andrew7shen/ar_science

ar_science is the code release accompanying a research paper about a method called Analogical Reasoning, or AR, which helps generate creative approaches to scientific problems. Given a problem, a language model runs two steps: first it extracts analogies to the problem from other domains along with explicit mappings between the objects involved, then it searches for real, existing solutions within each analogous domain. The paper tested this on biomedical problems, but the released prompts and skill are described as domain neutral, meaning they can be applied to any scientific problem. There are two ways to run it. The first is as a Claude Skill inside Claude Code: the skill folder is copied into the user level skills directory, after which typing a slash command followed by a problem description triggers the two step process. Adding a save flag stores the full output as JSON, including the analogies, their mappings and rationales, and the solutions found along with supporting details like relevance and citations. The second way is running a Python script directly from the command line, which defaults to using Claude but can be switched to OpenAI or Gemini by changing a constant near the top of the file. This path requires setting up API keys in a local environment file that is kept out of version control. The repository also includes the dataset used to evaluate AR, stored as a JSON file with its own separate documentation, along with the full evaluation pipeline and case studies from the paper. These cover topics like predicting the effects of biological perturbations, modeling interactions between brain regions, predicting properties of oligonucleotides, and analyzing cell to cell communication, each implemented as its own case study folder within the repository.