Executive Summary

Life sciences AI has crossed from isolated modeling tasks into the operating layer of biomedical discovery. Protein structure prediction, protein design, genome models, single-cell foundation models, molecular generation, self-driving laboratories, and agentic research workflows now affect how teams choose experiments.

The practical risk is not only overclaiming. It is using the wrong evidence for the wrong decision.

Core Takeaways

• Protein structure prediction is now a routine research input for many proteins, but structure is not function, mechanism, safety, or clinical value.
• Protein design and antibody design are experimental disciplines. Designed sequences need expression, binding, specificity, stability, immunogenicity, and manufacturability review.
• Genome and variant models are strongest when tied to measured functional genomic outputs. Clinical or organism-level claims require additional evidence.
• Therapeutics AI adds value when it improves a decision in target selection, chemistry, trial design, or evidence generation. It does not remove attrition.
• Cellular AI is moving from annotation toward perturbation prediction. Virtual cell claims should be read as task-specific unless proven otherwise.
• Self-driving labs and agentic workflows require protocol standards, instrument reliability, provenance, and human authorization gates.
• Information hazards are handled through deliberate disclosure, reproducibility planning, and review of release details.

Reading Rule

Every major capability claim in the handbook should land in one of three tiers: Demonstrated, Theoretical, or Beyond current capabilities.

Practical Use

Use this handbook to evaluate a model, design a validation plan, brief a research team, review a vendor claim, or decide whether a paper changes a program decision.