Designing Biomedical Research Protocols
Given a research topic, produce a structured protocol skeleton:
- Research question (PICO/FINER-style, falsifiable)
- Study design (in vitro / in vivo / cohort / cross-sectional / RCT)
- Key variables (exposure, outcome, confounders)
- Methods & endpoints (assays, sequencing, imaging, statistics)
- Risks & limitations (bias, off-target effects, translational gap)
Example invocation: "Design a study on gut microbiome composition and anti-PD-1 response in melanoma patients" → output below follows this exact structure.
Progress:
- [ ] Clarify the biological question and therapeutic/diagnostic context
- [ ] Define study design and justify choice (feasibility, ethics, cost)
- [ ] Specify population/model system (human cohort, mouse model, cell line, in silico)
- [ ] Define primary/secondary endpoints and measurement methods
- [ ] Identify confounders and control strategy
- [ ] Specify statistical/analytical plan (power, correction for multiple testing)
- [ ] List limitations, translational caveats, and ethical considerations
- [ ] Summarize into protocol document
Step 1 — Clarify the question. Convert a broad topic into a specific, testable hypothesis. Example: "microbiome and immunotherapy" → "Does relative abundance of Akkermansia muciniphila at baseline predict response to anti-PD-1 therapy in NSCLC patients?"
Step 2 — Choose design matched to question type:
- Mechanism (e.g., CRISPR off-target) → in vitro + in vivo, whole-genome/exome off-target sequencing (GUIDE-seq, CIRCLE-seq)
- Association (e.g., epigenetic aging) → longitudinal cohort with repeated methylation sampling (e.g., Horvath/GrimAge clocks)
- Device development (e.g., graphene biosensor) → iterative engineering + analytical validation (sensitivity, specificity, LOD) → clinical correlation study
- Toxicology (e.g., microplastics/endocrine) → dose-response animal study + human biomonitoring correlation
Step 3 — Define model/population with inclusion/exclusion criteria, sample size via power calculation, and relevant controls (germ-free mice for microbiome work, isogenic cell lines for CRISPR).
Step 4 — Endpoints: separate primary (e.g., objective response rate, off-target indel frequency, glucose correlation coefficient tear vs. blood) from secondary (toxicity, cost, durability).
Step 5 — Confounders: diet/antibiotics (microbiome), age/sex/smoking (epigenetics), batch effects (sequencing), delivery vehicle immunogenicity (exosomes).
Step 6 — Stats: specify test, correction (FDR/Bonferroni for omics), and power/effect size assumptions up front — not post hoc.
Step 7 — Limitations: always state species-translation gap, causality vs. correlation, generalizability of cohort.
Example 1: Input: "Оценка долгосрочных последствий применения CRISPR-Cas9 на нецелевые участки генома" Output:
- Question: Do AAV-delivered Cas9 constructs produce detectable off-target edits in long-term (>12 month) treated hematopoietic stem cells in vivo?
- Design: Longitudinal in vivo mouse model + validation in human HSC xenografts
- Methods: GUIDE-seq/CIRCLE-seq at baseline; whole-genome sequencing at 1, 6, 12 months; clonal hematopoiesis assay (CHIP panel)
- Confounders: guide RNA specificity variance, PAM site density, vector dose
- Stats: mixed-effects model for repeated measures, FDR correction across genomic loci
- Limitations: mouse-to-human extrapolation, limited long-term human data, detection sensitivity floor of sequencing (~0.1% VAF)
Example 2: Input: "Разработка биосенсоров на основе графена для мониторинга глюкозы в слезной жидкости" Output:
- Question: Does a graphene field-effect transistor (GFET) sensor achieve clinically actionable correlation (r>0.85) between tear and blood glucose across physiological range (2–20 mmol/L)?
- Design: Phase 1 analytical validation (LOD, selectivity vs. lactate/ascorbate interference) → Phase 2 human pilot (paired tear/blood sampling, n≥30)
- Endpoints: sensitivity, specificity, Clarke Error Grid analysis
- Confounders: tear volume variability, evaporation, circadian glucose lag (~5–15 min tear-blood delay)
- Limitations: contact-lens form factor biocompatibility, long-term biofouling of graphene surface
- Always state the hypothesis as falsifiable before choosing methods.
- Match statistical power calculations to the primary endpoint only.
- For omics/sequencing studies, pre-register multiple-testing correction method.
- Distinguish mechanistic claims (needs intervention/knockout) from associative claims (cohort/correlation suffices).
- For translational topics (mouse → human), explicitly flag the translation gap in limitations, not as an afterthought.
- For novel delivery systems (exosomes, nanomaterials), include stability/scalability/immunogenicity as go/no-go criteria early.
- Don't propose a cohort/correlational design when the question is mechanistic (e.g., "does CRISPR cause off-target effects" needs functional validation, not just sequencing association).
- Don't ignore batch effects and confounders in microbiome/epigenetic studies — these are the most common cause of irreproducibility.
- Don't set endpoints after the fact; this invites p-hacking in omics-heavy designs.
- Don't present preclinical (mouse/cell) results as directly generalizable to humans without caveats.
- Don't underestimate sample size needs for microbiome/epigenome studies — high inter-individual variability requires larger cohorts than typical clinical trials.