Session 06 · No-Code & Agentic AI for Life Sciences

Deep Learning
Prediction Methods

Biomarker classification · Model evaluation · Responsible AI

Md. Jubayer Hossain · Founder & CEO
DeepBio Limited · DeepBio Academy

No Code & Agentic AI for Life Sciences · Session 06 · June 2026

Agentic Orchestration

Prepare Claude for AI Workflows

Transforming a chat assistant into a reliable prediction engine requires three core layers:

  • Claude Chat: Model selection and prompt engineering for zero-shot predictions.
  • Claude Co-Work: Agentic design and workflow orchestration.
  • Claude Code: The P-A-E-I framework for iterative model building.

🎯 Model Selection

Claude 3.5 Sonnet: Best for reasoning and code generation.
Claude 3 Haiku: Fast, cheap classification tasks.

"The P-A-E-I loop allows agents to fix their own training code until the model meets performance metrics."
Conceptual

Neural Network Intuition

Think of a neural network as a series of adjustable filters that learn to recognize complex patterns.

  • Input Layer: Biological features (Gene expression, image pixels).
  • Hidden Layers: Abstract representations (Pathways, edges, textures).
  • Output Layer: Prediction (Healthy vs. Disease).

Layers learn from features to higher-level concepts.

Clinical Use Case

Biomarker Classification

🧬 The Task

Building a model to predict drug response based on a specific gene signature or protein profile.

🛠 Tools

  • Google Teachable Machine: Fast, no-code vision/audio/pose models.
  • Scikit-learn (via Agents): Random Forests and SVMs for tabular omics data.
# Teachable Machine Workflow 1. Collect samples (e.g., Histology slides: Normal vs. Malignant). 2. Train in browser. 3. Export model for use in clinical decision support apps.
Validation

Evaluation: Beyond Accuracy

Accuracy is misleading in biomedical data where one class (e.g., rare disease) is infrequent.

  • Precision: "Of all predicted positive, how many were actually positive?"
  • Recall (Sensitivity): "Of all actual positive, how many did we find?"
  • AUC-ROC: Overall ability of the model to distinguish between classes.

📊 The Bio-Tradeoff

In cancer screening, we prioritize **Recall** (don't miss any disease) even if **Precision** suffers (more false positives for review).

Pitfalls

The Overfitting Trap

⚠️ Small Data, Big Models

Biomedical datasets often have few samples (n < 100) but many features (p > 20,000 genes). This is a recipe for overfitting.

Solutions

  • Cross-Validation: Testing on multiple "folds" of data.
  • Regularization: Penalizing complex models (L1/L2).
  • Feature Selection: Only using the most relevant biomarkers.
Ethics

Responsible AI: Bias in Prediction

🚨 Historical Bias

Models trained on data from specific populations (e.g., European ancestry) may fail for others.

⚖️ Algorithmic Fairness

Ensuring predictive performance is equal across gender, race, and socioeconomic status.

"A diagnostic model is only as 'intelligent' as the data it was fed. In clinical AI, bias can be a matter of life and death."
Summary

Session 06 Key Takeaways

  • Claude Chat, Co-Work, and Code power agentic DL workflows.
  • Neural networks learn hierarchical bio-features.
  • Teachable Machine enables fast no-code classifier training.
  • Evaluate bio-models using Precision, Recall, and AUC.
  • Prevent overfitting using CV and feature selection.
  • Responsible AI requires auditing for population bias.

Next Session: Drug Discovery & Image Analysis — Target identification and AlphaFold structure prediction.

No Code & Agentic AI for Life Sciences

Thank You

Get in Touch

Md. Jubayer Hossain

bio.link/hossainlab

© 2026 Md. Jubayer Hossain · No Code & Agentic AI for Life Sciences — Session 06