Bioinformatics Studio

how to use

1. install uv if needed: https://docs.astral.sh/uv/getting-started/installation/
2. install python environment: uv sync
3. Fill in API keys in .env file
4. Run streamlit via uv run streamlit run main.py

choices made

• data is not sent to LLM, just the description of the data
• LLM is instructed to generate code, code is run in user's python environment with exec, which is not the best practice but is the only way I know how to get the code to run

todo

• handle multiple data files for one analysis: data.csv, sample_info.csv
• files workspace should look more like a regular file browser ex) NerdTree, Finder...
• fix right pane output -- currently uses streamlit's st.pyplot, which captures latest output; need to save output to disk in agreed layout (folder structure) so results are portable
• implement chat / code / output history
  • everything should be saved to disk
  • streamlit should not have in-memory state wrt data / code / output, it should be pure function of workspace
  • use case: user wants to modify data, they can do it themselves or ask LLM to do it and save it
• option to use current code output as context, or start from scratch: currently every generation starts from scratch
• error handling: when visualization fails (usually due to error in code), agent should automatically try to rerun, at least a couple times
• new library -- handle PyComplexHeatmap, which generates nicer heatmaps but requires
• currently, code executor just returns plt.gcf() and any print statements. Is there a good way to get one agent to write code to save figures to disk, and another agent to consolidate output to a report?
• include documentation for most common figures (heatmap, clustermap) as context
• add mermaid plot in explanation tab
• 사용자가 작성한 채팅에 대해 ‘ex) heatmap 그려줘 등‘ AI가 파악하여 heatmap에 필요한 정보들을 사용자에게 힌트 목록을 제시해준다.
• 참고할 레퍼런스(그래프나 그림, 글?)를 삽입하여 정확도를 높인다. 보통 연구자들은 논문에 구현된 figure를 따라 그리기 때문
• implement letter by letter writing animation for generated code
  • enable streaming in API calls
  • set up server endpoint to stream response
  • create frontend component to rcv stream and update display char by char
  • css transitions
• voice input -- talk to LLM