AI summary by TwinMind
Daphne Koller: Biomedicine and Machine Learning | Lex Fridman Podcast #93
Lex Fridman
1h 12m 3s
Background and Introduction
- Daphne Koller: Stanford CS professor, Coursera co-founder with Andrew Ng, CEO of Insitro
- Insitro: company at intersection of machine learning and biomedicine
- Focus on using data-driven ML methods for drug discovery and development at scale
- Conversation recorded before COVID-19 outbreak
Philosophical Questions on Disease and Longevity
- Curing all major diseases: very challenging due to damage accumulation by discovery time
- Few diseases actually "cured" - mostly treatments available
- Understanding disease mechanisms: 0-100 scale varies by disease
- Some diseases (70s-80s understanding): certain cancers
- Close to zero understanding: Alzheimer's, schizophrenia
- Mixed understanding: type 2 diabetes (insulin resistance mechanisms known, others unknown)
- Diseases are heterogeneous collections of mechanisms, not single entities
- Breast cancer example: multiple cellular mechanisms causing uncontrolled proliferation
- Longevity vs disease: overlapping but not identical
- Disease risk increases exponentially year-on-year after age 40
- Aging processes: DNA damage accumulation, misfolded protein aggregation, cellular wear and tear
- Health span vs lifespan: goal of living to 120 in perfect health quality
Machine Learning Role in Healthcare
- Historical limitation: insufficient large-scale, high-quality datasets
- Recent changes: emergence of technologies producing data at scale
- Insitro approach: flip traditional paradigm
- Traditional: generate data for scientific discovery, then apply ML as afterthought
- Insitro: design data generation specifically for ML applications
- Goal: create predictive models addressing fundamental human health problems
Personal Motivation and Background
- Early interest (2000s): biology datasets more inspiring than "20 newsgroups dataset"
- Father's death 12 years ago: autoimmune disease, only treatment was prednisone
- FDA wouldn't approve prednisone today due to side effect/benefit ratio
- Now 4-5+ drugs available for autoimmune diseases
- Current "golden time" in drug discovery: ability to create safer, more effective drugs
- Missing piece: insufficient biological understanding to know where to aim
Disease-in-a-Dish Models vs Animal Models
Animal Models Limitations
- Create disease phenotype copy, not mechanism copy
- Mice don't naturally get human diseases (Alzheimer's, diabetes, atherosclerosis, autism, schizophrenia)
- Cures in animals don't translate to humans
- Most drug failures occur due to poor translation from mouse to human
Disease-in-a-Dish Technology
- Enabled by technologies <5-10 years old
- Yamanaka factors: revert any cell (e.g., skin) to pluripotent stem cell status
- Stem cells can differentiate into specific cell types (neurons, cardiomyocytes, hepatocytes)
- Maintains individual genetics while changing cell type
- Can observe genetic disease burden manifest in specific cell types
- CRISPR gene editing: introduce known pathogenic mutations for controlled comparisons
Stem Cell Technology Details
- IPS (induced pluripotent stem) cells: 5,000-10,000 exist worldwide
- Initially finicky, now industrialized with contract research organizations
- Good success rate, though some "memory" of original cell state remains
- Genetic diversity important for capturing ethnic background differences
- Disease burden: polygenic risk scores show 10-12x difference between highest/lowest risk deciles
Data Generation and Measurement
Quantitative Cell Measurement Revolution
- Microarrays (late 90s/early 2000s): measure gene expression across entire genome
- Enabled discovery of molecular disease subtypes (e.g., breast cancer subtypes)
- Single cell RNA sequencing: measure gene activity at individual cell level
- Digital transcript counting: converts "squishy" biology to digital data
- Super-resolution microscopy: sub-cellular structure analysis below diffraction limit
Machine Learning Applications
- Pattern recognition in large cellular datasets
- Identify disease subtypes similar clinically but distinct molecularly
- Test interventions (drugs, CRISPR) to revert disease cells to healthy state
- Less hypothesis-driven approach to discover new interventions
- Avoid targeting same pathways as competitors
Target Diseases and Approach Characteristics
- Diseases with strong genetic basis more likely to manifest in stem cell models
- Need reproducible, robust cellular models
- Diseases contained in one/few cell types easier to model in dish
- Systemic, multi-organ diseases more challenging
- Technology advancing: organoids (cerebral, liver, kidney), multi-organ systems emerging
Coursera and MOOC Revolution
Origin Story (Late 2000s)
- Multiple Stanford efforts: Andrew Ng's CS229 everywhere, Daphne's flipped classroom model
- Fall 2011: first Stanford MOOCs launched
- 100,000+ students within weeks from single New York Times article
- Decision point: continue academic careers vs capitalize on opportunity
Educational Insights
- Short content better: 5-7 minutes vs 15+ minutes
- Shorter courses: 15-week semesters don't work for working adults
- Compression and flexibility crucial for self-paced learning
- Quick feedback through micro-quizzes increases engagement
- A/B testing revealed gender bias effects: female instructors increase women's participation in STEM
Future of Education
- MOOCs won't replace face-to-face teaching entirely
- Learning is social experience - study groups improve outcomes
- Continuing education primary use case: skills change rapidly, formal education not accessible
- Foundation learning crucial before advanced topics
Machine Learning Philosophy and Beauty
Most Beautiful Ideas
- End-to-end training: raw data to outcome without manual feature engineering
- Transfer learning: representations learned for one task accelerate learning on different tasks
Surprises in Neural Networks
- Success in finding meaningful representations in high-dimensional spaces with large datasets
- Early belief: needed constrained, knowledge-rich search spaces for small datasets
- Architecture still matters: different networks for images, language, speech, biology
- Question remains: will universal learning machine emerge or will domain insight always be needed?
Uncertainty in Machine Learning
Current Limitations
- Models provide answers without calibrated confidence measures
- Poorly calibrated uncertainty: 0.8 vs 0.9 confidence not meaningful
- Worse performance and higher confidence on out-of-distribution data
- Critical for medical diagnosis: wrong confident diagnosis can kill patients
- Autonomous driving: need "I don't know" capability for unknown objects
Approaches to Uncertainty
- Bayesian deep learning
- Gaussian processes
- Ensemble methods with different training subsets/random starts
- Active area of research
AI Safety and AGI Perspectives
AGI Timeline and Feasibility
- Still very far away despite progress since 1950s
- Current ML: excellent pattern recognition in specific domains with sufficient training data
- Fails completely when moved to different problem versions
- Nowhere close to human toddler's versatility and context-switching ability
- Not worried about machines taking over universe
Real AI Risks
- Trouble from much simpler systems already occurring
- Complex, poorly understood systems with unpredictable ripple effects
- Similar to electric grid, financial market software complexity
- Need better interpretability and generalizability testing
- Technology abuse potential (like projectile weapons → guns, nuclear power → bombs)
- Gene editing potentially more dangerous than ML for creating harmful viruses
Social Implications and Optimism
- Overall optimistic: most people mean well
- Importance of social norms where doing good correlates with peer approval
- Easy to create dysfunctional societies where atrocious behavior is rewarded
- Concern about some societies moving backward in this regard
Life Philosophy and Meaning
Personal Mission
- Steve Jobs quote influence: "make a dent in the universe"
- Goal: leave world better than when entered
- Greater burden for those born to privilege
- Advantages: educated family, loving parents, great education, food security
- Teaching children about responsibility that comes with privilege
- Using opportunities to make world better place
Action Items
- Continue developing disease-in-a-dish models for genetic diseases
- Expand organoid technology for multi-organ system modeling
- Improve machine learning uncertainty quantification methods
- Develop better generalizability testing for ML systems
- Maintain social norms that reward positive contributions to society
