# AI-Designed Biology

Canonical URL: https://menfem.com/kb/genomics/concepts/ai-designed-biology
Knowledge base topic: [Genomics](https://menfem.com/kb/genomics)
Frontier status: breakthrough
Tags: ai-genomics, generative-biology, protein-language-models, crispr, opencrispr, de-novo-design

---

AI-designed biology is the generation of functional biomolecules entirely from scratch using protein language models trained on evolutionary diversity — not discovery, not engineering of natural variants. The flagship result is **OpenCRISPR-1** (Profluent, Nature 2025), the first fully AI-designed gene editor to successfully edit the human genome. Every component — nuclease, guide RNA, deaminase — was generated computationally. OpenCRISPR-1 is 403 mutations away from SpCas9 and 182 mutations away from the nearest natural variant, yet it matches SpCas9's on-target efficiency (55.7% vs 48.3%) while cutting off-target editing by 95% (0.32% vs 6.1%).

This is adjacent to but distinct from [AI-Genomics Convergence](./ai-genomics-convergence.md) (CRISPR-GPT as a copilot for experiment design). CRISPR-GPT helps a researcher use an existing editor; OpenCRISPR-1 *is* a new editor that an AI created. The first category accelerates human work; the second replaces the discovery process with generation.

The platform implication is larger than CRISPR. Profluent trained on the CRISPR-Cas Atlas (5.1M proteins, 2.7× natural diversity expansion) and produced 4M candidate sequences. The same approach generalizes to any protein family with enough training data — kinases, antibodies, enzymes for industrial catalysis, receptors for drug discovery. The 2024 public release of OpenCRISPR-1 created a precedent: AI-designed biomolecules can be open-sourced, raising both democratization upside and dual-use-risk concerns.

## Key Claims

- **First fully AI-designed gene editor to edit the human genome** — every component (nuclease, gRNA, deaminase) is AI-generated, not engineered. *Evidence: strong* ([OpenCRISPR-1 paper](../../raw/opencrispr-1-ai-designed-gene-editor.md))
- **Matches SpCas9 on-target** — 55.7% vs 48.3% editing efficiency. *Evidence: strong* ([OpenCRISPR-1](../../raw/opencrispr-1-ai-designed-gene-editor.md))
- **95% reduction in off-target editing** — 0.32% vs 6.1% — the safety margin that makes this more than a proof-of-concept. *Evidence: strong* ([OpenCRISPR-1](../../raw/opencrispr-1-ai-designed-gene-editor.md))
- **403 mutations from SpCas9, 182 from the nearest natural variant** — operates in genuinely novel sequence space. *Evidence: strong* ([OpenCRISPR-1](../../raw/opencrispr-1-ai-designed-gene-editor.md))
- **CRISPR-Cas Atlas: 5.1M proteins** — 2.7× natural diversity expansion; 4.1× for Cas9 specifically. *Evidence: strong* ([OpenCRISPR-1](../../raw/opencrispr-1-ai-designed-gene-editor.md))
- **Open-sourced sequence** — public release in April 2024, prior to Nature publication. Tens of thousands of researchers accessed it. *Evidence: strong* ([OpenCRISPR-1](../../raw/opencrispr-1-ai-designed-gene-editor.md))
- **Profluent raised $106M, Nov 2025** — market validation of the platform thesis. *Evidence: moderate* ([OpenCRISPR-1](../../raw/opencrispr-1-ai-designed-gene-editor.md))

## Benchmarks & Data

| Metric | OpenCRISPR-1 | SpCas9 |
|---|---|---|
| On-target editing | **55.7%** | 48.3% |
| Off-target editing | **0.32%** | 6.1% |
| Distance from SpCas9 | 403 mutations | — |
| Distance from nearest natural variant | 182 mutations | — |

| Training Data | Scale |
|---|---|
| CRISPR-Cas Atlas | 5.1M proteins |
| Cas9 natural variants | 238,917 |
| Generated candidates | 4M |
| Metagenome data mined | 26 terabases |

## Open Questions

- Does the in vitro 95% off-target reduction hold up in vivo?
- Will regulators treat fully AI-designed editors as Class III novel entities requiring separate pre-clinical packages, or as equivalent to engineered naturals?
- Does public release of AI-designed biomolecules accelerate science more than it accelerates dual-use risk?
- Which protein family is next? (Zinc fingers, TALENs, base editors, deaminases, prime editing pegRNAs?)
- What's the ceiling — can a model trained on all known proteins *design a new class* of editor that doesn't exist in nature?

## Related Concepts

- [AI-Genomics Convergence](./ai-genomics-convergence.md) — CRISPR-GPT is the copilot complement; OpenCRISPR-1 is the creation counterpart
- [CRISPR Clinical Translation](./crispr-clinical-translation.md) — the trajectory AI-designed editors must cross

## Backlinks

*Pages that reference this concept:*
- [OpenCRISPR-1 paper](../../raw/opencrispr-1-ai-designed-gene-editor.md)
- [Profluent Bio](../entities/profluent-bio.md)

## Changelog

- **2026-04-17** — Initial compilation from the Profluent Nature 2025 paper.

## Sources

- opencrispr-1-ai-designed-gene-editor

---

Cite as: MenFem Knowledge Base — https://menfem.com/kb/genomics/concepts/ai-designed-biology