Preface
I wrote this handbook because life sciences AI needs a practical evidence framework. Clinical AI, public health AI, and biosecurity each answer different questions. This book focuses upstream on biological discovery and engineering: molecules, cells, tissues, organisms, environments, translation, therapeutics, and the research systems that connect them.
The goal is not to rank every model or chase every release. The goal is to give researchers a stable way to read claims, design evaluations, and decide when an AI output deserves experimental attention.
Why This Handbook Exists
Three observations led to the handbook.
Life sciences AI needed a practical reference at the right scale. The clinical literature is well-served by physician-facing AI references that focus on FDA-cleared tools, medico-legal exposure, and workflow integration. The public-health literature is well-served by surveillance- and forecasting-focused references that emphasise deployment-context evaluation. The biosecurity literature is well-served by frameworks for dual-use governance and AI-bio convergence. The life sciences layer (protein structure prediction, protein design, genome models, single-cell foundation models, systems biology, plant and ecological AI, therapeutics, autonomous laboratory work) sat between those domains and was treated by none of them in depth.
Marketing was outrunning evidence. Frontier-lab announcements, biotech press releases, and vendor pitches were promising autonomous discovery, virtual cells, virtual organisms, AI-discovered drugs at clinical readout, and end-to-end pipelines. The peer-reviewed evidence was more measured. Independent evaluations (Ahlmann-Eltze et al., Nature Methods 2025; Boiarsky et al., Nature Machine Intelligence 2024; the PoseBusters work by Buttenschoen et al., Chemical Science 2024) changed how those claims should be weighted. Working researchers needed a reference that placed capability claims beside independent evidence.
Researchers needed a stable framework, not a release tracker. Model releases are weekly. Benchmark updates are monthly. The pace makes any single edition obsolete. What does not change as quickly is the framework for reading capability claims: name the biological scale, name the biological object, name the validation experiment, place the claim in an evidence tier, design the validation that turns a prediction into evidence. The handbook uses that durable framework so next year’s models can be assessed with the same discipline.
The reader-facing taxonomy follows that logic. The book moves from foundations to molecular discovery and design, then to cells, tissues, and systems biology, then to organisms and environments, then to therapeutic discovery and translation, and finally to research systems, practice, and governance. The structure is meant to keep future additions from forcing URL churn or part-level rewrites.
Who This Handbook Is For
The shared question across the audience is: when does an AI output deserve experimental attention?
The reader profiles vary. Computational biologists need capability-tier reading and known failure modes per model class. Biotechnology team leads need build-vs-buy framing, license diligence, and what the open-source alternatives are when a frontier release is restricted. Drug discovery scientists need to know where AI shifts a stage gate and where it does not. Plant biologists, ecologists, neuroscientists, and aging researchers need their fields treated as core life sciences rather than edge cases. Physician-scientists need translation between bench AI and clinical decision-making. Synthetic biologists need design-tool selection plus dual-use awareness. Graduate students need canonical citations and conceptual entry points. Research program leaders need the capital-allocation framing.
How Evidence Is Handled
The handbook uses two repeated layers throughout. The first is a field guide. It gives readers enough context to enter a domain without pretending that every reader is a beginner. Major chapters answer:
- What is this field trying to solve? The biological or research problem that defines the area.
- What is the core idea? The assumptions, constraints, and terms that prevent misreading the field.
- What is the current state of the field? The active methods, model classes, and evidence landscape.
- What do we know, and what remains open? The settled reference points, unresolved questions, and limits of current evidence.
- Why does this matter? The scientific, translational, or program implication.
The second layer is evidence calibration. Major chapters are also written to answer:
- What is demonstrated? Evidence supported by published work, official documentation, reproducible benchmarks, or direct experiments.
- What is theoretical? Plausible claims that fit current methods but are not yet established for routine use.
- What is beyond current capability? Claims not supported by credible evidence with current systems.
- What would make this more promising? The experiment, benchmark, field study, validation cohort, or regulatory evidence that would make the area stronger or reveal its limits.
- What should researchers, biotech teams, funders, and program leaders do with this? The action layer: how to evaluate, adopt, defer, or test the capability.
The headings are phrased as questions because that is how readers actually use the book. The point is not to memorize labels. The point is to leave each chapter knowing what is real, what remains speculative, what is not yet possible, what would make the area stronger, and what action follows.
Citations are checked against Crossref, PubMed, publisher records, or official organizational sources where available. Citation verification matters because hallucinated DOIs and author-attribution errors have damaged reputations in adjacent fields. Where appropriate, both the original preprint and the peer-reviewed version are cited. Where a system remains restricted-release (AlphaFold 3 server access, AlphaProteo arXiv preprint), the restriction is named explicitly so the reader can weight the evidence accordingly.
How This Handbook Is Updated
Continuously. Major papers and benchmark results land in Nature, Science, Cell, and adjacent venues at a steady cadence; frontier-lab releases are frequent. The handbook is maintained as a working reference rather than as a frozen edition. Each chapter carries the date of last substantive update. When a new system supersedes an older one, the new system is cited and the older one retained as historical context.
The same updating discipline applies to corrections. When a citation is found to be wrong (a chimera, a misattribution, a journal mix-up), the correction is made and noted. When an independent evaluation changes the evidence weight of a capability claim, the chapter is updated.
What This Handbook Is Not
It is not a tutorial on machine learning. It is not a textbook of biology. It is not a guide to clinical practice, public health operations, biosafety, farm operations, environmental permitting, or regulatory submissions. It is not an endorsement of any vendor or platform. It is not a frozen reference to a specific moment in the field.
It is a working reference for evaluating life sciences AI claims against experimental evidence, with a scale-based structure that can absorb new model classes without changing the book structure.
Acknowledgments
Thanks to the open-source communities that make the underlying work possible: Crossref and the rest of the citation infrastructure, the Rosetta Commons and Baker lab ecosystem, the EvolutionaryScale and Meta FAIR ESM lineage, the Arc Institute Evo team, the Chan Zuckerberg Initiative CELLxGENE and Virtual Cells Platform, and the many academic groups whose papers form the backbone of the chapters that follow.
Thanks also to the researchers whose independent-evaluation work (Ahlmann-Eltze, Boiarsky, Buttenschoen, Luecken, Walsh and colleagues) keeps the field honest. Every serious reference on life sciences AI benefits from their work.
Bryan Tegomoh, MD, MPH
May 2026