Integrating Bio and AI — Putting Parts #1 to #14 into Your Lab Workflow

Why This Part?

If you've followed parts #1 to #14 in order, you already understand the following:

• The inherent nature of LLMs as probabilistic machines and their training and alignment principles (Parts #1-4).
• The internal structure of Transformers, embeddings, and attention (Parts #5 & 6).
• Prompts, RAG, agents, and context management (Parts #7-10).
• The limitations of hallucinations and alignment (Part #11).
• The practical landscape of PyTorch, Hugging Face, commercial APIs, and coding agents (Parts #12-14).

The goal of this part: How to integrate all of this into a real-world bio research lab workflow. It's the culmination of the 15-part track and the starting point for you to actually incorporate AI into your research, teaching, and lab.

This part does not introduce any new principles. It brings together the previous 14 parts into a practical roadmap.

The Five Stages of Bio Research

Let's assume your research generally follows these five stages:

1. Literature Review and Idea Generation: Reviewing the latest literature, formulating hypotheses.
2. Experiment Design: Protocols, reagents, statistical power.
3. Experiment Execution: Performing the actual experiment, collecting data.
4. Data Analysis: Statistics, visualization, interpretation.
5. Paper Writing and Presentation: Organizing results, preparing papers and presentations.

Here's how AI can be integrated into each stage, along with the corresponding part number.

Stage 1: Literature Review and Idea Generation

Traditional Method: PubMed search → Scan abstracts of 100 papers → Read 20 relevant papers in detail → Organize notes → Formulate ideas. Takes days to weeks.

AI-Integrated Method:

A. Literature Search and Summarization Assistant (Parts #7 & 8).

• RAG pipeline: PubMed, bioRxiv, and lab paper repository in a vector database.
• Embeddings: Domain-specific (BioBERT, SPECTER) or powerful general-purpose (text-embedding-3-large).
• LLM: Summarization using Claude or GPT-4o. Strict adherence to groundedness prompt rules.

B. Literature QA (Parts #8 & 11).

"What are the recent advances in improving the success rate of multiple gene editing using CRISPR base editing?"

RAG system retrieves relevant paper snippets → LLM generates a comprehensive answer + citations. A citation verification layer (Part #11) confirms the existence of the citations.

C. Idea Brainstorming (Part #7, Chain of Thought).

"My research topic is the regulation of T cell exhaustion. For each of the following five approaches, please provide the pros and cons and recent evidence: (1) TOX inhibition, (2) NFAT regulation, ..."

Ensure diversity with a temperature of 0.7.

D. Journal Club Preparation (Part #10, Hierarchical).

Hierarchically summarize a large review paper of 100+ pages. Section-by-section summary → overall synthesis.

Time Savings: Days → hours. However, it's essential to always verify citations and check the original source when making important decisions.

Stage 2: Experiment Design

A. Protocol Draft (Parts #7 & 9).

Agent tools:

• search_lab_papers: Search for similar experiments in our lab.
• search_pubmed: Search for standard protocols.
• get_reagent_stock: Check reagent stock.

This is the complete code from Scenario 1 in Part #9.

B. Statistical Power Calculation (Parts #7 & 9).

"For this experiment, calculate the required sample size with alpha=0.05, power=0.8, and effect size Cohen's d=0.5 for each condition."

The agent calls statsmodels.stats.power → result table.

C. Safety Review (Part #7, Scenario 2).

Review BSL level, toxic reagents, and PPE requirements. Role prompting + Chain of Thought.

D. Review of Potential Problems.

Use a role prompt that captures the insights of an experienced senior researcher to predict potential pitfalls in the protocol. "What are the five most likely causes of failure in this experiment?"

Stage 3: Experiment Execution

A. Real-time Lab Notebook Assistant.

Voice-to-text (Whisper, speech-to-text API) + LLM summarization. Automatically record spoken content during the experiment in a structured format in the lab notebook.

B. Automatic Logging of Images and Measurement Data.

Automatically detect instrument files → extract metadata → register in LIMS. File Watcher Agent from Part #9.

C. Anomaly Detection.

Automatically detect unusual phenomena during the experiment (abnormal cell morphology, spiking measurement values). VLM (Vision-Language Model, e.g., GPT-4o Vision) performs immediate image analysis.

D. Chatbot Interface.

During the experiment, quickly answer questions like "What was the doubling time of this cell line?" or "Summarize yesterday's experiment results" in Slack or an app. Lab knowledge chatbot (Scenario 1 in Part #8).

Stage 4: Data Analysis

A. Pipeline Development (Parts #12 & 14).

Develop analysis scripts using Claude Code and Cursor. Utilize Scanpy, Seurat, and Bioconductor.

The Snakemake workflow automation from Scenario 1 in Part #14 is applicable here.

B. Exploratory Analysis (Jupyter + Cursor).

In the notebook, use Cursor Chat to instantly process requests like "Add new metadata colors to this UMAP," or "List the top 20 differentially expressed genes in this cluster."

C. Statistical Analysis Verification.

"Review this analysis code to check if multiple testing correction, normality assumption, and effect size calculation are appropriate."

Code review workflow from Part #14.

D. Utilization of Domain-Specific Models (Part #13).

• Analyze protein sequences with ESM.
• Predict structures with AlphaFold Colabfold.
• Determine cell types with scVI and CellTypist.
• Foundation models (scGPT, Geneformer) for downstream tasks.

Load each model from Hugging Face/official repositories.

Stage 5: Paper Writing and Presentation

A. Result Organization (Part #7).

Structured output to summarize experimental results in a paper-style format. Generate initial drafts of figure legends and table descriptions.

B. Figure Creation (Part #14).

Use Cursor to iteratively improve figures using matplotlib, plotly, and ggplot2. Specify styles such as "paper style, colorblind-safe, sans-serif."

C. Draft Writing.

The Methods section can be automatically generated from the actual protocol. The Results section can be generated from data summaries. The Discussion and Introduction sections require human input (creativity and argumentation are needed).

D. Reference Management.

Automatically insert citations using Zotero or Mendeley APIs. The LLM can suggest citation locations.

E. Presentation Materials (Part #14).

Automatically generate presentation drafts (Marp, revealjs). Automatically arrange figures and tables.

F. Reviewer Response (Parts #7 & 11).

Generate initial drafts of responses to reviewer comments. Strengthen the arguments and citations. Human review and editing are required.

Practical Example: Fully Integrated Workflow

Let's say lab member Jisoo is starting a new project. Here's a typical 6-month workflow.

Month 1 — Literature Review.

• Build a RAG pipeline (lab papers + PubMed, last 3 years) using Part #8.
• Use commercial RAG tools like Elicit and Perplexity in parallel.
• Summarize 100 papers with Claude Sonnet → select 20 papers for detailed reading.
• Automate preparation for lab meeting journal clubs.

Month 2 — Idea and Design.

• Brainstorm using the Chain of Thought method (Part #7) to generate 4 hypotheses.
• Meet with the supervisor → select 1 hypothesis.
• Generate a draft of the experiment design (protocol, reagents, stock) using the agent from Part #9.
• Calculate statistical power.
• Complete the safety review → obtain IRB and IACUC approval.

Month 3-4 — Experiment.

• Use the real-time lab notebook assistant.
• Automatically back up data daily + register in LIMS.
• Receive anomaly detection alerts.
• Actively use the lab knowledge chatbot in Slack.

Month 5 — Analysis.

• Build a Snakemake pipeline using Claude Code.
• Perform exploratory analysis in a Jupyter notebook using Cursor.
• Utilize domain-specific models such as ESM and scVI (Part #13).
• Verify statistical analysis.

Month 6 — Paper and Presentation.

• Generate figure drafts using Cursor.
• Generate a draft of the methods section using the LLM.
• Automatically insert and verify citations.
• Prepare a presentation for the conference.

Time Savings: 30-50% compared to traditional methods, especially in literature review, code writing, and drafting papers.

Quality Improvement: Citations and statistical validations are explicitly managed.

Risks: Hallucinations, over-reliance, and insufficient verification. The mitigation strategies from Part #11 are essential.

Lab-Wide Implementation Roadmap

When implementing AI at the lab level, rather than on an individual basis.

Phase 1: Individual Use (Months 1-3)

• Each lab member has a Claude and GPT API account.
• Individual use of Cursor and Claude Code.
• Share prompt tips (Slack channel).

Investment: $20-$50/month per person. Several hours of learning.

Phase 2: Shared Infrastructure (Months 3-9)

• Lab paper and protocol RAG service (Part #8).
• Slack bot integration.
• Shared CLAUDE.md and project context.
• Document standard prompts and workflows.

Investment: Server infrastructure: $100-$500/month. Time to develop the RAG pipeline.

Phase 3: Agent Automation (Month 9+)

• Experiment design agent (Part #9).
• Data pipeline automation.
• Paper draft automation.
• Integration with LIMS and electronic lab notebooks.

Investment: Ongoing development and maintenance. Approximately one part-time developer per lab.

Phase 4: Domain Model Development (Long-Term)

• LoRA fine-tuning using lab data (Part #13).
• Custom MCP server (Part #9).
• Training of small, custom models for specific tasks.

Risks and Limitations

We've emphasized this repeatedly in Parts #1-14, but let's summarize it again.

Hallucinations (Part #11)

• Generation of non-existent papers, genes, or interactions.
• Creation of plausible but incorrect statistical values or manipulation of protocol details.
• Cannot be completely eliminated. Multi-layered verification is essential.

Privacy and Confidentiality

• Do not transmit patient data to APIs (HIPAA, GDPR).
• Use local models (Part #13) for unpublished experimental results.
• Use enterprise plans for patents and competitive information.

Reproducibility

• Record the prompt, model, and seed used for each experiment and analysis.
• Ensure that the final results are reproducible using code that is independent of the model.
• Do not rely solely on LLM responses as definitive results.

Skill Degradation

• Prevent degradation of basic skills in students and lab members.
• Combine understanding of principles with AI utilization. One without the other is not sufficient.
• The principle parts (Parts #1-15) of this track provide that balance.

Bias

• The bias in the training data is reflected in the responses.
• Bias towards Western and English-language literature and protocols.
• Asian and Korean-specific experimental practices may be underrepresented.

Cost Explosion

• API costs can increase rapidly with large-scale use.
• Monitor costs proactively.
• Utilize prompt caching (Part #10).

Future Prospects

Foundation Models for Biology. The emergence of state-of-the-art models for specific domains.

• AlphaFold 3: Predicts protein, ligand, and nucleic acid complexes.
• ESM3: Generates and designs proteins.
• scGPT & Geneformer: Foundation models for cell types.
• Nucleotide Transformer family: Understanding the genome.
• Evo & CaLM: DNA language models.

Multimodal Integration. Combining text + images + sequences + structures into a single model. This will bring significant changes to pathology, structural biology, and molecular discovery.

Autonomous Scientific Agents. Autonomous design and execution of experiments (using robots), analysis, and writing of papers. Still in its early stages, but rapidly evolving. Early examples include Cognition Devin and Sakana AI Scientist.

Democratization. Tools previously exclusive to large labs are now accessible to individuals and small labs. Open models and affordable serving are accelerating this trend.

Regulatory Framework. The FDA and EMA are strengthening regulations on AI-driven decision support systems. Clinical and pharmaceutical applications require more rigorous validation.

Final Advice

For those of you who have completed Episodes #1 through #15:

First, don't forget the fundamentals. Tools are constantly changing. Cursor may be replaced by a different IDE, and Claude may be replaced by a different model. However, the principles of probabilistic machines, transformers, and attention that you learned in Episodes #1-6 will remain unchanged. When you encounter new tools, using these principles to understand them will dramatically shorten the learning curve.

Second, cultivate a habit of verification. No matter what AI says, you must verify it yourself before it is included in your experiments, presentations, or publications. This is not just a skill, but an attitude. Hallucinations will not disappear.

Third, maintain your own perspective. AI is a statistical average of its training data. The unique insights, domain knowledge, and research intuition that your lab possesses cannot be replaced by AI. AI is a tool to augment these qualities, not to replace them.

Fourth, continue to refer to Episodes #1-14. This track is not a one-time learning experience, but a reference guide. Make it a habit to revisit the relevant episodes when you encounter new problems or tools, and review the fundamentals.

Fifth, enjoy the process. While we cannot predict exactly how bio research in the age of AI will evolve, it is certain to be rapid and exciting. I hope you will be an active participant in this change.

Reference Guide for Episodes #1-14

Which episode to revisit in each situation.

Acknowledgements and Conclusion

The BioPlayground's ai-native track of 15 episodes concludes here.

This track was built on the following principles:

• Originality: All scenarios, analogies, and figures were developed by BioPlayground. We did not simply translate symbolic representations from specific external authors.
• Bio Identity: We understand AI principles through the lens of cells, genes, proteins, and laboratory scenarios. This lowers the barrier to entry for bio researchers while also establishing BioPlayground's unique identity.
• Parallelism of Principles and Applications: Phase 1, which focuses on principles, explains why the techniques in Phase 2 work. We did not simply list tips without explaining the underlying principles.
• Depth and Breadth: The appendices provide mathematical and system details that will satisfy graduate-level readers, while the main text is accessible to undergraduate students and newcomers.

I hope this track is helpful to your research and growth.

Master Reference (Entire Track)

• Phase 1: Principles: what-is-llm, neural-network-basics, how-nn-learns, backpropagation-intuition, transformer-and-embedding, attention-mechanism
• Phase 2: Applications: prompt-engineering, rag-and-context, agent-and-tool-use, context-window-management, hallucination-and-alignment
• Phase 3: Tools: pytorch-basics, huggingface-and-openai, claude-code-and-cursor
• Phase 4: Integration: bio-ai-integration (this episode)

📐 Appendix — Practical Checklist for Lab Implementation

Difficulty: Practical Target Audience: Lab, research institute, and corporate R&D team implementation managers

A.1 Individual Usage Checklist

• Secure accounts for Claude, GPT, and Gemini.
• Install and familiarize yourself with the basics of Cursor or Claude Code.
• Learn prompt engineering tips (Episode #7).
• Gain experience with at least one real-world task automation.
• Establish a habit of verifying citations and facts.

A.2 Lab Infrastructure Checklist

• Set up a shared vector database (Chroma, Qdrant).
• Index your lab's publications and protocols.
• Build a Slack bot or web UI chatbot.
• Standardize CLAUDE.md and project context.
• Deploy an MCP server (PubMed, UniProt, etc.).

A.3 Regulations and Safety Checklist

• Policies for processing personal and patient data.
• API key management (separate personal and lab accounts).
• Audit logs and data retention policies.
• Policies for conflicts of interest and copyright.
• AI usage guidelines for student supervision.

A.4 Budget Plan

Small lab (5 people):

• Individual tools: $50/person/month x 5 =$250/month.
• Infrastructure: $200/month.
• Total: $450/month ≈$5400/year.

Medium lab (20 people):

• Individual tools: $50/person/month x 20 =$1000/month.
• Infrastructure: $500/month.
• Fine-tuning and special models: $2000/month.
• Total: $3500/month ≈$42000/year.

Enterprise:

• Enterprise API contract.
• In-house GPU infrastructure.
• Part-time AI engineer.
• $10K-100K/month per lab.

A.5 Performance Metrics

• Reduction in literature review time.
• Code development speed (LOC/day).
• Time to complete a draft manuscript.
• Reduction in experimental failure rate (due to improved design).
• Lab member satisfaction survey.
• ROI (time savings x lab member cost).

A.6 Risk Matrix

A.7 Recommended Implementation Order

1. Individual API usage (Week 1).
2. Cursor/Claude Code individual adoption (Weeks 2-4).
3. RAG pilot service (Months 2-3).
4. Agent automation (Months 4-6).
5. Custom model fine-tuning (Months 6-12).
6. Full integration and automation (Year 2+).

A.8 Learning Roadmap

Week 1: Episodes #1 and #7 (LLMs and prompt engineering). Week 2: Episode #8 (RAG). Week 3: Episodes #9 and #10 (Agents and context windows). Week 4: Episode #11 (Hallucinations and safety). Month 2: Episodes #12 and #13 (PyTorch, Hugging Face, and APIs). Month 3: Episode #14 (Coding agents). Months 4-6: In-depth review of the principles in Episodes #2, #3, #4, #5, and #6 (as needed).

Start with the application episodes and then delve into the principles for faster adoption in practice.

A.9 Community and Resources

• BioPlayground (this track): Continuously updated.
• Hugging Face Bio: Open-source models and data.
• Anthropic Cookbook and OpenAI Cookbook: Official examples.
• AI in Biology Newsletter: The Sequence, Elicit blog, etc.
• Korean Community: Domestic AI and biology study groups and conferences.

A.10 Next Steps

If you have completed this track:

• Apply it to real-world projects.
• Deepen your knowledge in a specific domain (e.g., protein design, genomics analysis, image diagnostics).
• Contribute to open-source projects.
• Share your knowledge with your team and lab.
• Continue learning (the technology is rapidly evolving).

References

The content, scenarios, analogies, and figures in this episode were all developed by BioPlayground. The following are external references that may be helpful for learning the concepts.

• AlphaFold: alphafold.ebi.ac.uk
• ESM: github.com/facebookresearch/esm
• scGPT: github.com/bowang-lab/scGPT
• Geneformer: huggingface.co/ctheodoris/Geneformer
• Evo (DNA LM): github.com/evo-design/evo
• Hugging Face Bio: huggingface.co/spaces/bigbio
• Elicit: elicit.com — AI research assistant
• Consensus: consensus.app — AI literature search
• Sakana AI Scientist: sakana.ai/ai-scientist
• Anthropic Cookbook: github.com/anthropics/anthropic-cookbook
• OpenAI Cookbook: github.com/openai/openai-cookbook
• BioPlayground DryBench: bioplayground.com/drybench

The BioPlayground's ai-native track of 15 episodes concludes here. I hope this track serves as a useful guide for your research journey, and that BioPlayground will continue to support your next chapter.