Appendix B — TL;DR Compilation
AI for the Life Sciences
Life sciences AI is not one field. It is a set of modeling practices that share biological data constraints, experimental validation requirements, and high error costs. The first rule is to ask what biological object the model represents and what experiment would falsify the output.
Biological Data Infrastructure
Better models do not rescue poorly specified biological data. AI-ready data require provenance, assay context, versioning, licensing, and negative controls. The most useful model card is often a dataset card.
Foundation Models for Biology
A foundation model is useful when pretraining improves a downstream biological task under realistic validation. Model size, modality count, and dataset volume matter less than task transfer, assay fidelity, and external testing.
Evaluation Principles for Biomedical Discovery AI
The core evaluation question is not whether a model performs well on held-out rows. The core question is whether it improves a real experimental decision under the distribution where it will be used.
Protein Structure Prediction
Structure models are now routine inputs to biology, but they are not substitutes for experiments. Confidence, conformational state, ligand geometry, and biological context determine whether a predicted structure supports a downstream decision.
Protein Design and Engineering
Protein design is strongest when the target is structurally specified and the success assay is direct. Claims become weaker as design moves from fold, to binding, to catalysis, to cellular phenotype.
Antibody and Biologic Design
AI design methods help generate and prioritize binders. Therapeutic biologics still require assay cascades, liability screening, cell-based testing, and manufacturing review.
Nucleic Acid and Genome Models
DNA and RNA models are strongest when the output is tied to measured functional genomic assays. Variant interpretation remains difficult when disease mechanism, cell context, and long-range regulation are uncertain.
Variant Effect Prediction
Variant models help prioritize variants and hypotheses. They do not replace segregation evidence, functional assays, population frequency, disease mechanism, and clinical interpretation.
Target Identification and Prioritization
AI-assisted target selection is useful when it integrates evidence transparently. The winning target is not the top-ranked node. It is the target with a testable mechanism, feasible modality, safety rationale, and disease-relevant assay path.
Small Molecule Generation and ADMET
The useful output is not a molecule that looks novel. The useful output is a prioritized set of compounds with rationale, feasibility, assay plan, and acceptable risk across potency, selectivity, ADMET, and chemistry.
mRNA, RNA, and Vaccine Design
RNA and vaccine AI is strongest when the model output is tied to a measurable endpoint: expression, stability, antigenicity, manufacturability, or immune response. Program success still depends on delivery, dosing, safety, and clinical evidence.
Clinical Trial AI for Translational Research
AI in trials is safest when the context of use is explicit. Recruitment support, site selection, endpoint extraction, enrichment, and synthetic controls carry different evidentiary and regulatory burdens.
Translational Evidence and Failure Modes
A model that improves a proxy endpoint may still harm the program if the proxy is poorly linked to disease biology or developability. Failure analysis belongs near the start of the workflow, not after candidate nomination.
Single-Cell Foundation Models
Single-cell foundation models are useful representation systems, not general virtual cells. Evaluation must account for cell type, donor, batch, disease state, and perturbation split.
Spatial Omics and Tissue Models
Spatial AI is most useful when it links molecular signals to tissue structure with clear resolution limits. It is not enough to assign labels to spots or cells. The biological question is whether spatial organization changes mechanism or decision.
Cell Painting and Image-Based Phenotyping
Cell images are rich biological measurements, but morphology is not mechanism by itself. High-content imaging requires careful controls, segmentation quality checks, and orthogonal validation.
Perturbation Prediction and Virtual Cells
Virtual cell work is promising when framed as perturbation prediction for defined outputs. It becomes misleading when a transcriptomic forecast is treated as a full model of the cell.
Microbiome and Multi-Omics AI
Multi-omics models are useful when each modality has clear provenance and the validation endpoint is explicit. Integration can hide weak measurements if the workflow does not track missingness and batch effects.
Self-Driving Laboratories
A self-driving lab is an experimental system with a model in the loop. It needs reliable instruments, machine-readable protocols, calibration, error handling, and human review of objectives and stopping rules.
Robotic Lab Automation and Cloud Labs
Automation improves repeatability only when protocols, reagents, instruments, and data capture are explicit. A robot executing a vague protocol only scales ambiguity.
Synthetic Biology Design Tools
Synthetic biology AI is useful when design output is tied to a build and test plan. Sequence novelty alone is not engineering progress.
Agentic Science Workflows
Agentic systems are useful as research operating layers when tasks are bounded, sources are checked, and lab actions require authorization. They are risky when fluent plans are treated as validated science.
Benchmarks for Bio AI
Benchmarks matter when they are hard to game, close to the intended decision, and paired with failure analysis. A leaderboard is not a validation plan.
Reproducibility and Open Science
Open code and open weights help, but life sciences reproducibility also needs protocol detail, reagent provenance, data versioning, and assay records. Scientific openness and risk management must be handled together.
Information Hazards in Capability Research
Responsible capability research keeps enough detail for verification while avoiding unnecessary operational detail that raises misuse risk. The standard is not secrecy by default. The standard is deliberate disclosure.
Workforce, Compute, and Institutional Readiness
The minimum viable team includes biological domain expertise, data engineering, machine learning, experimental validation, and governance. Compute without experimental judgment creates expensive noise.