- This episode is an in-person conversation at MIT between host Curt Jaimungal and computational biologist Manolis Kellis (MIT and the Broad Institute), centered on Kellis’s effort to articulate a unifying principle that connects biology, cognition, language, and AI—not through rigid laws like physics, but through recurring, emergent patterns of adaptive, layered information processing.
Biology is fundamentally different from physics
- Physics has a small number of laws written ~13.8 billion years ago that apply everywhere; biology has no such fixed rules, only evolving patterns constantly rewritten to fit ecological niches.
- Biology “creates its own rules” through evolution, so unification in biology means finding recurring principles and patterns, not exceptionless laws.
DNA, evolution, and the genome as a language
- The four DNA bases (A, C, G, T) are like binary digits—meaningless alone, but powerful through combinatorics and layered interpretation.
- The genetic code (64 triplets → 20 amino acids + stop codons) is nearly universal because horizontal gene exchange between bacteria and even across species (e.g., viruses infecting humans) requires software compatibility; organisms that changed their translation table became isolated.
- Meaning in DNA arises at many layers: nucleotide motifs recognized by proteins, amino-acid sequences folding into functional 3D structures, and regulatory regions (≈98.5% of the genome) that control when and where proteins act.
- DNA is packaged with histones that can be chemically modified (acetylation, methylation, ubiquitination), adding analog-like “font” modifications (bold, italics, strikethrough) on top of the digital ACGT code.
Layers of abstraction as a unifying principle
- Both biology and human technology build complexity through layers of abstraction: quarks → atoms → molecules → nucleotides/amino acids → protein folds → cells → brains → language → mathematics → computers → programming languages → AI.
- Each layer abstracts away lower-level detail (nucleotides don’t “care” about quantum effects), enabling reuse and recombination of building blocks.
From evolution to cognition
- Cognition emerges when organisms must integrate multiple conflicting sensory streams (chemical, light, heat) into a single model of the world and act on one decision.
- Central nervous systems allow integration across many senses and many competing goals (hunger, thirst, fear, long-term planning), producing increasingly complex internal models.
- Humans have largely shifted from vertical genetic evolution to horizontal cultural/meme evolution—ideas spreading rapidly across the globe and co-evolving with societal niches.
The unifying idea: function-fitting through tunable parameters
- Genomes, protein structures, brains, and AI systems all work by tuning thousands to billions of parameters to fit functions to data or environments.
- Evolution “tinkers” with subtle changes rather than making large jumps; this tinkering capacity itself improves over time (evolvability), explaining the accelerating complexity after multicellularity.
- Classical AI used explicit if-then rules designed by human experts; modern AI (neural networks, convolutions, transformers) learns reusable representations—edges → shapes → objects; motifs → regulatory grammars; words → contextual meanings—mirroring biology’s layered abstraction.
Language, meaning, and creativity
- Language is not just a lossy transmission channel; it is a creative tool that provides primitives for building more complex thought.
- Misunderstanding is productive: both the speaker’s compression of high-dimensional thought into words and the listener’s re-expansion into their own mental space can generate new ideas.
- Teaching forces researchers to re-explain foundations, revealing gaps in their own understanding and driving deeper insight.
Future of medicine and biology
- A convergence of cheap genome sequencing, single-cell and spatial omics, protein-structure prediction (AlphaFold/ESM Fold), electronic health records, and AI is making it possible to represent every gene, disease, and patient through many complementary facets (literature, knowledge graphs, protein structure, imaging, clinical notes).
- Complex diseases (Alzheimer’s, schizophrenia, cardiovascular disease) can be decomposed into reusable pathological “building blocks” (cholesterol transport, amyloid accumulation, tau pathology, neuroinflammation, etc.) that recur across different conditions.
- Truly personalized medicine becomes economically feasible when drugs target shared pathways combinatorially rather than creating one pill per individual.
Advice to students and researchers
- Build broad foundational knowledge across disciplines; the current unification across biology, AI, and medicine rewards people who can think in multiple frameworks.
- Learn to code and use tools like ChatGPT as “rubber ducks” to break down complex papers and codebases.
- Be fearless: dive into advanced material early, even if you don’t yet understand all the prerequisites, and use modern resources (YouTube, arXiv, bioRxiv, AI tutors) to accelerate learning.
- Effort compounds multiplicatively: skill = talent × effort, achievement = skill × effort, so effort counts twice.