LatentCellBio × AI

The Book

Every post is a chapter. This is the whole arc — foundations to frontier, with people as interludes — filling in a few each week.

6 of 100 chapters written

1 Why Bio × AI Now

Part I · Foundations

  1. Why bio × AI is the story of the decadepublished
  2. What is a foundation model? — the shared vocabularyplanned
  3. Attention is the engine: how one 2017 idea ate all of AIplanned
  4. An image is just 16×16 words: why one model reads text, pictures, and cellsplanned
  5. Scaling is not a metaphor: what happens when you make models biggerplanned
  6. Transfer learning: why you no longer train from scratchplanned
  7. Zero-shot: how a model does a task nobody trained it to doplanned
  8. Do LLMs actually know science, or just talk like it?planned
  9. The awe argument: cell ↔ galaxy, why scale is the storyplanned
  10. The human owns the why; the machine owns the howplanned

2 The Central Dogma, Reread by Machines

Part I · Foundations

  1. Ronald Vale — molecular motors and the measurement traditionpersonpublished
  2. DNA → RNA → protein for the AI readerplanned
  3. The genome as a sequence modelplanned
  4. What is a cell? the machinery, for the AI readerplanned
  5. Gene regulation as a prediction problemplanned
  6. Predict the molecule: biology as instruction-followingplanned
  7. Can an LLM reason about what a cell does when perturbed?planned
  8. Plausible ≠ predictive: a fluent biological story isn’t a correct oneplanned
  9. Clinical knowledge as a learned representationplanned
  10. The measurement tradition: why data quality is the real foundationplanned

3 Proteins: Structure, Language & Design

Part II · Reading Biology

  1. From AlphaFold to co-folding — what structure prediction actually doespublished
  2. Protein language models: ESM / ProGen / what they learnplanned
  3. RFdiffusion2 & atom-level active-site scaffoldingplanned
  4. De novo design & the protein functional universeplanned
  5. Reading a fold backwards: inverse folding & ProteinMPNNplanned
  6. Diffusion came for proteins first (Chroma / SE(3))planned
  7. Binders on demand: design a protein that grabs any targetplanned
  8. Antibodies, the hard case: atom-by-atom design yet?planned
  9. It takes two: predicting protein complexesplanned
  10. Proteins move: neural nets that emulate molecular dynamicsplanned

4 The Cell as a Model

Part II · Reading Biology

  1. Single-cell foundation models: scGPT / Geneformerplanned
  2. How to build the virtual cellplanned
  3. One embedding for every cell: a universal coordinate systemplanned
  4. The uncomfortable benchmark: do cell FMs beat linear regression?planned
  5. Perturbation prediction (GEARS / Perturb-seq)planned
  6. 100 million cells: giga-scale perturbation atlasesplanned
  7. Cells in context: spatial-omics foundation modelsplanned
  8. Reverse-engineering the wiring: GRNs from single-cell dataplanned
  9. Beyond RNA: foundation models for the single-cell epigenomeplanned
  10. Rebuilding an embryo in silicoplanned

5 Seeing Disease

Part II · Reading Biology

  1. Immunotherapy and the guided missile — where AI actually helpsplanned
  2. A whole-slide foundation model for pathologyplanned
  3. Multimodal histology + genomicsplanned
  4. How AI learned to read a pathology slide (a primer)planned
  5. Skip the sequencer: reading mutations off a $5 H&E slideplanned
  6. “Screen them all”: dozens of biomarkers from one slideplanned
  7. AI maps the tumor microenvironment at pathologist levelplanned
  8. Fusing slide + proteome + chart: multimodal oncologyplanned
  9. Agentic AI at the bedside: does it earn physician trust?planned
  10. Prognosis from the paper trail: Bayesian EHR + geneticsplanned

6 Designing Drugs

Part III · Acting on Biology

  1. From a slime mold to a heart drug — and what AI does nextpersonpublished
  2. Computational approaches streamlining drug discoveryplanned
  3. You don’t have to score every molecule: active learning at billion scaleplanned
  4. Boltz-2: one model that folds the complex AND predicts affinityplanned
  5. AlphaFold’s first real drug: a CDK20 inhibitorplanned
  6. FEP-quality affinity without the supercomputerplanned
  7. Drawing molecules straight into the pocket: flow-matching designplanned
  8. Teaching a model to optimize molecules with RLplanned
  9. Docking reality check: DiffDock & PoseBustersplanned
  10. ML speeds up antibody (biologics) discoveryplanned

7 Agents That Discover

Part III · Acting on Biology

  1. Building a drug-discovery agent — what actually chains, and where the wall ispublished
  2. AI agents are doing science — where they help, and where they don’tpublished
  3. Autonomous labs & the make-test-decide loopplanned
  4. Graph / orchestration abstractions for agents (LangGraph)planned
  5. The anatomy of a bio-AI company (the TechBio thesis)planned
  6. The AI co-scientist that proposes novel hypothesesplanned
  7. Robin: a multi-agent system that ran a discovery loop end-to-endplanned
  8. Biomni: one general-purpose biomedical research agentplanned
  9. Coscientist: the LLM that ran its own chemistry labplanned
  10. “The AI Scientist” writes its own papers — claim vs realityplanned

8 Editing Life & Curing the Rare

Part III · Acting on Biology

  1. CRISPR-GPT: when the AI designs your gene editplanned
  2. Smaller scissors: compact nucleases that fit in a virusplanned
  3. Rewriting DNA one letter at a time with repurposed RNA editorsplanned
  4. Knowledge graphs as orphan-drug matchmakersplanned
  5. Before the cure, the diagnosis: AI variant interpretation for rare diseaseplanned
  6. Programmable cells: RNA sensors that compute before they treatplanned
  7. Cellular reprogramming (Yamanaka / OSKM), the toolkitplanned
  8. Prime editing: search-and-replace for the genomeplanned
  9. In-vivo CRISPR reaches patients (Casgevy / Verve)planned
  10. Worked example: pick a rare disease, gene → cure pathplanned

9 The Craft & The Frontier

Part IV · The Craft

  1. Correlation isn’t causation — and under small N, design is everythingplanned
  2. Learning from data you’re not allowed to see — privacy-preserving AI in the clinicplanned
  3. The evolution of agentic engineering — from prompts to self-improving loopsplanned
  4. Hallucination is not a glitch — it’s what the objective rewardsplanned
  5. Contamination-free or it didn’t happen: why benchmark scores inflateplanned
  6. The leaderboard illusion: how arena rankings get gamedplanned
  7. Can a model grade itself? LLM-as-verifierplanned
  8. Is “reasoning” real or an illusion? what breaks when problems get hardplanned
  9. Do agents need world models?planned
  10. The validation bottleneck: conjecture faster than science can checkplanned

10 Longevity: Aging as an Engineering Problem

Part IV · The Craft — capstone

  1. The hallmarks of aging: the engineering specplanned
  2. Aging clocks: can we measure biological age?planned
  3. Precious3GPT: one transformer for the whole aging machineplanned
  4. Which organ is aging fastest? organ-specific proteomic clocksplanned
  5. Partial reprogramming reaches the clinic — and hits a wallplanned
  6. Rapamycin wins, metformin doesn’t: the geroprotector scoreboardplanned
  7. Inflammaging: why your immune system ages youplanned
  8. Not all senescent cells die: the case for senosensitizersplanned
  9. Aging as an engineering problem — my perspective on living to 200planned
  10. Books that shaped my view + can we engineer a new human? the ethicsplanned

The order isn’t fixed — it’s the map, not a contract. Published chapters link through; the rest are on the way.