[Atlas] Cross-reference tabular datasets with KG: bidirectional entity linking done analysis:5 coding:5 safety:9

[Atlas] Cross-reference tabular datasets with KG: bidirectional entity linking

Goal

Create a seamless bridge between the graph world model and tabular observations. When a researcher browses a gene in the KG, they should see which datasets contain measurements for that gene. When browsing a dataset, they should see rich KG context for each column.

Bidirectional Linking

KG → Datasets ("What data exists for this entity?")

When viewing a KG entity page (e.g., gene TREM2):

Datasets containing TREM2:
┌─────────────────────────────┬──────────────┬──────────┬────────────┐
│ Dataset                      │ Column       │ Rows     │ Values     │
├─────────────────────────────┼──────────────┼──────────┼────────────┤
│ SEA-AD Differential Expr.    │ gene_symbol  │ 15,000   │ log_fc=2.3 │
│ Allen Brain Cell Atlas       │ gene         │ 50,000   │ expr=high  │
│ AD GWAS Summary Stats        │ gene_name    │ 8,000    │ p=1.2e-8   │
└─────────────────────────────┴──────────────┴──────────┴────────────┘

Implementation:

def get_datasets_for_entity(entity_name, entity_type=None):
    """Query artifact_links and column metadata to find datasets
    containing this entity. Returns dataset artifacts with the 
    specific column and summary statistics for that entity."""

Datasets → KG ("What does the KG know about this data?")

When viewing a tabular dataset page:

Column: gene_symbol (linked to: gene entities)
  Known genes in KG: 12,450 / 15,000 (83% coverage)
  Top entities by KG connectivity:
    TREM2 (142 edges) | APOE (138 edges) | APP (95 edges) | MAPT (89 edges)
  
Column: cell_type (linked to: cell_type entities)
  Known cell types in KG: 8 / 8 (100% coverage)
  microglia (256 edges) | astrocyte (198 edges) | neuron (312 edges)

Implementation:

def get_kg_context_for_dataset(dataset_artifact_id):
    """For each linked column in the dataset, resolve entities against
    the KG and return coverage stats and top entities by connectivity."""

Auto-Linking on Registration

When a tabular dataset is registered (via register_tabular_dataset):

For each column with linked_entity_type:

- Sample N values from the column
- Resolve each against KG entities of that type
- Create artifact_links from dataset to matched entities

Compute and store coverage statistics in dataset metadata

Flag columns with low coverage for review

UI Components

• KG entity page: "Related Datasets" section with table of matching datasets
• Dataset detail page: "KG Coverage" section per linked column
• Click-through: clicking a dataset row for an entity navigates to the dataset filtered to that entity

Acceptance Criteria

Dependencies

• atl-ds-01-GROW (KG edge extraction pipeline)

Dependents

• d16-22-TABD0001 (demo: gene expression cross-referencing)

Work Log