Lex Fridman PodcastManolis Kellis: Biology of Disease | Lex Fridman Podcast #133
CHAPTERS
- 0:00 – 2:30
Podcast setup: computational biology deep dive and why disease is hard
Lex introduces Manolis Kellis and frames the episode as a science-heavy exploration of genetics, molecular biology, and disease. He sets expectations for a fast-moving discussion that jumps between fundamentals and cutting-edge research.
- •Episode context: third conversation with Kellis, focused on science and genetics
- •Lex’s motivation: awe at biological complexity and implications for intelligence/engineering
- •Transition from intro to the core question: challenges in understanding disease
- 2:30 – 15:31
From model-organism knockouts to human genetics as nature’s perturbation engine
Kellis explains how disease biology historically relied on perturbing genes in model organisms, then mapping insights to humans. He argues that modern human genetics flips the workflow: massive natural variation in humans now drives discovery of basic biology and disease mechanisms.
- •Perturbation as the scientific method: poke the system and model the response
- •Genetic epidemiology provides direction of causality vs observational correlations
- •Humans carry millions of variants—natural experiments at planetary scale
- •Modern cohorts/EHRs enable genotype-by-phenotype matrices across many traits
- 15:31 – 26:47
Layer-by-layer causality: from variants to enhancers, genes, cells, organs, and behavior
They discuss breaking the long path from DNA to disease into measurable intermediate steps. Kellis emphasizes multi-level phenotyping (epigenome, expression, cellular properties, imaging, behavior) and why earlier molecular readouts often yield stronger, cleaner signals.
- •Trade-off: early molecular effects are larger but further from disease endpoints
- •Intermediate phenotypes: epigenomic activity, gene expression, cellular physiology
- •Single-cell profiling as an experimental ‘peeling of complexity’ strategy
- •Expanding phenotypes: cognition, behavior, digital biomarkers beyond clinic basics
- 26:47 – 31:38
What diseases matter most: quality-of-life vs mortality, and societal levers
Kellis ranks disease importance using multiple metrics: lifestyle impact, quality of life, and deaths. They discuss leading causes of death and broaden the view to solvability via policy, prevention, and societal interventions alongside biomedical research.
- •Two importance metrics: well-being/quality-of-life and deaths/years lost
- •Major killers: heart disease, cancer, accidents, respiratory disease, Alzheimer’s, stroke
- •COVID as a high-impact lifestyle and mortality disruptor
- •Societal interventions: nutrition, exercise access, pricing externalities, education/empathy
- 31:38 – 41:35
Genes, environment, and what ‘strong genetic effects’ really mean
The conversation digs into heritability, why genetics can illuminate mechanism even when environment matters, and how “strong-effect” variants can be misleading. Kellis explains that lack of variation can signal essential genes and that pharmacology can exploit genes where evolution only tolerates small changes.
- •Genetics → mechanism → actionable intervention points (even against environmental drivers)
- •Heritability example: Alzheimer’s heavily genetic with meaningful environmental modulation
- •Constraint as signal: essential genes tolerate few mutations, so GWAS may miss them
- •Distinguishing: variant→gene effect vs gene→disease effect; targeting the best leverage points
- 41:35 – 45:42
Unexpected disease biology revealed by genetics: AMD, schizophrenia, and immune pathways in the brain
Kellis gives examples where genetic results overturn intuitive tissue assumptions. Age-related macular degeneration and schizophrenia implicate complement/immune pathways, and microglia-mediated synaptic pruning emerges as a surprising mechanistic bridge.
- •Personal example: AMD risk and how genetics changes both understanding and choices
- •Complement pathway links eye disease (AMD) and psychiatric disease (schizophrenia)
- •Microglia as immune cells in the brain involved in synaptic pruning
- •Genetics as an ‘unbiased’ way to discover hidden causal biology
- 45:42 – 53:20
Roadmap epigenomics and tissue-of-action mapping: Alzheimer’s points to microglia, not neurons
Kellis describes using epigenomic maps across tissues to see where disease variants concentrate. A major surprise: Alzheimer’s loci show no enrichment in bulk brain samples dominated by neurons/astrocytes/oligodendrocytes, but strong enrichment in microglia, enabling more targeted therapeutic strategies.
- •Epigenomic ‘chromatin state’ maps reveal where noncoding disease variants act
- •Trait-by-tissue enrichment examples: immune traits in immune cells; T2D in pancreatic islets
- •Alzheimer’s surprise: zero enrichment in bulk brain samples → points to microglia
- •Knowing tissue/pathway of action guides assays, screens, and therapeutic development
- 53:20 – 1:03:26
Genetics unifies disease research: collapsing departmental silos into genome circuitry
Kellis argues that genetics is dissolving traditional disease boundaries by revealing shared circuitry and cross-tissue mechanisms. He describes why his lab spans many diseases while focusing on the common subproblem: decoding regulatory circuitry across cell types and tissues.
- •Genetics reveals cross-disease links (immunology ↔ neurodegeneration, brain ↔ metabolism)
- •Need for a ‘circuitry’ core shared across disease domains (dynamic programming analogy)
- •Kellis lab structure: circuitry foundation + disease-area collaborations + translation to pharma
- •Long-term goal: map function for every nucleotide across tissues/cell types
- 1:03:26 – 1:22:18
The modern pipeline from GWAS hit to mechanism: why 93% of disease variants are noncoding
Kellis outlines a staged workflow for going from association to causal biology and therapy. A key challenge is that most disease-associated variants sit outside protein-coding regions, so mapping a locus to the true target gene requires understanding long-range regulation.
- •Polygenic diseases require scalable, multi-stage understanding (not one postdoc per gene)
- •GWAS often links a genomic region, not a gene, because most hits are noncoding
- •Long-range regulation means ‘nearest gene’ assumptions can be wrong
- •FTO obesity locus as a cautionary tale: association within one gene can control distant genes
- 1:22:18 – 1:37:14
Linking variants to target genes: 3D genome folding, eQTLs, activity links, and CRISPR perturbations
They walk through the toolkit for assigning noncoding variants to their functional targets. Kellis explains Hi-C using the ‘glued noodles’ analogy, describes genetic and activity-based links, and then provides a detailed CRISPR overview—from bacterial immunity to dCas9 activation/repression.
- •Four linkage strategies: physical (Hi-C), genetic (eQTL), activity correlation, causal perturbation
- •Hi-C intuition: crosslinked proximity ligation reveals 3D contacts and domains
- •CRISPR fundamentals: RNA-guided targeting; repair pathways enable editing; prime editing as refinement
- •dCas9 enables functional tests without cutting—turn enhancers on/off to establish causality
- 1:37:14 – 1:47:50
Obesity mechanism case study: IRX3/IRX5, mitochondria, thermogenesis vs fat storage
Kellis completes the FTO story by tracing the causal variant through enhancer disruption to transcription-factor binding and downstream metabolic programs. The key biological switch is whether adipocyte precursors bias toward energy storage (lipogenesis) or energy dissipation (thermogenesis) via mitochondrial programs.
- •Causal chain: variant → motif disruption → regulator binding loss → IRX3/IRX5 upregulation
- •IRX3/IRX5 associate with lipid metabolism (up) and mitochondrial biogenesis (down)
- •Thermogenesis as the ‘energy loss’ term in the energy balance equation
- •CRISPR editing of a single nucleotide can flip adipocyte phenotypes in vitro
- 1:47:50 – 1:58:41
Scaling discovery: automation, MPRAs, and pooled perturbation readouts
Kellis describes the shift from solving one locus over years to testing thousands systematically. He explains massively parallel reporter assays (MPRA) and variants (SHARPR-MPRA, HYDRA, STARR-seq-style self-transcribing reporters), plus the philosophy of using computation to choose the best hypotheses to test.
- •Robotics/automation for high-throughput screening
- •MPRA concept: test thousands of enhancer/variant hypotheses in parallel with barcoded readouts
- •Method extensions to increase resolution (tiling) and scale (millions of fragments)
- •Not brute force over the whole genome: use priors and data integration to pick hypotheses
- 1:58:41 – 2:17:06
Single-cell revolution and multimodal atlases: bubbles, barcodes, and 10M-cell brain datasets
They move to single-cell RNA sequencing and how barcoding enables massive scaling—from wells to droplets to combinatorial indexing. Kellis then connects single-cell readouts with CRISPR perturbations and describes a huge multi-disorder human brain dataset integrating RNA and epigenomic accessibility.
- •Single-cell progression: wells → droplets → combinatorial ‘bottle’ barcoding
- •Droplet barcoding enables thousands–millions of cells with cell-of-origin tracking
- •Coupling CRISPR libraries with single-cell RNA readouts for pooled causal screens
- •Multimodal profiling (RNA + chromatin accessibility) to reconstruct enhancer–gene circuitry at cell-type resolution
- 2:17:06 – 2:34:57
Where the field is going: schizophrenia regulators, Alzheimer’s progression models, mosaicism, and systems medicine
Kellis outlines near-term research directions and long-term medical transformation. Topics include unified regulator models in schizophrenia, region-by-region progression modeling in Alzheimer’s for early intervention, somatic mosaicism in the brain, AI-driven drug design, and network-level therapeutics to avoid on-target side effects.
- •Schizophrenia: regulators bridging early development and synaptic pruning in excitatory/inhibitory neurons
- •Alzheimer’s: pseudotime/progression models across regions; early biomarkers before cognitive symptoms
- •Somatic mosaicism: post-zygotic mutations as a distinct signal vs common/rare inherited variants
- •Future therapy: cell-type-specific delivery, synthetic biology logic, multi-target network interventions (systems medicine)
