6  Guide 3 — AI Research Experiment Agent

6.1 “Virtual Research Assistant”

6.2 The problem

Researchers — especially in biology, medicine and agriculture — struggle with:

  • experimental design (controls, replicates, power),
  • data analysis (QC, normalisation, differential analysis),
  • coding (Python/R pipelines),
  • statistics (which test, multiple-testing correction, effect sizes),
  • reporting (figures, tables, a defensible Methods section).

A lab typically waits weeks for a bioinformatician or statistician.

6.3 The solution

An agent that manages the experiment from plan to report:

Agent Role Output
Experiment Planner Turns a one-line description into a design: groups, replicates, covariates, power notes, analysis plan design.json
Analysis (Bioinformatics/Data) Agent Runs QC, normalisation, differential analysis (e.g. control vs. treatment) using pandas/scipy tables + figures
ML Agent Trains baseline models (logistic regression, random forest), cross-validation, feature importance ml_results.json, figures
Statistics Agent Checks assumptions, corrects for multiple testing, flags bias/leakage, interprets effect sizes stats_report.md
Report Writer Produces the Methods + Results text and a reproducible notebook methods_results.md, analysis_notebook.ipynb

6.3.1 Example

Input: “I have RNA-seq data from heat-stressed rice (control vs heat, n=3 each).”

Output: design (control vs heat stress, n=3), QC summary, differential-expression table (log2FC, p, FDR), volcano plot, a random-forest classifier with CV accuracy, a statistics sanity report, and a Methods/Results section — plus a Jupyter notebook that reproduces everything.

6.4 Architecture — code does the math, the LLM does the reasoning

description ─► Experiment Planner (LLM, structured) ─► ExperimentDesign
                                                          │
 data.csv (or synthetic demo) ─► Analysis Engine (pure Python: pandas/scipy/sklearn/statsmodels)
                                                          │ numbers, tables, figures
                                                          ▼
                              Statistics Agent (LLM reads the numbers → assumptions, warnings)
                                                          │
                                                          ▼
                              Report Writer (LLM → Methods & Results) + Notebook generator (code)

Why this split matters: LLMs are unreliable at arithmetic and p-values. All statistics are computed by deterministic Python; the LLM only plans, interprets and writes. This keeps results reproducible and defensible.

6.5 Build steps

Step File What you learn
1 data.py Loading real CSVs or generating a synthetic omics-style dataset (genes × samples)
2 analysis.py QC, normalisation (log/CPM-like), t-tests per feature, BH-FDR, volcano plot, RF classifier with CV
3 agents.py Planner (structured), Statistics reviewer (structured), Report writer
4 notebook.py Generating a runnable .ipynb with nbformat
5 main.py CLI: --description, --data path.csv, --group-col, --demo

Run (demo data):

python agent_03_research_experiment\main.py --description "RNA-seq of heat-stressed rice, control vs heat, n=3" --demo

Run (your data):

python agent_03_research_experiment\main.py --description "..." --data mydata.csv --group-col condition

6.6 Product extensions

  • Domain plugins: RNA-seq (DESeq2 via rpy2 / pydeseq2), GWAS, metabolomics, field trials (ANOVA/mixed models), clinical trials (survival analysis), social science (regressions/SEM).
  • Sandboxed code execution so the agent can write and run custom analysis code safely (Docker / Claude code-execution tool).
  • Figure styling to journal specs; automatic figure legends.
  • Lab data connectors: LIMS, Google Sheets, S3, sequencing cores.
  • Power analysis & sample-size calculator at design time.

6.7 Market value

Customer Value Price
Biology / medical / agriculture labs Replace weeks of waiting for an analyst $50–100/month per researcher; lab $5k/year
Core facilities & bioinformatics units Handle 5× the projects $20k–50k/year
Universities Standardised, reproducible analysis for theses $50k–150k/year
Pharma / agri-biotech R&D Private deployment, custom pipelines, audit trails $150k–500k/year

This is the highest-value enterprise agent in the set because the alternative is hiring specialists.

6.8 How to sell it

  1. Free “analyse my CSV” demo (synthetic or uploaded, capped size).
  2. Core-facility partnership: give the facility a dashboard; they resell analysis to labs.
  3. Grant-budget line: position at $5k/year as a “bioinformatics support” line item in grants.
  4. Pharma/agri pilots: 3-month pilot on one dataset with a reproducibility report.
  5. Training: “Data analysis for biologists with AI agents” workshops (Codanics-style) → lab plans.

6.9 Risks & mitigations

Risk Mitigation
Wrong statistical test chosen Statistics Agent checks assumptions and explains alternatives; rule-based defaults per data type
Data leakage in ML Explicit CV pipeline; agent flags leakage patterns
Users trust numbers blindly Every result carries method, n, test, correction; report includes “limitations”
Data privacy (clinical) On-prem / VPC option; no raw data sent to the LLM — only summary statistics

6.10 90-day roadmap

  • Weeks 1–2: CLI with synthetic pipeline (this repo) → Weeks 3–6: CSV/Excel upload UI, figure export → Weeks 7–10: RNA-seq & field-trial plugins → Weeks 11–13: sandboxed custom code + notebook editor.

Codanics — Python ka Chilla: Build AI Agents

  • Course: https://codanics.com/courses/python-ka-chilla-build-ai-agents/
  • Website: https://codanics.com · YouTube: @codanics
  • WhatsApp channel: https://whatsapp.com/channel/0029Va7nRDq3QxRzGqaQvS3r
  • Complete free course: https://www.youtube.com/watch?v=UE3o-7KRNXs&list=PLaQmdklCcX40