Final Project — Exploring and Modeling Life Science Data with AI Agents

⚠️ Note: This is a draft project plan. Details (including deadlines, deliverables, and evaluation criteria) may change. Updates will be announced on the course website and in class.

Welcome to the final project of DDLS 2025.
This is your opportunity to demonstrate what you have learned throughout the course:

  • Working with real biological datasets.
  • Using AI agents (e.g., Gemini-CLI, ChatGPT, Perplexity) to assist in exploration, coding, and analysis.
  • Designing and evaluating models for meaningful scientific tasks.
  • Making your work more accessible and reusable by wrapping it as a web application or MCP tool.
  • Critically reflecting on AI’s role in scientific discovery.
  • Practicing FAIR and open science principles in your work.

Working Arrangements

Individual vs. Pair Work:

  • You may choose to work individually or in pairs (teams of 2 students).
  • If working in pairs: Projects can complement each other and students can collaborate on data exploration, model development, and technical implementation.
  • Individual accountability: Each student must complete all three phases of the project (exploration, modeling, and accessibility wrapping) and demonstrate individual mastery of the learning objectives.
  • Individual submissions: Each student must submit their own project report that clearly highlights their specific contributions, analysis, and reflections.
  • Pair coordination: If working in pairs, coordinate your project plans during Step 1 to ensure complementary approaches (e.g., different models, evaluation metrics, or accessibility implementations on related datasets or scientific questions).

The project is divided into two steps:

Step 1 — Project Plan & Validation (Due: Oct 6, 2025)

Before you begin, you must submit a project plan for validation.
This ensures that your project is relevant, novel, feasible, and not too close to what has already been done in the computer labs.

Project Choice Criteria

Your project plan will be approved only if it is:

  • Not too easy (must go beyond trivial workflows or toy problems).
  • Not too close to computer lab exercises.
  • Not too hard / compute-heavy (should run on reasonable hardware, e.g., Google Colab).
  • Technically sound (analysis steps and methods make sense).
  • Feasible (can be completed in ~2–3 weeks).
  • Relevant to a real scientific problem.
  • Shows novelty (not just reproducing a tutorial, but bringing something creative and realistic).

What to Submit (max 3 pages + appendix)

  1. Scientific Question & Dataset (combined)

    • Start with a rough idea or problem you are interested in.
    • Use a GenAI tool (Gemini, Perplexity, ChatGPT, etc.) to brainstorm possible scientific questions and relevant datasets, use “deep research” features to do broad investigation to understand the landscape.
    • Arrive at a refined dataset + scientific question to work on.
    • You will need to also investigate:
      • What datasets are available and accessible?
      • What is their license/ethics status?
      • Why is the problem scientifically relevant?
      • What kind of task (classification, clustering, regression, etc.) is appropriate?
    • Clearly document the dataset name, source (URL/reference), size, and chosen subset (if needed).
    • Appendix requirement: attach the AI “deep research” transcript/report showing your brainstorming process, with both your inputs and the agent’s outputs.

  2. Exploration Goals

    • At least 3 dataset exploration questions that are important to understand before working on the model training.
    • Examples: data size, types, value range, missing values, distributions, class balance, correlations.

  3. Proposed Model & Evaluation

    • Planned baseline model, and a few potential improvements.
    • Planned evaluation metric(s).
    • What “good performance” would mean for your task.

  4. Accessibility Plan

    • How will you make your final workflow accessible to others?
      You can choose between:
      • Option A: Wrapping your pipeline as a web application to allow user load data, run inference, display inference and evaluation results.
      • Option B: Wrapping the final workflow into a MCP toolset with tools such as:
        • Load data
        • Run model inference
        • Show results
        • Evaluate model
    • Important: Your web app or MCP tools, as well as your GitHub repository, must be made public.
    • These outcomes will be showcased on the course website and used as demonstrations for SciLifeLab researchers and beyond.

  5. Feasibility Check

    • Justify that the project is doable in the timeframe.
    • State expected runtime.
    • Mention potential risks (large dataset, convergence issues).

  6. Submit the plan before Oct 6, 2025 to get feedback from the teachers

Submit Your Document

Step 1 Evaluation

  • Pass → You may proceed to Step 2.
  • Revise → You must update your plan before starting.
    Plans may be rejected if the dataset/task is too trivial, too close to a lab, overly ambitious, license-restricted, or missing the AI deep research appendix.

Step 2 — Final Project (Oct 6–29, 2025)

Once your plan is approved, you will complete the project in three phases:

Phase I — Dataset Exploration

  • Use Python (optionally with AI agent support) to explore the dataset.
  • Summarize distributions, imbalances, and other key features.
  • Split the dataset into train, validation, and test sets.
  • Deliverable: Data exploration visualizations, showing distribution for each dataset split.

Phase II — Model Training, Evaluation & Improvement

  • Train a baseline model (with or without AI agent help).
  • Evaluate on validation/test data.
  • Attempt at least one improvement step (better preprocessing, tuning, or different model).
  • Deliverable: A working model with measurable performance and evidence of improvement attempts.

Phase III — Accessibility Wrapping

  • After you have a working pipeline, make it accessible to others by choosing one of two options:
    • Option A: Build a simple web application wrapping your workflow (similar to computer lab 2).
    • Option B: Create a minimal MCP toolset (load data, run inference, show results, evaluate model, similar to computer lab 4, 5, 6).
  • Deliverable: A working wrapper (web app or MCP) with basic documentation.
  • Outcome: Your wrapper and GitHub repository will be publicly available and showcased on the course website as demonstrations to SciLifeLab researchers and external audiences.

FAIR Data and Open Science Recommendations

To promote the FAIR principles (Findable, Accessible, Interoperable, Reusable) and support open science:

  • Prefer public datasets. If using a private dataset, explain how it could be made available in the future.
  • You must create a public GitHub repository for your project.
    • Add the topic ddls-course-2025 to make your project findable (GitHub Topics Guide).
    • Include a README file with an overview and usage instructions.
    • Document your code properly.
    • Add a permissive license (e.g., MIT license) to allow reuse.
    • If a public dataset was used, provide a script or instructions to download it.
    • Make your project reproducible; upload trained model weights to GitHub Releases or Zenodo if possible.
    • Consider publishing your report (e.g., on Zenodo).
  • At the end of the project, we will try to deploy your web app and mcp tools to make it more accessible, e.g via the course website.

Final Deliverables

You will submit:

  1. Final Project Report (max 5 pages main text, unlimited appendices)
    The report should include:

    • Abstract (≤100 words): problem, method, results.
    • Background and Motivation: why you chose this dataset/task.
    • Dataset Summary: data source, preprocessing, splits, distributions.
    • Method Description: workflow, models, evaluation metrics.
    • Results: figures, tables, performance metrics vs. baseline.
    • Conclusion & Discussion: findings, limitations, future directions.
    • Data and Code Availability: links to dataset and repo (per FAIR guidelines).
    • Acknowledgments: contributions, support, and note on GenAI tools used.
    • References: relevant literature.
    • Appendices: AI deep research log (mandatory), prompts, agent transcripts, extra figures.

  2. Accessibility Wrapper & Repository

    • Either a simple web app or a minimal MCP toolset wrapping your final workflow.
    • A public GitHub repository containing your code, documentation, and reproducibility instructions.
    • Published outcome: These deliverables will be linked on the course website and demonstrated to SciLifeLab researchers and beyond.

  3. Accessibility Demo Video (3–5 minutes)
    Choose the appropriate demo format based on your accessibility option:

    • Option A (Web Application): Screen recording showing your web app in action, demonstrating all available features and that it works as intended.
    • Option B (MCP Toolset): Screen recording (or asciinema) showing an AI agent (e.g., Gemini-CLI) using your MCP tools for dataset exploration, model inference, evaluation, and result visualization.
    • Note: You may fast-forward or cut long waiting times (e.g., model training/inference) in your final submission video to match the length requirement while still demonstrating full functionality.

Submit Your Final Project

Pass Criteria

✅ Project plan includes scientific question + dataset refined through AI deep research.
✅ Data exploration performed and documented.
✅ A baseline model is trained and evaluated.
✅ At least one attempt to improve performance is demonstrated.
✅ A working wrapper (web app or minimal MCP tool) is provided and made public.
✅ A 3–5 minute demo video showing the accessibility wrapper in action (web app features or agent using MCP tools).
✅ A public GitHub repository with README, license, and reproducibility instructions.
✅ Student-authored prompts for agent interactions are included.
✅ Report includes dataset analysis, model justification, evaluation, and accessibility plan.

Stretch Goals (for high grades)

⭐ Compare multiple model types.
⭐ Provide advanced evaluation (ROC curves, feature importance, error analysis).
⭐ Strong critical reflection on AI use and limitations.
⭐ Clear reproducibility (clean Colab, stable results).
⭐ Public release of dataset/code/report (GitHub/Zenodo).

Timeline

  • Oct 6, 2025 — Project plan + AI deep research log due.
  • Oct 8, 2025 (13:00–14:00) — Consultation session 1.
  • Oct 20, 2025 (13:00–14:00) — Consultation session 2.
  • Oct 29, 2025 (23:59) — Final project submission (report, accessibility demo video, wrapper, repo).
  • Oct 31, 2025 (9:00-12:00) — Oral presentations (Master’s students only).

Oral Presentations

Who must present:

  • Master’s students: Oral presentations are mandatory.
  • Other course participants: You are welcome to present as well, as long as time allows. Please sign up using the following link: https://forms.gle/zZWUaG2uhH3fiB9q7

Who can attend:

  • All course participants are welcome to join and observe the presentations.

Format:

  • Duration: 10 minutes for your presentation, followed by a 5-minute Q&A session.
  • Platform: The presentations will take place online via Zoom (see the calendar event for details).
  • Demo (optional): A live or recorded demo of your accessibility choice (web app or MCP toolset) is not required, but you are welcome to include it. If you choose to demo, your presentation time can be extended by 3–5 minutes.

What to present:

  • Overview of your scientific question and dataset.
  • Key findings from your data exploration.
  • Your model approach, evaluation, and improvement attempts.
  • Your accessibility wrapper (web app or MCP toolset).
  • Reflections on AI use and lessons learned.

Note on GenAI Tools

You are encouraged to use ChatGPT, Gemini, Perplexity, or similar for brainstorming, coding assistance, and report writing.
However:

  • You must verify all outputs and remain responsible for correctness.
  • You must document how you used these tools (in the report and acknowledgments).
  • You must include your AI deep research log in the appendix (mandatory).
  • You must include agent transcripts and your authored prompts.

Use AI responsibly — see AI in Life Sciences: Power with Responsibility.

Last updated on Oct 30, 2025

Edit this page