AI summary by TwinMind
Human Stories in AI: Tommy Tang
StatQuest with Josh Starmer
36m 54s
Background and Early Life
- Tommy Tang from rural southern China, born in low-income family
- Father: middle school teacher, mother: homemaker caring for Tommy and sister
- Family struggled financially - couldn't afford meals, ate soup with soy sauce and green onions
- Mother's advice: "study hard, only knowledge can change your life"
- Ranked #1 in county for high school, attended top university in Shanghai
- August 8, 2008: came to US for PhD at University of Florida
PhD Experience and Transition to Computational Biology
- Initially trained in wet lab molecular biology/cancer biology
- No prior data analysis skills - first year advisor asked to analyze public RNA sequencing data
- Data crashed Excel - realized need for computational skills
- Self-taught computational biology through online courses (edX, Coursera, Udacity)
- Spent evenings after lab work learning programming, doing homework, Googling
- Last 2 years of PhD: learned basic Unix commands, R, Python
- PhD took 6 years total: first 4 years wet lab publications, last 2 years computational learning
MD Anderson Cancer Center Postdoc
- Full computational biology postdoc under Dr. Roel Hak (Cancer Genome Atlas project)
- Analyzed large-scale genomic data sets (TB size)
- Learned multiple sequencing data types: whole exome, whole genome, RNA-seq, ChIP-seq, ATAC-seq, reduced bisulfite sequencing for DNA methylation
- Developed expertise in workflow language SnakeMake
- Dr. Kunal Rai's lab: high-throughput ChIP sequencing
- Profiled 6 different histone modifications across hundreds of samples
- 6 modifications + input control = 7 × hundreds = thousands of samples
- Built SnakeMake pipeline for uniform processing
Harvard and Dana-Farber Cancer Institute
- Harvard Faculty of Arts and Sciences Informatics Division: senior bioinformatician (1.5 years)
- Analyzed single cell RNA sequencing and single cell ATAC sequencing data
- Collaborated with neuroscience labs
- Dana-Farber: lead scientist on NIH-funded consortium project CIDC (Cancer Immunological Data Commons)
- 4 cancer centers: MD Anderson, Dana-Farber, Mount Sinai, Stanford
- Immunotherapy clinical trials with Next Generation sequencing profiling
- Technologies: whole exome, whole genome, TCR sequencing, RNA sequencing
- Collaborated with Ethan Cerami's software engineering team
- Processed clinical trial data securely on cloud, packaged into web portal for clinical teams
Immunitas Therapeutics - Current Role
- Director of Computational Biology
- Single cell genomics company using single cell RNA-seq and TCR-seq data
- Mission: find new therapeutic targets for cancer patients
- Antibody engineering team generates antibodies for wet biology team testing
- Leading program: IM9 (CD161 antibody)
- CD161: T cell and NK cell modulator
- Currently in Phase 1 clinical trial
- 7-8 patients dosed, showing safety profile
- Tommy's role: identify new targets using same single cell platform, support biomarker analysis from clinical trials
Technical Methodology and Machine Learning Approaches
- 10x Genomics machine for single cell RNA-seq and TCR-seq data
- Single cell provides: gene expression profile + T cell receptor sequence
- T cell expansion analysis: when T cells recognize tumor antigens, they expand with identical TCR sequences
- Counting TCR occurrences indicates expansion level
- Correlate gene expression profiles with T cell expansion phenotypes
- Machine learning approaches: simple, interpretable methods
- Logistic regression for expanded vs non-expanded T cell classification
- Random forest for correlation analysis
- Focus on feature weights to identify genes positively/negatively correlated with expansion
PD-1 Checkpoint Biology Example
- PD-1: immune checkpoint on T cells, increases when T cells activated
- Tumor cells express PD-L1 (PD-1 ligand) to block PD-1
- Result: T cells bypass cancer cells (tumor immune evasion)
- Anti-PD-1 antibodies: occupy PD-1 pocket, prevent PD-L1 binding
- PD-1 shows up in analysis as positively correlated with T cell expansion (validation of methodology)
- Biology complexity: PD-1 promotes expansion but also immune system dampening (balance mechanism)
- PD-1 also marker of immune exhaustion when T cells activated too long
Data Management and Infrastructure
- Google Cloud platform for computing and storage
- Single cell datasets: millions of cells, require large RAM and disk space
- In-house library preparation, external sequencing at Broad Institute
- FastQ files uploaded to Google Cloud virtual machines
- SnakeMake pipelines for data processing and analysis
- Pre-processing: 10x Cell Ranger (tens of hours) vs alignment-free tools like Kallisto/Salmon (1 hour)
Cell Type Annotation Challenges
- Major challenge: classifying different cell types in tumor samples
- Tumor contains: different immune cells, cancer cells
- Methods: marker-based or machine learning approaches
- Problem: immunologists always change automated annotations
- No standardized nomenclature for different cell states
- Cells exist on continuous state spectrum, not discrete categories
- Resolution parameter tuning: higher resolution = more clusters, lower = fewer clusters
- Ground truth problem: no definitive cell type classifications in real biological samples
Learning Philosophy and Advice
- Meta-skills: skills that enable learning anything
- Deep learning example methodology:
- Read technical books with code examples (non-biology applications)
- Watch educational videos for high-level understanding
- Apply to specific biological questions with example datasets
- Implement neural networks on biological data
- Complete projects using newly learned skills
- Teaching others: write blog posts to solidify understanding
- Blog: "Diving into Genetics and Genomics" (divingintogeneticsandgenomics.com)
- Currently writing series on learning deep learning
