## Inspiration

For researchers and scientists, it takes considerable time to research just one particular disease. They often need to go through the introduction, abstract, results, and more. This is time-consuming, and sometimes the papers they find aren't even useful. Our solution addresses this directly: why not reduce the time spent on this stage of research ? We had the idea of building an AI tool that could gather several reliable papers together and extract the key information about a disease automatically. This helps scientists, researchers, and even students save time on the research phase. So, they can invest that time into analysis and the next stage of their work instead. Users simply enter a disease name and indicate how many papers they'd like included, and the AI handles the overview. Importantly, all information is drawn directly from real sources, it is not generated or made up by the AI.

## What it does

An AI tool is used to summarize key findings across multiple sources for any searched disease. The summary includes key information such as the disease's mechanism, treatment approaches, genetic implications, and the differences between papers. Including gaps in current knowledge.

*## How we built it *

AI was used throughout both the planning and the technical set-up.

First, we set up a connection to PubMed's API, the system that let our code request real research papers and receive them back in a structured format (JSON) rather than as a webpage. Second, we created an account with Claude's API. This was necessary because Claude.ai (the normal chat website) is built for humans typing messages. But we needed our own Python code to send messages and receive replies automatically, without a person doing the typing. The API is a separate doorway into the same AI, built specifically for software to use. To make that connection work, we generated an API key. This acts like a password that authorises our specific code to communicate with Claude's servers. From there, we used AI in two ways. First, to extract information (gene, pathway, treatment, open questions) from each individual paper. Second, to synthesise those extracted facts across multiple papers into one summary. This includes identifying where papers disagreed or left questions unanswered. In the remaining days, we built a simple web interface using Streamlit and tested the app for errors. We also used Claude as a learning guide throughout this process, since we were both completely new to AI. Rather than treating this as something to just get working, we used it as a learning opportunity. We came away having actually understood how each piece works.

## Challenges we ran into

Honestly, we didn't run into too many major challenges, but we did hit two real technical issues. The first was that clicking the "search" button twice in quick succession caused an error. Since it sent two overlapping requests at once. The second was that, searching a very broad disease (e.g cancer), returned too many sources for the AI to process in one go, which caused timeout errors. We had to go back into the code and add extra handling for both cases. Just to prevent double-submissions and adjusting how many sources could be processed at once to make the system more reliable.

## Accomplishments that we're proud of

This is our very first project, and we're proud that we saw it through from start to finish. We actually built something that addresses a real-life problem ! Beyond just finishing it, we're proud that it works.

## What we learned

We learned an enormous amount personally. We might not be able to write an entire codebase from scratch yet, but we now understand how API requests work. We're also much more comfortable with terms like "import" and "JSON format" that used to be unfamiliar. We became confident using the terminal to run commands, including launching Streamlit locally. We learned how the Claude Console works and how Python communicates with it to send and receive instructions. Coming into this with no AI experience, we leave having actually understood how each piece of the system works, not just copied codes that ran.

## What's next for Disease Mechanism Explorer

We would like to improve the design and add more features going forward. As a biomedical science student, I would genuinely like to keep using this for my own research, since I've found it useful firsthand. We would like to make it publicly available for others to use too, but since every search uses Claude API credits, that's a cost consideration we need to solve before opening it up more broadly. Maybe by looking into more sustainable hosting or usage limits.