Session 03 · No-Code & Agentic AI for Life Sciences
Agentic
Literature Review
Biomedical discovery · Systematic workflows · Multi-paper synthesis
Md. Jubayer Hossain · Founder & CEO
DeepBio Limited · DeepBio Academy
No Code & Agentic AI for Life Sciences · Session 03 · June 2026
Conceptual Foundation
What are AI Agents?
An AI agent is an autonomous system that uses an LLM as its "brain" to complete multi-step goals.
🔄 The Agentic Loop
- Perception: Reading literature, data, or files.
- Plan: Deciding which papers to read next.
- Act: Searching PubMed, extracting data.
- Memory: Remembering findings across steps.
🤖 Chat vs. Agent
Chat answers a question. An agent executes a workflow (e.g., "Find and summarize all papers on XY since 2024").
🧠 Persistent Context
Agents maintain state, allowing them to iterate on their own findings without starting from zero.
Search Tools
Consensus for Systematic Review
Consensus is an AI-powered search engine that extracts findings directly from peer-reviewed literature.
- Evidence Synthesis: Aggregates findings across multiple studies.
- Consensus Meter: Visualizes the scientific agreement on a query.
- Quality Filters: Filter by study type (RCT, Meta-analysis), sample size, and journal impact.
- Direct Citations: Every claim is linked to a DOI.
💡 Use Case
"Does metformin improve longevity in non-diabetic adults?" Consensus scans 200M+ papers to give you the current "Meter."
🔍 Better than Google Scholar?
Scholar gives you a list of links. Consensus gives you the answers inside the links.
Automated Workflows
PubMed Monitor → Claude Summarizer
Building a custom surveillance system for your research niche.
1. Monitor
Use a tool (or script) to watch PubMed RSS feeds for specific keywords (e.g., "Liquid Biopsy Oncology").
2. Filter
Automatically pass new abstracts to Claude with a system prompt to filter for high-impact findings.
3. Deliver
Get a weekly "Digest" email or Slack notification with 1-paragraph summaries and DOIs.
Prompt: Act as a senior oncologist. Review the following abstract. If the study involves a Phase III trial or a novel biomarker with p < 0.001, provide a 3-sentence summary. Otherwise, return "Skip".
Knowledge Management
Claude Projects as Literature Assistant
Claude Projects allow you to upload up to 500MB of PDFs to create a specialized knowledge base.
- Persistent Context: The model "remembers" your library in every new chat within the project.
- Custom Instructions: Set a project-level persona (e.g., "Help me find contradictions in these papers").
- Fast Retrieval: Ask questions across 20+ PDFs simultaneously.
📦 The "Mini-Review" Workflow
Upload 10 key papers. Ask: "Generate a table comparing the study designs, sample sizes, and primary outcomes of all these papers."
⚠️ Limitation
Claude only sees the text you provide. It won't search the web unless you use a connected tool or the "Search" feature.
Advanced Synthesis
NotebookLM for Multi-Paper Synthesis
Google's specialized research tool that uses Gemini 1.5 Pro's massive context window.
- Source-Grounded: Every answer includes citations that highlight the exact passage in your PDFs.
- Audio Overview: Generate a "Podcast" summary of your research library (Deep Dive).
- Auto-Notebook: Automatically suggests FAQs, Study Guides, and Tables of Contents for your papers.
🌟 Key Strength: Grounding
NotebookLM is significantly less likely to hallucinate because it is strictly forced to cite your uploaded sources.
🎧 The "Commute" Summary
Turn a 50-page systematic review into a 10-minute conversational audio summary to listen to on the go.
Quality Control
Critical Appraisal with AI Agents
🔍 Checking the Methodology
"Identify the blinding strategy used in this study. If not stated, mark as 'High Risk of Bias' according to Cochrane standards."
📊 Statistical Audit
"Look for any discrepancies between the p-values reported in the Abstract and the Results table. Highlight any inconsistencies."
The "Skeptic" Persona
Don't ask AI to summarize; ask it to critique. "Act as a critical peer reviewer. What are the top 3 weaknesses in the discussion section of this paper?"
Framework-Based Toolkit
AI Prompts for Literature Review
Each prompt anchored to an established methodology — students practice the thinking, not just the prompting. AI is scaffolding, not a shortcut.
1 Scope question
2 Search strategy
3 Screen & select
4 Critical appraisal
5 Extract & matrix
6 Thematic synthesis
7 Gap analysis
8 Theory mapping
9 Write narrative
10 Self-audit
⚠ Non-Negotiable Ground Rule
Verify Every Citation
AI models fabricate citations, authors, journal names, even plausible-looking DOIs.
✅ The Rule
Every reference must be independently verified in a real database before it appears in any written work.
🔎 Where to Verify
Google Scholar · Scopus · Web of Science · PubMed. Treat AI output as a draft to verify, never a source to cite directly.
1
Scoping the Research Question
PICO(S/T) — clinical
SPIDER — qualitative
SPICE — service/policy
FINER — feasibility
🧠 How to think
- PICO = Population · Intervention · Comparison · Outcome.
- Ask for narrow / moderate / broad — then judge the scope yourself.
- Run a FINER gut-check before committing.
Prompt 1.1 · PICO
Structure a focused PICO question, then give 3 phrasings — narrow / moderate / broad — and flag any too broad for one review.
Populationadults with type-2 diabetes
Interventionintermittent fasting
Comparisonstandard calorie diet
OutcomeHbA1c reduction
You get → three scoped questions + a scope-creep warning.
2
Search Strategy Design
Boolean AND / OR / NOT
Controlled vocab (MeSH)
PRISMA-S reporting
🧠 How to think
- Expand every concept: synonyms, variants, MeSH terms.
- Combine with Boolean AND / OR / NOT.
- Audit like a peer reviewer — gaps, over-restriction, reproducibility.
Prompt 2.1 · Keywords
For each concept in my question, build the search vocabulary and a ready-to-paste Boolean string.
Conceptintermittent fasting
DatabasePubMed → adapt for Scopus
- Synonyms, related terms & spelling variants
- Controlled vocabulary (e.g., MeSH terms)
- Boolean AND / OR / NOT string — flag ambiguous terms
You get → a database-ready search string + ambiguity flags.
3
Screening & Selection
PRISMA 2020 flow: identify → screen → eligibility → include
🧠 How to think
- Draft inclusion / exclusion criteria as a defensible table.
- Cover population, design, pub type, dates, language, outcomes.
- Use cautiously: students screen first, then compare.
Prompt 3.2 · Screening (cautious)
Judge whether one study meets my criteria — give a verdict and a reason per criterion.
Title + abstractpaste the record
CriteriaRCTs · adults · 2015+ · English
- Verdict: Include / Exclude / Unclear
- Justify against each criterion individually
- Flag anything you can't judge from the abstract
You get → a per-criterion verdict — never a final decision.
4
Critical Appraisal of Methodology
CASP checklists
GRADE — certainty
Cochrane Risk of Bias
🧠 How to think
- Apply the design-specific CASP checklist, item by item.
- Scan bias: selection, measurement, confounding, publication, reporting.
- Separate bias authors acknowledge from bias they don't.
Prompt 4.1 · CASP critique
Apply the design-specific CASP checklist and judge rigor — do not summarize.
Methods sectionpaste the text
DesignRCT / cohort / qualitative
- Go through each checklist item in turn
- Rate: adequate / partial / not addressed
- One-line justification per item
You get → an item-by-item rigor verdict, not a summary.
5
Data Extraction & Synthesis Matrix
Garrard's matrix method
🧠 How to think
- Extract into a table — explicit content only, no inference.
- Cols: Author/Year, Design, Sample, Variables, Findings, Limitations.
- Cross-tab themes × studies → convergence vs. contradiction.
Prompt 5.2 · Pattern matrix
Turn my extraction table into a themes × studies matrix and read the patterns.
Extraction tablepaste — rows = studies
- Themes with strong convergence across studies
- Themes with contradictory findings — name the studies
- Themes in only 1–2 studies (under-researched)
You get → a themes × studies map of agreement & gaps.
6
Thematic / Narrative Synthesis
Thomas & Harden — descriptive → analytical themes
🧠 How to think
- Step 1: line-by-line descriptive codes, close to findings.
- Step 2: group codes into descriptive themes.
- Step 3: analytical themes that go beyond the studies.
Prompt 6.1 · Inductive themes
Build themes from the ground up — show your reasoning at each step.
Findingssummaries from N studies
Questionmy review question
- Step 1 → line-by-line descriptive codes
- Step 2 → group codes into descriptive themes
- Step 3 → analytical themes answering my question
You get → a coded trail from raw findings to new themes.
7
Gap Analysis & Future Questions
Robinson et al. — PICO-based gap mapping
🧠 How to think
- Population gaps — groups understudied.
- Intervention / Outcome gaps — variants & measures untested.
- Methodological gaps — designs underused.
Prompt 7.1 · Find the gap
Run a PICO-based gap analysis and turn each gap into a researchable question.
Synthesiswhat's known so far
Topicintermittent fasting & T2D
- Population / intervention / outcome / method gaps
- Why each gap matters
- One researchable question per gap
You get → mapped gaps, each with a researchable question.
8
Theoretical / Conceptual Mapping
Theory-family mapping across the literature
🧠 How to think
- Group studies by theoretical lineage / family.
- Identify where theories compete vs. complement.
- Judge if an integrated framework is justified — sketch it.
Prompt 8.1 · Map theories
Map the theories used across my studies and judge whether to integrate them.
Studies + theory"Smith 2019 – SCT; Lee 2021 – TAM"
- Group studies by theoretical lineage / family
- Where do theories compete vs. complement?
- Is an integrated framework justified? Sketch it
You get → theory families + a candidate integrated framework.
9
Writing the Review Narrative
Ask for structure & critique — not finished prose
🧠 How to think
- Get a paragraph-by-paragraph order for the synthesis.
- Decide where contradictions are addressed.
- The analytical writing stays yours.
Prompt 9.1 · Structure only
Plan the synthesis section — give an outline, not prose. I write the analysis.
Themesfrom Stage 6
Gapsfrom Stage 7
- Paragraph-by-paragraph theme order
- Where contradictions are addressed
- How the gap leads into my research question
You get → a section outline to write against — not the prose.
10
Self-Audit Before Submission
Integrity & rigor check
🧠 How to think
- Catch claims that don't match what the source found.
- Flag overgeneralizations beyond the cited evidence.
- Surface missing counter-evidence & logical gaps.
Prompt 10.1 · Final check
Audit my draft for integrity — list issues by paragraph, don't rewrite.
Draft reviewpaste full text
- Claims that don't match the cited source
- Overgeneralizations beyond the evidence
- Missing counter-evidence & logic gaps
You get → a paragraph-indexed issue list to fix yourself.
Quick Reference
Framework Cheat Sheet
| Stage | Framework | Best for |
|---|
| Question scoping | PICO(S/T) | Clinical / quantitative questions |
| Question scoping | SPIDER | Qualitative questions |
| Question scoping | FINER | Feasibility check |
| Search strategy | PRISMA-S | Reproducible search reporting |
| Screening | PRISMA 2020 | Selection flow / reporting |
| Appraisal | CASP | Design-specific quality checklists |
| Appraisal | GRADE | Certainty of evidence across studies |
| Synthesis | Thomas & Harden | Thematic synthesis (qualitative) |
| Gap analysis | PICO-based gap mapping | Identifying researchable gaps |
Instructor tip: pair each stage with a reflection — "What did the AI oversimplify, and how would you know without checking the source?"
Standalone Handout
NotebookLM Prompts for Literature Review
Pairs with the framework toolkit — but works on its own. NotebookLM reasons only over what you upload.
📂 One notebook per topic
Upload full text (not abstracts) of every included study as a separate Source, named like Smith_2021.pdf so citations map back cleanly.
🔗 Citation-tracked answers
Every claim links to a passage in your sources — it can't fabricate a finding outside your set.
🚫 No outside search
Won't tell you what the wider field says. Still need Scholar/Scopus to build the source set.
🎛️ Studio panel
Audio/Video Overview, Mind Map, Study Guide, Briefing Doc, FAQ, Timeline — mostly one-click, not freeform prompts.
NotebookLM · Chat Prompts
Familiarize & Synthesize
A · Per uploaded study
Summarize this source's research question, methodology, sample/setting and key findings — cite the section or page for each point, and note limitations the authors acknowledge.
B · Cross-source synthesis
Across all sources, what are the most consistent findings on Theme? Name which sources support each claim — include nothing not directly stated in the sources.
Themeweight-loss outcomes
NotebookLM · Chat Prompts
Contradictions & Comparison Table
C · Contradiction detection
Which sources in this notebook disagree on Finding? Name the sources involved and summarize what each one claims.
Findingfasting's effect on HbA1c
D · Methodology table
Build a comparison table across all sources, and cite each row to its source.
- Source · Research design
- Sample / setting · Key variables
- Main finding
NotebookLM · Chat Prompts
Gap Identification
E · Gaps within your set
Based only on the sources in this notebook, what aspects of Topic are not addressed, underexplored, or measured in only one source?
Topicintermittent fasting & T2D
⚠ Caveat
Reveals gaps within your uploaded set only — not gaps in the broader literature you haven't found yet. That still needs database search skills.
NotebookLM · Studio
Using Studio Outputs
📄 Study Guide / FAQ / Briefing
First-pass orientation before deep reading — not a substitute for reading the primary sources.
🗺️ Mind Map
Auto-generated, not promptable. Click a node, then ask in chat: "Which sources support this sub-theme — do any contradict it?"
🎧 Audio Overview
Steerable with a focus instruction → turn it into a critical-appraisal conversation.
Audio Overview · focus steer
Focus on methodological strengths and weaknesses across these sources, not just a summary of findings. Treat it as a critical-appraisal conversation — where do the studies disagree on rigor, and why?
Before You Rely On It
Limitations to Know
🧪 Grounding ≠ validation
It won't invent citations outside your sources — but it does not check whether the uploaded papers are accurate, peer-reviewed, or unbiased. Critical appraisal stays your job.
📊 Hard caps
Free tier limits Audio Overview generations per day; individual sources have size limits. Plan uploads for large review sets.
Source-grounding stops fabrication outside your set — it does not make the set itself trustworthy.
Summary
Session 03 Key Takeaways
- AI agents plan and execute literature tasks autonomously.
- Consensus provides answer-first, evidence-based search.
- PubMed monitors automate surveillance of your niche.
- Claude Projects create persistent lit assistants.
- NotebookLM is the king of source-grounded synthesis.
- Use AI as a "Critical Skeptic" to audit paper quality.
- Anchor every prompt to a framework — verify every citation.
Next Session: Omics Data Analysis — No Code Pipeline (Bulk RNA-seq) and Seurat/Scanpy workflows.
No Code & Agentic AI for Life Sciences
Thank You
© 2026 Md. Jubayer Hossain · No Code & Agentic AI for Life Sciences — Session 03