Combine scientific data and code without erasing provenance

Scientific data does not become decision evidence merely because an API returned it. Identity, provenance, method and coverage determine whether a record is fit to change a model.

Acquire before interpreting

ARiDA separates live acquisition from deterministic computation. Connectors retrieve molecular, target, trial, regulatory, patent and market records. Sources such as ChEMBL expose structured bioactivity data (ChEMBL); Open Targets separates evidence by data source and type (Open Targets). ARiDA normalizes records while preserving the original source, retrieval time and ledger identity.

Entity resolution comes next. A molecule name, development code and salt form may refer to related but non-identical objects. A target symbol can be ambiguous across species or complexes. Unresolved identity remains a gap rather than becoming a convenient join.

Execute science under a method contract

Governed compute profiles run sanctioned scripts in controlled sandboxes. The asset-science lane uses real RDKit, scikit-learn and SciPy capabilities. A molecular profile can yield descriptors, structural alerts and fingerprints. RDKit documents the Morgan fingerprint implementation and its chemical feature variants (RDKit).

The compute record includes the engine identifier, inputs, source links and artifact hash. Downstream logic checks that contract. A descriptor row without source references and ledger identifiers cannot adjust PoS. A similarity matrix without full portfolio coverage cannot change correlation.

Let computation change a decision carefully

Scientific engines feed explicit business mechanisms. Druggability can modify benchmark transition probabilities. Molecular similarity can inform covariance. Target evidence can change advance or pass. Pathway enrichment and translational readiness can change indication rank. Uncertainty can enter information-value calculations.

Each transformation is inspectable and bounded. The source record remains distinct from the derived feature; the derived feature remains distinct from the accepted assumption; the assumption remains distinct from the final decision.

This layered design is slower to describe than “we combine data and AI.” It is also the reason the result can be reviewed.