated and used by AiChemist Doctoral Candidates (DCs), how public and proprietary data will be handled, which types of data can be made available to the public, where the data will be published and under what circumstances. As the DCs progress in their individual research projects, the DMP will be updated and will go into more detail on standards for storing data, reusability and interoperability, and individual datasets and models will be described in the DMP. Which data can be made publicly available? Each DC, in consultation with their supervisors, is responsible for making their developed models and the respective data publicly available upon publication of results, as long as (1) the source data is already publicly available or (2) in cases where proprietary data is used, the company that owns the data allows said data to be published. Purchased data (e.g. from Reaxys) and data linked to patents and IP cannot be made publicly available. How should data be made publicly available? Datasets and models related to publications will be made available on GitHub and/or Huggingface. LICENSE:MIT DOI: - 03/11/2025 Best practices defined by GitHub/Hugginface will be used to make the data/models findable. Each repository will feature a comprehensive description of data, models, nomenclature and file structure in the README/Model Card, and will incorporate appropriate keywords into the README/Model Card. https://docs.github.com/en/contributing/writing-for-github-docs/making-contentfindable-in-search. https://huggingface.co/docs/hub/model-cards. For longer-term storage of data/models, Zotero will be used. Public data related to molecular properties, activities or toxicities, will be stored in OCHEM https://ochem.eu/. Every student working on molecular property and/or activity prediction is responsible for uploading their data/models to OCHEM. How should data be stored? Each DC will store their data on their company/institutional hardware and will regularly make back-ups according to their company/institutional policy. Contact The project manager, Katya Ahmad, is responsible for monitoring compliance to FAIR data principles, updating the DMP, and answering any questions related to data management within the AiChemist Project. Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 Funder Grant European Commission||EC Explainable AI for Molecules - AiChemist (corda_____he::101120466) Researchers Karoline Schjelde (orcid:0009-0002-7381-3435), Dina Khasanova (orcid:0009-0003-7477-7664), Mateusz Iwan (orcid:0000-00015151-4659), Fabian Per Krüger (orcid:0009-0005-6420-2175), Vasilii Fastovskii (orcid:0009-0008-5391-3432), Eric Alcaide (orcid:00000002-7367-0369), Bob Wilhelmus van Schendel (orcid:0009-00084334-3720), Jaehyeon Park (orcid:0009-0004-7459-809X), Ghaith Mqawass (orcid:0009-0005-2514-5999), SUBASHINI KENNEDY (orcid:0000-0002-1024-8152), Andrea Hunklinger (orcid:00090008-6107-3859), Marcel Hiltscher (orcid:0009-0001-6096-3584), Matthew Ball (orcid:0009-0008-8777-1778) Organizations Bayer (Germany), Technical University Eindhoven, AstraZeneca (Sweden), Barcelona University, Barcelona, Spain, Pfizer (Germany), Universiteit Leiden, Molecular Networks (Germany), Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 Universita della Svizzera Italiana, Helmholtz Zentrum München, Ecole Polytechnique Fédéral de Lausanne, Ecole Normale supérieure de Paris, Technische Universität München, Centre for Genomic Regulation, University of Copenhagen, Sanofi (France), Istituto di Ricerche Farmacologiche Mario Negri IRCCS (IRFMN) Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 1. Main Info Title of DMP: AiChemist Data Management Plan Description: Data Management Plan (DMP) for the Horizon Europe MSCA "Explainable AI for Molecules - AiChemist". This plan defines which kind of data will be generated and used by AiChemist Doctoral Candidates (DCs), how public and proprietary data will be handled, which types of data can be made available to the public, where the data will be published and under what circumstances. As the DCs progress in their individual research projects, the DMP will be updated and will go into more detail on standards for storing data, reusability and interoperability, and individual datasets and models will be described in the DMP. Which data can be made publicly available? Each DC, in consultation with their supervisors, is responsible for making their developed models and the respective data publicly available upon publication of results, as long as (1) the source data is already publicly available or (2) in cases where proprietary data is used, the company that owns the data allows said data to be published. Purchased data (e.g. from Reaxys) and data linked to patents and IP cannot be made publicly available. How should data be made publicly available? Datasets and models related to publications will be made available on GitHub and/or Huggingface. Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 Best practices defined by GitHub/Hugginface will be used to make the data/models findable. Each repository will feature a comprehensive description of data, models, nomenclature and file structure in the README/Model Card, and will incorporate appropriate keywords into the README/Model Card. https://docs.github.com/en/contributing/writing-for-github-docs/making-contentfindable-in-search. https://huggingface.co/docs/hub/model-cards. For longer-term storage of data/models, Zotero will be used. Public data related to molecular properties, activities or toxicities, will be stored in OCHEM https://ochem.eu/. Every student working on molecular property and/or activity prediction is responsible for uploading their data/models to OCHEM. How should data be stored? Each DC will store their data on their company/institutional hardware and will regularly make back-ups according to their company/institutional policy. Contact The project manager, Katya Ahmad, is responsible for monitoring compliance to FAIR data principles, updating the DMP, and answering any questions related to data management within the AiChemist Project. Researchers: Karoline Schjelde (orcid:0009-0002-7381-3435) Dina Khasanova (orcid:0009-0003-7477-7664) Mateusz Iwan (orcid:0000-0001-5151-4659) Fabian Per Krüger (orcid:0009-0005-6420-2175) Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 Vasilii Fastovskii (orcid:0009-0008-5391-3432) Eric Alcaide (orcid:0000-0002-7367-0369) Bob Wilhelmus van Schendel (orcid:0009-0008-4334-3720) Jaehyeon Park (orcid:0009-0004-7459-809X) Ghaith Mqawass (orcid:0009-0005-2514-5999) SUBASHINI KENNEDY (orcid:0000-0002-1024-8152) Andrea Hunklinger (orcid:0009-0008-6107-3859) Marcel Hiltscher (orcid:0009-0001-6096-3584) Matthew Ball (orcid:0009-0008-8777-1778) Organizations: Bayer (Germany) Technical University Eindhoven AstraZeneca (Sweden) Barcelona University, Barcelona, Spain Pfizer (Germany) Universiteit Leiden Molecular Networks (Germany) Universita della Svizzera Italiana Helmholtz Zentrum München Ecole Polytechnique Fédéral de Lausanne Ecole Normale supérieure de Paris Technische Universität München Centre for Genomic Regulation University of Copenhagen Sanofi (France) Istituto di Ricerche Farmacologiche Mario Negri IRCCS (IRFMN) Contact: Ahmad Katya Ahmad ([email protected]) 2. Funding Funding organizations: European Commission||EC Grants: Explainable AI for Molecules - AiChemist (corda_____he::101120466) Project: Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 3. License License: MIT Access Rights: Public 4. Templates Descriptions Personalized Cardiotoxicity Datasets The dataset is derived from the FAERS pharmacovigilance database and includes drug– adverse event reports mapped to active ingredients and SMILES strings. Cardiotoxicity labels were assigned using disproportionality analysis metrics (ROR, PRR, IC) based on selected MedDRA terms. Demographic factors (sex, age, weight) were incorporated through stratified subsets. Multiple dataset versions were created to reflect variations in inclusion criteria and labeling methods. Label certainty scores are provided for use as sample weights in modeling. Dataset sizes range from ~2,000 to ~40,000 entries. Template: Data Management Plan Template - University of Bologna Type: Dataset 1 Your research output 1.1 Contact details and responsibilities 1.1.1 Contact person(s) Contact Name Contact ID E-mail address Affiliation Iwan, Mateusz 0000-0001-51514659 mateusz.iwan@hot mail.com Mario Negri Institute for Pharmacologial Research Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 1.1.3 Research output creator(s) and contributor(s Contributor Name Contributor ID E-mail address Affiliation Role Iwan, Mateusz 0000-00015151-4659 mateusz.iwan@ hotmail.com Mario Negri Institute for Pharmacologic al Research Creator Roncaglioni, Alessandra 0000-00021734-0939 alessandra.ronc aglioni@marion egri.it Mario Negri Institute for Pharmacologic al Research Project member Garcia de Lomana, Marina 0000-00029310-7290 marina.garciad elomana@baye r.com Bayer AG Project member Grisoni, Francesca 0000-00018552-6615 [email protected] Eindhoven University of Technology Project member 1.1.4 Institution(s) involved a. Istituto di Ricerche Farmacologiche Mario Negri IRCCS b. Bayer AG c. Eindhoven University of Technology 1.2 Resources for data management 1.2.1 What resources will be dedicated to data management and ensuring that data will be FAIR? As the dataset has already been collected, cleaned, and curated, no additional financial costs are expected for data collection or processing. Time resources were primarily dedicated during the data preparation phase, including standardization, Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 instructions for reproduction. - Paper: A peer-reviewed publication will serve as an authoritative reference, providing a comprehensive description of data sources, processing pipeline, quality checks, and intended use cases. Additional plain-text documentation may be included alongside downloadable datasets. 4.2.3 Keywords Pharmacovigilance,Cardiotoxicity,Disproportionality Analysis,Machine Learning, FAERS,QSAR 4.3 Accessibility 4.3.1 Is the research output openly accessible? Yes 4.3.3 Accessing the research output The finalized version of the dataset will be made publicly available through the OCHEM platform, ensuring long-term accessibility and integration with cheminformatics tools. In addition, a versioned archive of the dataset will be deposited on Zenodo, which will assign a persistent DOI and support standardized metadata for citation and discovery. Metadata and supporting documentation will be provided either directly on the OCHEM page or via a public GitHub repository that also hosts the associated code and preprocessing pipeline. This setup will ensure that users can easily access, understand, and reuse the dataset. 4.3.4 Is metadata openly accessible? Yes 4.4 Interoperability 4.4.1 Methodologies The dataset was prepared using a standardized pipeline to ensure interoperability. Drug data were mapped to active ingredients using open chemical databases (e.g., ChEMBL, PubChem), and structural information was represented via SMILES strings. Labels were derived using established Disproportionality Analysis metrics Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 (e.g., IC, ROR, PRR) and standardized MedDRA Preferred Terms. Demographic attributes were harmonized into predefined categories (e.g., Sex, Age groups, Weight classes). Data processing was conducted using widely adopted tools such as Python (pandas, RDKit, NumPy), ensuring reproducibility and compatibility. The dataset is stored in open formats (CSV), enabling easy integration with other systems and tools. 4.4.2 Vocabularies, taxonomies and other standards The dataset leverages several established vocabularies and standards to ensure semantic interoperability and facilitate reuse: - MedDRA (Medical Dictionary for Regulatory Activities): Used for the definition and grouping of cardiotoxicity-related adverse events through standardized Preferred Terms (PTs), ensuring consistent labeling across pharmacovigilance studies. - SMILES: Structural chemical information is encoded using SMILES notation, a widely adopted standard in cheminformatics. - FAERS standard fields: Original source data adheres to FDA FAERS schema and controlled vocabularies. 4.5 Reusability 4.5.1 Licensing CC BY 4.0, "Attribution", https://creativecommons.org/licenses/by/4.0/ 4.5.2 Documentation The dataset will be accompanied by a full peer-reviewed publication that describes the data preparation, processing steps, labeling methodology, and model development in detail. Additionally, a README file will be included with the dataset to summarize its structure, content, and usage instructions. All relevant code will be made available via a public GitHub repository to ensure full transparency and reproducibility. Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 5 Quality and security 5.1 Quality of the research output 5.1.1 Measures implemented to ensure quality Several measures were implemented to ensure the consistency and quality of the dataset: - Standardized data processing pipeline: All data were processed using reproducible Python scripts with strict validation checks at each stage (e.g., mapping, cleaning, labeling). - Use of controlled vocabularies: Drug names, adverse events, and demographic attributes were normalized using established vocabularies such as MedDRA and standardized chemical identifiers (SMILES). - Manual inspection: Sample subsets were manually inspected during critical steps (e.g., mapping drugs to active ingredients) to ensure accuracy. - Version control: All code and data transformations are tracked using version control (Git), enabling full traceability and reproducibility. - Cross-checking label consistency: Disproportionality analysis results were verified using multiple metrics (IC, ROR, PRR), and label confidence was calculated to highlight uncertain cases. 5.2 Security measures 5.2.1 Storage Local Hard Drive (PC, laptop) Comment: During the active phase of the research, all data were stored locally on two dedicated, access-controlled PCs. Code and data processing scripts were versioncontrolled and collaboratively managed via GitHub. To ensure redundancy and data security, key datasets and essential intermediate files were also backed up to two separate cloud storage services. These backups were sufficient to fully reconstruct the final dataset. Access to local and cloud resources was restricted to authorized team members and managed internally. Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 5.2.2 Backup During the research, intermediate files and processed datasets were regularly backed up to cloud storage platforms (Dropbox and Google Drive). Access to all storage locations—both local and online—was restricted to authorized project team members. In the event of data loss or corruption, files could be recovered from the most recent cloud backup or reconstructed from GitHub and preserved intermediate data. 5.2.3 Additional security The dataset does not contain any sensitive, personal, or proprietary information. All records are fully anonymized and derived from publicly available pharmacovigilance sources, primarily the FDA FAERS database. Given the nonsensitive nature of the data, no additional security measures beyond standard data protection, controlled access during the active research phase, and regular backups were deemed necessary. Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 Benchmark datasets molecular property prediction These datasets were used to benchmark representation learning methods based on their performance in tasks relevant to drug discovery. The datasets were collected from established platforms that curate benchmark datasets for machine learning applications in drug discovery and were subsequently cleaned and standardized to enhance their overall quality, coherence, and usability. Template: Data Management Plan Template - University of Bologna Type: Dataset 1 Your research output 1.1 Contact details and responsibilities 1.1.1 Contact person(s) Contact Name Contact ID E-mail address Affiliation Krüger Fabian 0009-0005-64202175 fabian.krueger@tu m.de Technische Universität München, Helmholtz Munich, AstraZeneca 1.1.3 Research output creator(s) and contributor(s Contributor Name Contributor ID E-mail address Affiliation Role Krüger, Fabian 0009-00056420-2175 fabian.krueger @tum.de Technische Universität München, Creator Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 Helmholtz Munich, AstraZeneca Tetko, Igor 0000-00026855-0012 [email protected] Helmholtz Munich Supervisor Engkvist, Ola 0000-00034970-6461 Ola.Engkvist@a strazeneca.com AstraZeneca Supervisor 1.1.4 Institution(s) involved a. Technische Universität München (TUM) b. AstraZeneca AB c. Helmholtz Munich 1.2 Resources for data management 1.2.1 What resources will be dedicated to data management and ensuring that data will be FAIR? N/A. The data and all methods of preparing it are published in an open access scientific journal article. Thus, all the costs for maintained public access are already paid 1.3 Identifying the research output 1.3.1 Work package and task number WP Task 1 1 1.3.2 Which type of research output are you describing? Dataset Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 2 Summary 2.1 Identification of the dataset 2.1.1 Are you re-using an already existing dataset? Yes 2.1.2 Re-using an existing dataset: why and how We made a collection of publicly available datasets to benchmark representation learning methods across a range of different tasks. We cleaned the data to ensure the data is standardized across tasks. All datasets we have collected are licensed for free reuse and re-distribution, therefore we do not have any potential restrictions to data re-use 2.2 Dataset characteristics 2.2.1 Dataset name Ames, BBB, hERG, DEL 2.2.2 Data types a. quantitative b. processed c. text to scalar 2.2.3 Data formats comma separated values 2.2.4 Size of data kilobytes (KB) 2.2.5 How is the data generated/collected We used four different datasets to predict pharmacologically relevant molecular properties. The datasets differ in size, task, and class imbalance. The first dataset is used for mutagenicity prediction [18, 19]. It contains Ames test results for 7,255 drugs. Of these, 54% show positive results. The second dataset assesses bloodbrain barrier permeability [17]. It contains 1,909 molecules, with 76% able to Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 penetrate the barrier. The third dataset provides information on the inhibition of the potassium ion channel encoded by the human Ether-à-go-go-Related Gene (hERG) [21]. Inhibition is defined as a half-maximal inhibitory concentration of less than 10 M. This dataset contains 306,341 compounds, with 4.5% being inhibitors. These three datasets were obtained from Therapeutics Data Commons [22]. The fourth dataset contains information on whether a molecule is enriched in a DNA-encoded library (DEL) for binding to carbonic anhydrase IX [20]. Positive enrichment is defined as the top 5% of enrichment scores. This dataset includes 108,528 molecules, with 4.9% showing enrichment after cleaning the data. 17 Martins Ines Filipa, Teixeira Ana L, Pinheiro Luis, Falcao Andre O (2012) A bayesian approach to in silico blood-brain barrier penetration modeling. J Chem Inf Model 52(6):1686–1697 18 Hansen Katja, Mika Sebastian, Schroeter Timon, Sutter Andreas, Ter Laak Antonius, Steger-Hartmann Thomas, Heinrich Nikolaus, Muller Klaus-Robert (2009) Benchmark data set for in silico prediction of ames mutagenicity. J Chem Inf Model 49(9):2077–2081 19 Congying Xu, Cheng Feixiong, Chen Lei, Zheng Du, Li Weihua, Liu Guixia, Lee Philip W, Tang Yun (2012) In silico prediction of chemical ames mutagenicity. J Chem Inf Model 52(11):2840–2847 20 Lim Katherine S, Reidenbach Andrew G, Hua Bruce K, Mason Jeremy W, Gerry Christopher J, Clemons Paul A, Coley Connor W (2022) Machine learning on dnaencoded library count data using an uncertainty-aware probabilistic loss function. J Chem Inf Model 62(10):2316–2331 21 Fang Du, Haibo Yu, Zou Beiyan, Babcock Joseph, Long Shunyou, Li Min (2011) Hergcentral: a large database to store, retrieve, and analyze compound-human ether-a-go-go related gene channel interactions to facilitate cardiotoxicity assessment in drug development. Assay Drug Dev Technol 9(6):580–588 Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 22 Huang Kexin, Tianfan Fu, Gao Wenhao, Zhao Yue, Roohani Yusuf, Leskovec Jure, Coley Connor W, Xiao Cao, Sun Jimeng, Zitnik Marinka (2021) Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development. arXiv preprint arXiv:2102.09548 2.2.6 Why is the data generated/collected We made a collection of publicly available datasets to benchmark representation learning methods across a range of different tasks (which is the objective of the project). 2.2.7 Dataset description This collection includes 4 datasets that reflect a range of common tasks in drug discovery. All are binary classification data. The first dataset describes whether molecules pass the blood brain barrier, the second describes the results of the ames test, the third whehter a molecule inhibits the hERG encoded potassium receptor, and the last one whether a molecule is enriched in a DNA-encoded library (DEL) for binding to carbonic anhydrase IX. 3 Ethics and legal aspects 3.1 Personal data 3.1.1 Are you handling personal data? No 3.2 Legal aspects 3.2.1 Are there any legal issues associated to your ouput? (e.g. IPR or valorization) No 3.3 Other ethical aspects 3.3.1 Are there other ethical issues associated to your ouput? No 4 Making research FAIR 4.1 Long-term preservation 4.1.1 Selecting what to preserve Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025 The data is published together with the corresponding research article in Journal of Cheminformatics (open access article): https://doi.org/10.1186/s13321-025-00982-w 4.1.2 Chosen repository/ies Name Link Is your repository OpenAIRE compliant? Is your repository catalogued in re3data? What type of repository is it? GitHub https:// github.com/ FabianKruger/ molprivacy Yes No Generalist Journal of Cheminformati cs https:// doi.org/ 10.1186/ s13321-02500982-w Yes No Generalist 4.2 Findability 4.2.1 Persistent identifier(s) Type of PID (e.g., DOI, Handle) PID (e.g., doi.org/10...) DOI https://doi.org/10.1186/s13321-025-00982w 4.2.2 Metadata All the meta data is in the published research paper: https://doi.org/10.1186/s13321-025-00982-w 4.2.3 Keywords Data Management Plan | AiChemist Data Management Plan LICENSE:MIT DOI: - 03/11/2025