TargetIntel-IO uses two deliberately separated planes.
The existing deterministic pipeline owns feature construction, biological role classification, therapeutic-intent scoring, ranking, benchmark evaluation, and report generation. It remains authoritative.
The graph plane represents genes, diseases, treatments, datasets, publications, cell types, cohorts, tissue compartments, clinical outcomes, pathways, and other explicit contexts. Scientific observations are represented as immutable assertions rather than flattened causal edges. Each assertion retains evidence identifiers, direction, support relation, review status, effect size, uncertainty, release metadata, limitations, and boundary flags.
Deterministic retrieval, path inspection, community construction, hybrid entity resolution, and provider-neutral GraphRAG exports operate only on graph snapshots. Any future candidate-discovery algorithm must emit a separate, auditable hypothesis bundle and may not promote a candidate into the productive ranking without an explicit validation and authorization gate.
- v0.6.0: graph-ready single-cell and spatial context; reviewed-evidence graph foundation; deterministic local retrieval and BYOG export.
- v0.7.0: patient/cohort/treatment/response/timepoint context, sealed development and holdout boundaries, and clinical GraphRAG research queries.
- v0.8.0: explicit graph discovery algorithms such as personalized PageRank, propagation, path ranking, community bridging, link prediction, popularity adjustment, source ablation, and temporal validation. GraphRAG explains retrieved evidence; it does not decide promotion.
- v0.9.0: multitumor transfer and external validation with disease-specific subgraphs and temporal holdouts.
- v1.0.0: validated multitumor platform with stable graph schema, interchangeable storage backends, API/CLI, and complete traceability.