AI Detecting Cancer with Graph Neural Networks: Stanford Internship by Venika Kakarla

Institution: Stanford University School of Medicine, Donna Saker Lab

Overview

High school student Venika Kakarla presented her summer research internship at Stanford University’s Donna Saker Lab, focusing on spatial analysis of hepatocellular carcinoma (liver cancer) using advanced data analysis techniques and machine learning.

Background

The Donna Saker Lab

Focus Areas:

• Gastroenterology research
• Digestive system studies
• Hepatocellular carcinoma (HCC)

Significance:

• HCC is the third leading cause of cancer-related deaths worldwide
• Critical need for better detection and prognosis methods
• Integration of technology with biological research

Personal Motivation

Venika’s research builds on her previous work detecting the cancer that killed her grandfather, demonstrating how AI can:

• Reduce computational requirements (2,500 hours → 5 minutes)
• Democratize research access (expensive clusters → Google Colab)
• Accelerate medical discoveries

Lab Procedures

1. Immunohistochemistry (IHC)

Purpose: Locate specific cell types within tumor regions

Method:

• Staining techniques highlight different cell types
• Brown vs. white areas indicate cell presence
• Spatial mapping of immune cells, tumor cells, etc.

Value: Helps researchers identify drug targeting locations

2. Cell Culture

• Growing cells in laboratory conditions
• Controlled research environments
• Detailed cellular study

3. Cryostat Sectioning

• Precise tumor sample cutting
• Microscopy preparation
• Spatial analysis enablement

4. Proteomics

• Protein expression analysis
• Understanding cellular manufacturing
• Functional insights

Research Project: Angiogenic Influence

Project Overview

Goal: Understand how blood vessel cells (endothelial cells) influence tumor growth in hepatocellular carcinoma

Significance:

• Endothelial cells form capillaries
• They deliver blood to tumors
• Critical for tumor progression

Four Analysis Components

1. Direct Interaction Study (30-50 micrometers)
2. Indirect Interaction Study (100+ micrometers)
3. MVD Analysis (Microvessel Density)
4. Clustering Analysis

Data & Methodology

Dataset

Source: Codex software (cell classification and spatial positioning)

Size: ~1.4 million cells across 124-128 regions

Data Points:

• Cell ID
• Cell type (13 different types)
• Cohort (Primary vs. Residual HCC)
• X, Y coordinates (spatial data)
• Protein expression levels

Cohort Definitions

• Primary HCC: First occurrence of tumor
• Residual HCC: Recurrent tumor (second+ occurrence)

Analysis 1: Direct Interactions (30-50μm)

Methodology

• Analyzed cells within 30-50 micrometer radius of endothelial cells
• Compared cell type distributions
• Statistical significance testing

Key Findings

Primary vs. Residual Differences:

• Increased fibroblasts as tumor cells decrease
• Significant differences in multiple cell types
• Green bars = statistically significant differences

Implication: Tumor microenvironment changes significantly with recurrence

Tools Used

• Python: Primary programming language
• PyCharm & Jupyter Lab: Development environments
• Matplotlib: Graph creation
• Pandas & NumPy: Excel/data handling
• SciPy.stats: Statistical analysis

Analysis 2: Indirect Interactions (100-200μm)

Methodology

• Examined cells 100-200 micrometers from endothelial cells
• Farther radius analysis
• Comparative study

Key Findings

Distance-Based Patterns:

• More tumor cells farther from endothelial cells
• More fibroblasts at distance
• Fewer immune cells reaching tumor

Implication: Immune exclusion - immune cells aren’t reaching tumors effectively, while blood delivery continues

Research Impact: Identifies potential therapeutic targets for improving immune cell access

Analysis 3: MVD (Microvessel Density)

Methodology

Grid-Based Approach:

1. Divide each region into 8x8 grid
2. Count endothelial cells per grid square
3. Normalize by total cells per square
4. Categorize as Low MVD vs. High MVD
5. Compare cell type distributions

Visualization

Region → 8x8 Grid → Count cells per square → Classify MVD → Compare patterns

Key Findings

MVD Impact on Cell Distributions:

• Significant differences in cell proportions
• High MVD areas show distinct patterns
• Box plots reveal distribution variations

Implication: Blood vessel density correlates with specific cellular environments

Value

Provides granular, small-scale insights into tumor microenvironment composition

Analysis 4: Clustering with Voronoi Tessellations

Voronoi Tessellation Method

Concept:

1. Take each cell as a data point
2. Find nearest neighboring cells
3. Calculate average distance
4. Draw boundaries at midpoints
5. Create spatial partitions

Result: Visual representation of cell clustering vs. dispersion

Interpretation

• Larger areas = cells are more dispersed
• Smaller areas = cells are clustered together
• Residual HCC shows greater dispersion around endothelial cells

Key Finding

Residual (recurrent) HCC demonstrates:

• Higher area measurements
• Greater cell dispersion
• Different spatial organization than primary HCC

Future Work: Graph Attention Networks (GAT)

Proposed Implementation

GAT Architecture:

• Nodes: Represent individual cells
• Edges: Spatial relationships
• Features: Cell type + patient data

Input Data Integration

Spatial Information:

• Cell positions
• Cell types
• Clustering patterns
• Proximity relationships

Clinical Data:

• Patient age
• Treatment history
• Previous therapies
• Recurrence status

Prediction Goals

1. Recurrence Probability: Likelihood of cancer returning
2. Progression Risk: Speed of cancer worsening
3. Treatment Response: Patient-specific predictions

Multimodal ML Potential

Integration Possibilities:

• Previous protein thermal stability research
• Spatial clustering analysis
• Clinical metadata
• Expression profiles

Goal: Comprehensive predictive model combining multiple data types

Previous Research Connection

Thermal Stability Research

Previous Work:

• Protein thermal stability analysis
• Cancer progression speed detection
• Breakthrough efficiency (2,500 hours → 5 minutes)

Integration Plan:

• Combine thermal stability data
• Add spatial analysis
• Incorporate clinical records
• Create multimodal ML model

Outcome: Enhanced prediction of:

• Cancer onset probability
• Recurrence likelihood
• Patient survival rates

Technical Skills Demonstrated

Data Analysis

• Large-scale dataset handling (1.4M+ cells)
• Statistical significance testing
• Spatial data processing
• Visualization techniques

Programming

• Python proficiency
• Jupyter notebook workflows
• Scientific computing libraries
• Statistical analysis tools

Research Methods

• Experimental design
• Hypothesis testing
• Results interpretation
• Scientific communication

Impact & Recognition

Stanford Internship

• Summer research position
• Donna Saker Lab
• School of Medicine placement
• Mentorship from university researchers

Community Mentorship

• Mentored by Colin McNamara
• Supported by AIMUG community
• Previous presentation in March 2025
• Ongoing research guidance

Publication Plans

• Research paper in progress
• Publication process underway
• Future sharing with community

Future Goals

• Continue Stanford research
• Potential full-time enrollment
• Further AI/ML integration
• Medical research advancement

Key Takeaways

1. AI Democratizes Research: Complex analysis accessible via Python/Colab

2. Spatial Patterns Matter: Cell positioning reveals cancer progression mechanisms

3. Multi-Scale Analysis: Different distance scales reveal different insights

4. Recurrence Has Signatures: Primary vs. residual HCC show distinct spatial patterns

5. ML Future: Graph Attention Networks can integrate spatial + clinical data

6. Young Researchers: High school students can contribute to cutting-edge research

Community Impact

Inspiration

• High school student doing university-level research
• Democratization of AI/ML tools
• Accessible research methodologies
• Mentorship model success

Knowledge Sharing

• Open presentation of methods
• Educational value for community
• Reproducible techniques
• Encouragement for others

Resources

Previous Work

• March 2025 AIMUG presentation (more AI-focused)
• LinkedIn profile with additional projects

Tools & Technologies

• Python: Data analysis
• Google Colab: Accessible computing
• Codex: Cell classification
• Matplotlib: Visualization
• SciPy: Statistical testing
• Pandas/NumPy: Data manipulation

Research Context

• Donna Saker Lab, Stanford
• Hepatocellular carcinoma research
• Spatial biology methods
• Graph neural networks

Questions & Answers

Q: How does this relate to your previous cancer detection research?

A: Both projects can be integrated into a multimodal ML model:

• Thermal stability data (previous work)
• Spatial clustering data (current work)
• Clinical data
• Combined prediction of recurrence and progression

Audience Reaction

• Standing ovation
• Community pride in student achievement
• Recognition of groundbreaking work
• Encouragement for continued research

Call to Action

For Researchers

• Explore spatial analysis techniques
• Consider Graph Attention Networks
• Integrate multiple data modalities
• Share methods openly

For Students

• Research opportunities available
• AI/ML accessible to beginners
• Community support available
• Real impact possible

For Community

• Support young researchers
• Provide mentorship
• Share knowledge
• Encourage innovation

---

Related Sessions:

• LangGraph Middleware
• A2A/AP2 Protocols
• March 2025 AI Cancer Detection Talk

Video: Watch the full presentation in the October 2025 showcase recording

Future Updates: Follow Venika’s LinkedIn for publication announcements and research updates