---
title: "Executive Summary – The Life Sciences AI Handbook"
source: "https://lifesciencesaihandbook.com/executive-summary.html"
author: "Bryan Tegomoh, MD, MPH"
last_updated: "2026-10-20"
doi: "10.5281/zenodo.22073135"
license: CC BY 4.0
generated-by: handbook-markdown-export
---

# Executive Summary

Life sciences AI now reaches beyond isolated modeling tasks into biological discovery, translational research, and research operations. Protein structures are predicted before they are solved. Cells are represented by foundation models. Organismal and environmental systems are becoming model targets. Generative systems can propose more molecules than a laboratory can synthesize and test. Autonomous chemistry runs closed loops on robotic platforms. The practical risk extends beyond overclaiming to using the wrong evidence for the wrong decision.

## Core Takeaways

**Protein structure prediction is a routine research input.** AlphaFold 2 produced a marked improvement on many single-chain targets at CASP14, with accuracy competitive with experimental structures for much of the benchmark ([Jumper et al., 2021](https://doi.org/10.1038/s41586-021-03819-2)). AlphaFold 3 extended prediction to biomolecular interactions involving nucleic acids, ions, and small-molecule ligands ([Abramson et al., 2024](https://doi.org/10.1038/s41586-024-07487-w)). DeepMind released AlphaFold 3 inference code in November 2024 ([Abramson et al., 2024, addendum](https://doi.org/10.1038/s41586-024-08416-7)). Its downloadable weights are limited to noncommercial use by noncommercial organizations ([Google DeepMind, AlphaFold 3 Model Parameters Terms of Use](https://github.com/google-deepmind/alphafold3/blob/main/WEIGHTS_TERMS_OF_USE.md)); commercial access is offered through Google Cloud ([Google DeepMind, AlphaFold 3 repository README](https://github.com/google-deepmind/alphafold3/blob/main/README.md)). Boltz and Chai released separately developed models ([Wohlwend et al., 2024, preprint](https://doi.org/10.1101/2024.11.19.624167); [Chai Discovery et al., 2024, preprint](https://doi.org/10.1101/2024.10.10.615955)). Structure prediction is now routine research infrastructure, but confidence scores and experimental validation determine which downstream claims are defensible.

**Protein design is an experimental discipline supported by generative models.** RFdiffusion generates backbones ([Watson et al., 2023](https://doi.org/10.1038/s41586-023-06415-8)). ProteinMPNN and LigandMPNN design sequences ([Dauparas et al., 2022](https://doi.org/10.1126/science.add2187); [Dauparas et al., 2025](https://doi.org/10.1038/s41592-025-02626-1)). ESM3, Chroma, and ProGen2 add multimodal, diffusion-based, and autoregressive generative approaches ([Hayes et al., 2025](https://doi.org/10.1126/science.ads0018); [Ingraham et al., 2023](https://doi.org/10.1038/s41586-023-06728-8); [Nijkamp et al., 2023](https://doi.org/10.1016/j.cels.2023.10.002)). RFantibody extends the lineage to peer-reviewed de novo antibody design ([Bennett et al., 2026](https://doi.org/10.1038/s41586-025-09721-5)). AlphaProteo from DeepMind reports strong binder generation but remains an arXiv preprint with restricted code ([Zambaldi et al., 2024, preprint](https://arxiv.org/abs/2409.08022)). Every designed protein still needs expression, purification, characterization, functional assay, and developability review before it counts as a candidate.

**Single-cell foundation models do part of what their papers claim.** scGPT, Geneformer, scFoundation, and scBERT produce reusable cell-state representations ([Cui et al., 2024](https://doi.org/10.1038/s41592-024-02201-0); [Theodoris et al., 2023](https://doi.org/10.1038/s41586-023-06139-9); [Hao et al., 2024](https://doi.org/10.1038/s41592-024-02305-7); [Yang et al., 2022](https://doi.org/10.1038/s42256-022-00534-z)). They support cell-atlas search and selected integration tasks. Nicheformer provides peer-reviewed evidence that pretraining on spatial as well as dissociated transcriptomic data can yield cell representations that capture spatial context ([Tejada-Lapuerta et al., 2025](https://doi.org/10.1038/s41592-025-02814-z)). Independent evaluations found that simple baselines can match these models: logistic regression matched or outperformed scBERT on cell-type annotation, and under common metrics none of seven deep perturbation models outperformed deliberately simple baselines such as PCA plus linear regression ([Boiarsky et al., 2024](https://doi.org/10.1038/s42256-024-00949-w); [Ahlmann-Eltze et al., 2025](https://doi.org/10.1038/s41592-025-02772-6)). A 2026 analysis argues that those metrics are poorly calibrated and that deep perturbation models can outperform the baselines under calibrated metrics ([Miller et al., 2026](https://doi.org/10.1038/s41587-026-03307-w)). The evidence supports single-cell foundation models for some tasks and leaves others unverified.

**Genome foundation models are early and active.** Evo demonstrated sequence modeling across DNA, RNA, and proteins ([Nguyen et al., 2024](https://doi.org/10.1126/science.ado9336)). Evo 2 extended genome modeling across all domains of life, with open model parameters, code, and OpenGenome2 data ([Brixi et al., 2026](https://doi.org/10.1038/s41586-026-10176-5)). Nucleotide Transformer, GET, Orthrus, and AlphaGenome show the field splitting into genome-wide, transcriptional, RNA-specific, and regulatory-variant models ([Dalla-Torre et al., 2025](https://doi.org/10.1038/s41592-024-02523-z); [Fu et al., 2025](https://doi.org/10.1038/s41586-024-08391-z); [Fradkin et al., 2026](https://doi.org/10.1038/s41592-026-03064-3); [Avsec et al., 2026](https://doi.org/10.1038/s41586-025-10014-0)). Production-quality variant interpretation still happens inside ACMG/AMP and ClinGen frameworks, with computational predictors as evidence inputs rather than classifications ([Richards et al., 2015](https://doi.org/10.1038/gim.2015.30); [Pejaver et al., 2022](https://doi.org/10.1016/j.ajhg.2022.10.013)).

**Systems-biology claims require causal discipline.** Cell-state embeddings, pathway scores, and network diagrams become useful when they name a mechanism and a falsifying perturbation. Gene regulatory network inference and whole-cell modeling are established research traditions. Observational edges are not causal maps. Virtual-organism claims are not larger virtual-cell claims; they require development, physiology, tissue coupling, environment, time, and prospective validation.

**Organismal and environmental biology belong in the core scope.** Neuroscience AI, aging clocks, plant and crop AI, ecological monitoring, environmental DNA, and virtual organisms occupy the scale rung between cellular biology and translation. The evidence problem is measurement validity: whether the model is learning biology or platform artifacts, geography, sampling effort, season, field conditions, or cohort structure.

**AI for therapeutics is delivering candidates, not proof of improved approval probability.** Rentosertib reached a randomized Phase 2a trial whose primary endpoint was safety and tolerability; forced vital capacity was a secondary endpoint, and the study was too small and short to establish registration-level benefit ([Xu et al., 2025](https://doi.org/10.1038/s41591-025-03743-2)). A 52-week Phase 3 trial with forced vital capacity decline as its primary endpoint began in China in September 2026 ([NCT07687459](https://clinicaltrials.gov/study/NCT07687459)). Public evidence reviewed on October 5, 2026 did not identify an FDA- or EMA-approved small molecule whose molecular discovery was attributed to generative AI. Comparative evidence that AI improves clinical success rates, attrition, or total program cost remains insufficient.

**Translation is broader than small molecules.** Chemical biology, target engagement, cell and gene therapy, diagnostics, biomarkers, trials, and real-world evidence each have different validation objects. A generated molecule, engineered cell, biomarker signature, or diagnostic feature becomes useful only when the evidence matches the decision it is supposed to support.

**Self-driving laboratories and agentic systems work for bounded tasks, but they are not the same category.** Self-driving laboratories close a machine-directed experimental loop through robotic execution. Coscientist and A-Lab provide chemistry and materials examples ([Boiko et al., 2023](https://doi.org/10.1038/s41586-023-06792-0); [Szymanski et al., 2023](https://doi.org/10.1038/s41586-023-06734-w)). Virtual Lab, Co-Scientist, and Robin are agentic or lab-in-the-loop research systems in which people perform or authorize key experimental steps ([Swanson et al., 2025](https://doi.org/10.1038/s41586-025-09442-9); [Gottweis et al., 2026](https://doi.org/10.1038/s41586-026-10644-y); [Ghareeb et al., 2026](https://doi.org/10.1038/s41586-026-10652-y)). Open-ended autonomous biology outside bounded settings remains beyond current capabilities.

**Information hazards are handled through deliberate disclosure and current institutional review.** The standard is enough detail for scientific verification without unnecessary operational detail that raises misuse risk. The July 2026 USG Policy for Stopping High-Risk Life Sciences Research replaced the 2024 policy on dual use research of concern and pathogens with enhanced pandemic potential, prohibited federal support for defined dangerous gain-of-function research, established international-research restrictions, and required agency implementation guidance ([White House, 2026](https://www.whitehouse.gov/wp-content/uploads/2026/07/USG-Policy-for-Stopping-High-Risk-Life-Sciences-Research_July-2026.pdf)). NIH states that previously paused activities remain paused until NIH-specific requirements are established ([NIH, 2026](https://grants.nih.gov/grants/guide/notice-files/NOT-OD-26-101.html)).

**Biosecurity evaluation requires more than a capability benchmark.** Scientific utility, hazardous capability, human or agent uplift, safeguard effectiveness, operational consequence, and lifecycle governance are different claims. A public benchmark can support one of them without establishing the others. Release decisions should therefore use configuration-specific testing, matched baselines, statistical uncertainty, legitimate-use evaluation, adversarial safeguard testing, independent review, and post-deployment monitoring. The Benchmarks for Bio AI chapter provides the evaluation framework.

**Reproducibility is computational and experimental.** DOME is a useful community checklist for reporting supervised machine learning in biology; it is not a universal professional standard ([Walsh et al., 2021](https://doi.org/10.1038/s41592-021-01205-4)). Model cards and dataset cards add documentation. Bridge2AI and the announced ARPA-H IGoR program provide federal infrastructure examples ([NIH Common Fund Bridge2AI, 2026](https://commonfund.nih.gov/bridge2ai); [ARPA-H IGoR, 2026](https://arpa-h.gov/explore-funding/programs/igor)). A notebook that reruns is necessary but not sufficient if the assay cannot be repeated.

**Workforce and compute are institutional, not individual.** A credible team is cross-disciplinary: biological expertise, data engineering, machine learning, wet-lab partnership, regulatory engagement when relevant, and governance. Compute without experimental judgment creates expensive noise.

## Reading Rule

Every major capability claim in the handbook lands in one of three tiers:

- **Demonstrated** for claims supported by peer-reviewed evidence, official documentation, or reproducible blind benchmarks.
- **Theoretical** for claims plausible under current methods but not yet established for routine use.
- **Beyond current capabilities** for claims not supported by credible evidence with current systems.

If a claim cannot be placed in a tier, the claim is not specific enough.

## Practical Use

The handbook is built for these uses:

- Evaluating a model claim before adopting it in a research program.
- Designing a validation plan that matches the biological decision the model informs.
- Briefing a research team on the current evidence for a given model class.
- Reviewing a vendor pitch against the published evidence in peer-reviewed venues.
- Deciding whether a newly published paper changes a program decision.

It is a reference for working biologists, biotechnology leaders, physician-scientists, drug-discovery scientists, synthetic biologists, plant biologists, ecologists, neuroscientists, aging researchers, graduate students, and research program leaders who need to evaluate AI systems against experimental evidence. It does not teach machine learning.

## The One-Page View

The capability frontier is real and uneven. Structure prediction and selected design tasks are demonstrated. Single-cell representation works; single-cell perturbation prediction is contested. Genome foundation models are early. Systems biology and virtual-organism claims require stronger perturbation evidence. Plant, ecological, neuroscience, and aging AI are now large enough to deserve explicit homes. AI for therapeutics is delivering candidates into trials but not yet approvals. Autonomous laboratories work for bounded tasks. Information hazards require institutional discipline. Reproducibility requires infrastructure investment.

The value of life sciences AI depends on more than the model. It depends on the validation plan, the wet-lab follow-up, the calibration discipline, and the willingness to treat every model output as a hypothesis with explicit uncertainty.
