Claude CRISPR-Like Enzyme Find: Mine Genomes Yourself

The #1 mistake after Claude discovers a novel enzyme system with CRISPR-like repeats

Mid-late September 2026, the post landed hard: Claude agents had flagged array-associated reverse transcriptases (ART) in bacteriophage DNA – an RT, a partner gene, and evenly spaced repeats that look CRISPR-adjacent. Plenty of people opened Claude and typed “find me the next CRISPR.”

That’s the mistake. Nobody one-shot a gene editor. Anthropic ran a multi-day, multi-agent genome-mining campaign on a narrow brief; humans handled wet-lab follow-up. Steal the workflow. Don’t treat the headline like a magic prompt, or you burn tokens on polished nonsense.

Why single-prompt “discovery” and news-only reading fall short

The bottleneck was never “can we BLAST proteins.” Classic mining already builds HMMs, walks huge clusters, scores neighbors. Humans stall on weird flanks. Scaling that attention is the actual win – not a brand-new bench method overnight.

Per Anthropic’s announcement (23 Sep 2026), the campaign sat around 950 agents, ~21 hours, ~210 million tokens: ~200k RT clusters, a few thousand scored partner families, then a short stack of human-readable reports. Existing write-ups recycle that funnel and the viral “tandem repeat array… CRISPR-like” agent line. They skim past what breaks on a laptop-sized rerun: path dependence, access tiers, and unknown biology.

If you only read the hype, you miss the brief. It asked for novel partner genes next to RTs. The CRISPR-like repeats were a side observation – after a worker dumped a bad partner hypothesis and still opened the raw flank DNA.

Recommended approach: reverse-engineer the ART campaign for your laptop

The same surfaces Anthropic points scientists at – Claude Science and Claude Code – plus, on their side, a multi-agent setup you won’t fully clone. You will not chew 1.9 billion protein clusters on a laptop. You can copy the logic on a public slice.

Step 1 – Scope a survey brief like theirs

Write a short research brief, not a chatty wish. Shape:

Goal: Survey reverse transcriptases (or another family) in public phage/metagenomic data.
1) Reproduce a known RT classification from literature on a small public set.
2) Build or reuse sequence profiles; cluster hits.
3) For each unusual locus, pull ±2-5 kb flanks as raw DNA text.
4) Score recurring neighbor ORFs as candidate partners.
5) Flag non-coding anomalies (tandem repeats, conserved spacers).
6) For top 5 loci, write a 1-page report: evidence, literature check, what would falsify it.
Reject annotation artifacts. Do not claim function without evidence.

Force flank DNA into context. The ART moment wasn’t a vibes-based “aha.” The agent read sequence as text, counted repeats, measured spacing, checked prior reports, then filed for humans.

Step 2 – Run it where code + databases meet

Claude Science beta: Pro/Max can start; Team/Enterprise need an admin toggle. Point it only at folders you approve. Turn on the genomics connectors you actually need. Leave provenance on. Prefer Claude Code? Drop FASTAs, HMMER/MMseqs scripts, and a notes folder into a repo and work there.

• Tiny public set first – a few phage genomes or a filtered MGnify slice – so every call is auditable.
• Make the agent re-derive one known result before any “novel” claim (method check).
• Split workers (search, flank parse, lit check) from a supervisor pass that kills weak candidates.

Most candidates die on second look. Normal. Thousands of scored families collapsed to a handful of partner stories and a few RT lineages worth writing up.

Step 3 – Budget for non-determinism

The catch is reproducibility theater. Ten identical campaign reruns: zero recovered the ART array, because no agent opened the critical upstream DNA. Hand capable models that DNA directly and recognition jumps – write-ups of the preprint cite ≥90%. Make them open files and drive the tools themselves and success can crater (figures as low as ~32%).

Pro tip: Encode the step. “For every shortlisted RT, fetch flanks and print a windowed sequence with a simple repeat detector before you write the report.” Do not wait for spontaneous curiosity.

Step 4 – Human gate before any “discovery” language

ART’s underlying RT was already in earlier papers. What got elevated was the system framing: array + accessory protein. Early bench work (as of the 23 Sep 2026 disclosure) shows the array as distinct short RNAs. Primary function? Still open. RT activity on those RNAs? Not shown. Programmable cut/copy/paste? Not established – even though only a handful of known systems share this feature set and those ones do perform DNA operations. Feng Zhang’s line after reading the preprint: “genuinely intriguing,” worth more work. Not a finished tool.

Agents propose. You – or a biologist – dispose. No wet lab, no claim.

Real mini-example: one locus, one afternoon

Grab a public jumbo-phage contig. Paste something like this into Claude Science:

Load contig X. Find ORFs with RT domains (PF00078 or equivalent).
For the top hit, extract 3 kb upstream as plain sequence.
1) Run a tandem-repeat scan (period 10-40 nt, ≥3 copies).
2) List nearest downstream ORFs and BLAST them lightly.
3) Compare architecture to retrons and CRISPR arrays in 3 review papers.
4) Output: table of repeat units + spacers, neighbor table, and a skeptical paragraph on whether this is a system or junk.

Output will feel like Anthropic’s intermediate artifact, not a journal paper. Good. Remember the flank mandate from Step 3? Here’s where it stops being optional. Scale only after the report format is boringly reliable.

It’s odd how much of modern biotech still starts with someone squinting at letters in a row. Restriction enzymes, Taq, CRISPR repeats – same habit, bigger hard drives. Agents make the squinting cheaper. They don’t make it optional.

Access, cost, and honest limits

Claude Science runs on your plan’s models. It is not a free Mythos enable. Deep biology with relaxed classifiers sits behind the Life Sciences Verification Program (announced 17 Sep 2026): verified Team/Enterprise orgs, Standard vs High-risk grants. Individual Pro/Max expansion was planned later – not at launch. Cap parallel workers. Write intermediate CSVs. Even a toy multi-session survey adds up when the reference run lived in the hundreds-of-millions-of-tokens neighborhood.

Genome mining isn’t new. Some CRISPR veterans read ART as a solid candidate report, not CRISPR 2.0. The technical write-up is a preprint PDF, not peer-reviewed gospel. Treat it like a methods notebook.

Two hundred million tokens for one careful survey. On your dime, that number should make you flinch – and then write a tighter brief.

Park that table next to any PI demo or blog draft. It blocks oversell.

FAQ

Did Claude invent a new CRISPR tool?

No. Phage RT system with a CRISPR-like repeat layout. Not a shipping editor.

Can a beginner reproduce something like the ART search?

Yes – the workflow, on a small public set in Claude Science or Claude Code: brief → profile search → mandatory flank read → partner/repeat flags → skeptical one-pager. Start with one genome family and one accession. You will not casually spin 950 agents across 1.9B clusters. Expect most ideas to die in review.

Why did reruns miss the array if the model “saw it by eye”?

Sequence already in context ≠ choosing to fetch that locus among thousands of candidates. The successful path included a follow-up after rejecting a partner story; identical high-level campaigns skipped that branch. Process bug. Fix with checklists: required flank windows, required repeat scan, required falsification line. Vibes won’t encode anomaly detection.

Next action: install Claude Science (or open Claude Code on a phage FASTA), paste the survey brief with one concrete accession, and don’t stop until the one-page report states what would falsify the hit.