# Routing benchmark

Ground truth is CIViC: an alteration, a therapy, and the statement that the therapy works. A routing decision counts as correct when the nominated target's approved drugs include that therapy. **The control that decides it is "route every gene to itself"** — the naive rule the whole synthetic-lethal layer exists to beat.

- 60 distinct alteration/therapy claims at evidence tier A or B
- 43 scorable; 17 excluded because the therapy is not in this library's drug table at all, which measures library coverage rather than routing

| Ranking | Top-1 accuracy |
| --- | --- |
| **OncoForge routing** | **0.837** |
| OncoForge, any proposed hypothesis | 0.837 |
| route the gene to itself | 0.395 |
| always propose PARP1 | 0.256 |
| random target | 0.070 |

| Relation used | n | Correct |
| --- | --- | --- |
| direct | 22 | 22 (100%) |
| pathway | 10 | 9 (90%) |
| synthetic_lethal | 5 | 5 (100%) |

**Verdict: beats every trivial control.** top-1 accuracy 0.837 against 0.395 for 'route the gene to itself'. The routing layer is doing something a lookup table does not.

### Where it goes wrong

| Alteration | Expected target | Proposed | Therapy |
| --- | --- | --- | --- |
| BARD1 Mutation | PARP1 | (nothing proposed) | Olaparib |
| CHEK1 Mutation | PARP1 | (nothing proposed) | Olaparib |
| CHEK2 mutation | PARP1 | (nothing proposed) | Olaparib |
| RAD51B Mutation | PARP1 | (nothing proposed) | Olaparib |
| RAD51D Mutation | PARP1 | (nothing proposed) | Olaparib |
| RAD54L Mutation | PARP1 | (nothing proposed) | Olaparib |
| v::ROS1 Fusion | NTRK1 | ALK | Entrectinib |