AI-Powered Enzyme Inhibition Experiment Design Guide

ISEF Category: Biochemistry

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

You can ruin an enzyme study with one missing control. That makes enzyme inhibition a great test case for AI, because the best answer is not a long answer, it is a careful one. Your project asks whether a retrieval-augmented LLM can draft those lean plans as well as an expert mentor. If it can, you get a new way to study how scientific design quality can be measured.

What Is It?

Enzyme inhibition happens when one molecule slows or blocks an enzyme, like putting a thumb on a spinning gear. In a real experiment, you do not just ask whether the enzyme slowed down. You also ask what control group, readout, and comparison will prove the slowdown came from the inhibitor and not from noise.

A retrieval-augmented LLM is a chatbot that checks a source bank before it answers. For this project, the agent does not just guess a lab plan. It pulls relevant papers, then proposes the smallest experiment that can test an inhibition idea. The benchmark asks a simple question: do its plans look like what an expert mentor would design for the same hypothesis?

Why This Is a Good Topic

This makes a strong science fair topic because you can test it with clear scoring rules, not gut feelings. The real-world link is drug discovery, enzyme diagnostics, and fast experiment planning. You can learn prompt design, literature search, benchmarking, and statistical comparison, all without needing a wet lab.

Research Questions

• How does adding retrieval from enzyme-kinetics reviews change the agreement between the agent's plan and expert mentor plans?
• What is the effect of forcing the agent to name controls first on the completeness of its design?
• Does the agent propose fewer unnecessary steps when you ask for a minimal experiment instead of a full protocol?
• To what extent do different inhibitor classes, such as competitive and noncompetitive cases, change the quality of the agent's plan?
• Which retrieval source, review articles, abstracts, or prior student examples, produces the closest match to expert designs?
• How does blind scoring by two raters change the ranking of prompt versions?

Basic Materials

• Laptop with reliable internet access.
• Python installed on your computer.
• Spreadsheet software for rubric scoring.
• A small set of enzyme inhibition papers or abstracts from PubMed.
• A rubric template for expert comparison.
• A note-taking app or lab notebook for tracking prompt versions.

Advanced Materials

• Access to a local open-source LLM or an API.
• A vector database or search index for retrieval.
• Python packages such as pandas, NumPy, and scikit-learn.
• A larger benchmark set built from review articles and mentor examples.
• Annotation software for blind human scoring.
• Access to a shared server or GPU workstation for repeat runs.

Software & Tools

• Python: Automates data cleaning, scoring, and analysis.
• Jupyter Notebook: Keeps prompt tests, model outputs, and notes in one place.
• Ollama: Runs local open-source models for repeatable prompt testing.
• Zotero: Organizes the papers and review articles behind each hypothesis.
• Google Sheets: Lets you score expert and model designs side by side.

Experiment Steps

1. Define what counts as a strong minimal enzyme-inhibition design, then turn that into a scoring rubric.
2. Build a held-out set of inhibition hypotheses and expert reference designs.
3. Decide what retrieval sources the agent will see, such as abstracts, reviews, or prior examples.
4. Compare prompt versions with the same benchmark so you can measure design quality, not luck.
5. Score each output for controls, feasibility, completeness, and match to the expert plan.
6. Analyze where the agent fails, then revise the retrieval setup or prompt constraints.

Common Pitfalls

• Mixing up the hypothesis and the readout, which makes the agent test the wrong thing.
• Letting retrieval pull in irrelevant papers, which pushes the model toward vague or noisy plans.
• Scoring only for similarity to expert plans, which rewards copying instead of good experimental logic.
• Using cases with different assay formats in the same benchmark, which makes the comparison unfair.
• Leaving out negative controls or baseline conditions, which makes a minimal plan unusable.

What Makes This Competitive

A strong version goes past "the model sounds smart." You compare it against a fixed expert set, score more than one quality metric, and test whether retrieval actually helps. The best entries also break failures down by hypothesis type, assay type, and inhibitor class. That turns a neat AI demo into a study about how scientific design quality can be measured.

Project Variations

• Switch from mentor-made reference plans to published protocol fragments and see whether the benchmark still holds.
• Test whether the agent performs better on competitive inhibitor cases than on noncompetitive inhibitor cases.
• Compare text-only retrieval with retrieval plus short paper summaries from review articles.

Learn More

• PubMed: Search for review articles on enzyme inhibition, assay design, and enzyme kinetics.
• NIH NCBI Bookshelf: Read free textbook chapters on enzymes, metabolism, and inhibition.
• MIT OpenCourseWare: Find free biochemistry lectures and problem sets on enzyme kinetics.
• PubChem: Look up inhibitor structures, target annotations, and bioactivity links.
• RCSB PDB: Explore enzyme structures and active-site geometry.