results | index | references
4.1 What the app adds
For this case study, the app gave the target-level evidence for all seven genes in one place, in minutes, where gathering it by hand means visiting six or more databases. Every number can be traced to a named source and recomputed, and the app labels which values are measured facts and which are model predictions.
4.2 Limitations
The app confirms; it did not discover. FAK ranks #293 overall. A user reading only the top of the ranking would not reach it. FAK stands out when the Dependency column is examined, which is how the case study found it.
Some claims need data outside the app. Not in the app:
- ERBB2/3 induction after KRAS blockade (perturbation data)
- SRC-module survival modelling
- Single-cell co-expression
Co-pilot recall varies. It missed one relevant abstract in both runs. Its output needs human verification, as slide 67 already says.
One disease. This check covers pancreatic adenocarcinoma only. The same workflow is available for the other loaded diseases (Alzheimer’s disease, glioblastoma, exocrine pancreatic carcinoma), but they have not been checked here.
4.3 Points to reconcile
| # | Discrepancy | Recommendation |
|---|---|---|
| 1 | The Ranking Board ranks FAK #293 (small-molecule weighted score). The Evidence tab, knowledge graph and co-pilot rank it #773 using a different evidence score. | The two ranks should be labelled so users know which score they are reading. |
| 2 | Slide 40 reports FAK failing a network hub test on a 427-gene single-cell network. The app’s network measure uses the STRING disease graph, where FAK is at the 97.1th percentile. | These measure different things and should be described together. |
| 3 | The case study reports FAK dependency across 47 lines (slide 40) and 48 lines (slide 38). | The app does not display its line count; it should. |
5. Conclusion
Disease2Target independently holds evidence for most of the pancreatic case study’s target-level claims. The strongest agreement is on:
- the FAK dependency (within 0.03 of the lab’s own value),
- SRC as a network hub rather than a dependency,
- ERBB2/3 druggability, and
- KRAS as the backbone.
The app’s scores are fully traceable to their sources. Its role is confirming and speeding up a hypothesis, while new mechanistic findings still come from experimental data.
Comments