Defining the Data set Defines the QSAR Claim.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Machine learning significantly advances quantitative structure-activity relationship (QSAR) modeling.
• However, critical methodological choices in chemical representation, endpoint definition, and evaluation design are often poorly documented.
• Inconsistent standardization and data leakage inflate performance metrics, masking real-world applicability weaknesses.

Purpose of the Study:

• To propose a standardized framework for documenting and validating QSAR modeling methodologies.
• To enhance the transparency, reproducibility, and trustworthiness of predictive claims in cheminformatics.
• To shift focus from model architecture to verifiable data processing and evaluation protocols.

Main Methods:

• Introduction of data set contracts: executable and auditable documents.
• Explicit declaration of chemical processing rules, endpoint definitions, and aggregation logic.
• Inclusion of data splits and leakage diagnostics tailored to specific prediction scenarios.

Main Results:

• Data set contracts are feasible using current open-source tools.
• The proposed framework facilitates transparent and reproducible QSAR modeling.
• This approach addresses the inflation of performance metrics due to inconsistent practices.

Conclusions:

• Data set contracts offer a robust solution for auditable and trustworthy QSAR modeling.
• The methodology promotes a paradigm shift towards verifiable scientific claims in machine learning applications.
• Adoption of these contracts will improve the reliability of QSAR models in practical applications.