AI summary by TwinMind
Human Stories in AI: Fabio Urbina
StatQuest with Josh Starmer
35m 30s
Background and Career Journey
- Fabio Urbina: Associate Director at Collaborations Pharmaceuticals
- Childhood interest in science, nature documentaries, early exposure to computers (Apple with green text, Oregon Trail, floppy discs)
- College: Bachelor's in Biology + Computer Science minor
- Internship at Massachusetts General Hospital working on familial dysautonomia (rare disease affecting ~400 people worldwide)
- PhD at UNC Chapel Hill in Stephanie Gupton's lab
- Focus: neuron cell biology and development
- Applied computational image analysis to cell biology imaging
- Graduated 2020 (delayed by pandemic)
Transition to Industry
- Decided against academia during final PhD year
- Found Collaborations Pharmaceuticals internship posting
- Emailed CEO Sean, got response within 1 hour, started next day
- Company characteristics when joined:
- <10 employees including CEO
- "Mom and pop operation" - Sean and his wife run company
- Focus on rare/neglected diseases
- Funding model: government grants vs traditional VC funding
- Mission: therapeutics for diseases considered "not profitable"
Machine Learning in Drug Discovery
- Traditional approaches:
- Medicinal chemist alters existing drug structures
- Brute force screening of large compound libraries
- ML approach at Collaborations:
- Train models on known active/inactive compounds
- Virtual screening to predict antimalarial activity (example)
- Narrow down thousands of compounds to 10-100 for testing
- 10-100x or 1000x better hit rates than random screening
Success Rates and Challenges
- Variable success rates:
- Best case: 3/3 compounds active (anti-Ebola/antiviral project)
- Good case: 7/10 compounds efficacious
- Worst case: 0/10 compounds active
- Main challenge: coverage/applicability domain
- Training sets often cover narrow chemical space
- Models overconfident when predicting outside training space
- Data scarcity: expensive to generate experimental data points
Small Dataset Machine Learning
- Typical dataset sizes: hundreds to ~100,000 compounds
- Extreme case: recent project with only 15 compounds training set
- Successfully engineered predictive model from 15 data points
- Approach: few-shot/zero-shot learning models
- Focus on extracting maximal information from tiny datasets
Prototypical Networks Model
- Model architecture: embedding model + prototypical network
- Embedding model:
- Converts compound structures to numerical vectors
- Similar structures → similar vectors
- Dissimilar structures → different vectors
- Prototypical network:
- Calculates average vectors (prototypes) for each class
- New predictions: find closest prototype
- Based on k-nearest neighbor algorithm
- Strategy: choose negative compounds structurally dissimilar from positives
Hardware and Computational Requirements
- Models run on modest hardware (2015 Mac, no GPU required)
- Advantage of small datasets: minimal computational resources needed
- Contrast with large language models requiring massive GPU clusters
- Most people can run these models on standard desktop hardware
Advice for Career Transitions
- Biggest barrier: discomfort with not knowing/understanding immediately
- Importance of persistence through uncomfortable learning phases
- Field switching challenges:
- Starting over from scratch after becoming expert
- Learning new lingo, acronyms, field practices
- Feeling like beginner again after years of expertise
- Key mindset: comfortable with being uncomfortable
- Learning capability transfers between fields
- Worth pursuing if genuinely interested in new direction
