Skip to content

About

Website for browsing reviews of GO-CAMs.

This contains an initial pass of reviews of GO-CAMs with >2 causal edges, performed by the aurelian agentic framework using claude 3.7.

THIS IS A DEMO FIRST PASS AND NEEDS FURTHER TUNING

  • Open the "Reviews" tab on the left to see all reviews
  • The guidelines are provided for reference
  • simple search in top right

Known Bugs

False positive warning about 'has input'

When I fed the GO-CAMs to the agent, the JSON was missing 'has input' to uniprot IDs (only CHEBI IDs includes). This means that the agent will sometimes mention that there is a missing input - this is not a hallucination, it's an upstream data bug.

Examples

Accidental inclusion of non-causal models

For simplicity, we include all GO-CAMs with >2 causal edges. This is generally sufficient to exclude gene-centric models, however, a few may slip through due to spurious causal annotations. The reviews for these are typically very negative, which is a nice confirmation things are working.

For example,

The review for this -- as expected complains that there needs to be a lot more causal connections, and overall this scores very low (2).

In addition there are some models that may be in-between, e.g.

Figures

A drawing agent was instructed to create an SVG of the GO-CAM (input is JSON). A review agent checks this to make sure the image is clear. It is meant to tell the draw agent to draw again if things are misaligned or text crosses boxes etc but this part does not work so well, and some images have floating chunks or incorrect legends, but overall the drawing agent seems to understand both the visual grammar of pathway figures and the intent of the model.

I deliberately left this quite unconstrained to see what kinds of styles the model would come up with, but we can easily constrain this to a house style, or give it a predefined set of components.

See:

Selected: