KGIndex

India --:--:-- --

All work

02 · Clinical evidence graph

Athena

A second opinion for rare cancers that cites its sources instead of guessing confidently.

When
2026
Role
Core engineer, team SecondLook
Built with
Python, FalkorDB, Cypher, CIViC, PubMed, AutoDock Vina
Source
healers-second-look/Athena ↗

See it

Athena case dashboard showing one flagged change and the live knowledge graph context
Case review: what changedRecording a newly observed LMNA::NTRK1 fusion in a demo case (congenital fibrosarcoma, stage IV) flags exactly one change, in red, above the live knowledge-graph context the evidence comes from and the case’s current state. Demo data, not a real patient.
Patient timeline with imaging, procedures, treatments and MRD markers
The patient timelineImaging, procedures, treatments and MRD results on one axis, with lab trends underneath. The demo uses a real published osteosarcoma reference dataset.

The problem

When an oncologist runs out of standard genomic leads, the tools left are search engines and chatbots. Neither tells you where a claim came from, and neither remembers what changed since the last tumour board.

Numbers, with dates

  • 191 / 217nodes and edges of curated CIViC sarcoma evidence in the graph (2,712 nodes once ClinicalTrials.gov is added)counted on live CIViC data, 28 Oct 2026
  • 1,388Python tests passing on our branch, plus 94 in the web clientrun 28 Oct 2026, after one real bug fix
  • 13 in 4tumour-board subsystems across four pull requests, all still openPRs #128 to #132, checked 28 Oct 2026

How it works

  1. 01 · case

    Start from the case

    Gene, variant, cancer type and history become a query, not a prompt.

  2. 02 · graph

    Pull a cited subgraph

    Cypher returns the evidence around the variant. Every edge carries the citation it came from.

  3. 03 · diff

    Diff what changed

    A new result is compared with the last state of the case; findings it contradicts are marked invalidated, not silently kept.

  4. 04 · board

    Hand it to the board

    The tumour board sees what is known, what changed and what is still unproven, in that order.

What I built

  • A knowledge graph over CIViC and PubMed evidence in FalkorDB, queried live for a gene, a variant and a cancer type.
  • Case memory that diffs new pathology, scans and labs against the last state and flags which earlier findings they invalidate.
  • A structural-biology pipeline (docking, AlphaFold and AlphaMissense lookups) for cases where the evidence graph comes back empty.

Team

Built with the healers-second-look team. The repository credits everyone who worked on it.