Preface
This is a working textbook about how foundation models — the same kind of large, pretrained models behind modern AI — are being used to read and design the molecules of life: proteins, genomes, and cells. It is written for someone who already knows machine learning — you can read a training loop, you know what an embedding is, you have used PyTorch or JAX — but who wants the current, load-bearing picture of what these models actually do in biology, and where they still break.
What this book assumes
You are comfortable with the mechanics of training a neural network and with the idea of a pretrained model whose representations transfer to new tasks. You do not need a biology degree. Where a biological concept matters — what an assay measures, what a variant is, why a benchmark number does or does not answer your question — we teach it, in enough depth to reason correctly and no more. The machine learning is assumed and refreshed quickly (Chapter 3); the biology is the part we slow down for.
The goal is not encyclopedic coverage. It is the load-bearing set of ideas: the ones that come up when you nominate a drug target, score a genetic variant, design a protein, or sit across from a wet-lab collaborator who asks whether your model's prediction is worth an experiment.
How it is organized
The book runs on one spine, chapter after chapter: the problem, the models that attack it, and what is still hard. That last part is not a formality — these models are genuinely useful and genuinely oversold, and telling the difference is the whole skill.
Part I builds the shared vocabulary: the landscape of problems, a machine-learning on-ramp, and the data and genetics you need. The problem chapters then split into two lobes. Part II is the molecular and therapeutic side — target discovery, property prediction, structure, design, and cell engineering. Part III is the genomic and regulatory side — predicting function from a stretch of DNA, and tracing a variant to a mechanism. Part IV is the bridge: multimodal models that reason across both lobes at once. Part V, "doing it for real," covers what a benchmark hides — messy data, honest evaluation, and closing the loop with the lab. Part VI looks at the frontier, and doubles as the conclusion.
Appendices hold the fields this edition leaves at the door — single-cell, spatial, and imaging models — plus a glossary, because the book crosses two jargon-dense worlds.
How to read it
The callouts do specific jobs, and you can lean on them:
Intuition
The one-sentence mental model. If you remember nothing else from a section, remember the intuition box.
Analogy
A concrete comparison. Analogies are lies that fit in your head; each one leaks somewhere, and we say where.
Collaborator
A question a wet-lab partner or a skeptical statistician would put to you, and the crisp answer behind it.
Note
A useful detail that would otherwise interrupt the flow.
Common trap
A place people reliably get it wrong.
Every jargon term is defined the first time it appears, and collected in the glossary at the back.
A note on how this book is made
This is a folder of linked HTML files generated from Markdown by a small Python build. The table of contents lives in one file (toc.py); the prose lives in content/. A chapter that is not written yet renders as a navigable stub from its planned outline, so the whole book is always browsable. See CLAUDE.md for the authoring workflow.
Nothing here is final. It is a scaffold meant to be argued with, corrected, and filled in.