I Gave Claude Science One Prompt. It Ran a Full Spatial Analysis.

By Lociven · SpatiaBio · July 2, 2026

I gave Claude Science AI Workbench — Anthropic's new scientific analysis platform — a single prompt and a dataset. Thirty minutes later, it handed me five publication-quality figures, a fully executable Jupyter notebook, and a reproducibility report.

---

What is Claude Science AI Workbench?

Released in June 2026, Claude Science is Anthropic's platform for running agentic scientific analysis inside a sandboxed Linux environment. You describe what you want in plain language. The agent installs packages, writes code, executes it, fixes errors, and returns figures and notebooks — all without you touching a terminal.

For spatial transcriptomics researchers, the pitch is simple: skip the environment setup, skip the debugging, get directly to the biology.

The prompt I used

Analyze the Squidpy IMC breast cancer dataset (sq.datasets.imc()):
1. Spatial neighbors graph — Delaunay triangulation
2. Neighborhood enrichment (sq.gr.nhood_enrichment, n_perms=1000)
3. Co-occurrence as a function of distance
4. Interaction matrix and centrality scores
5. Ripley's L for clustering vs. randomness

Use publication-grade conventions. Return a Jupyter notebook with all outputs embedded.

That's it. No code. No conda commands. The agent created a spatial conda environment (squidpy 1.8.2, scanpy 1.11.5), downloaded the Jackson et al. breast cancer IMC dataset (4,668 cells × 34 protein markers, 11 cell types), ran the full pipeline, and produced the figures below.

---

Figure 1 — Cell types in situ

The IMC dataset captures 11 cell types across a breast cancer tissue section. Apoptotic tumor cells (cyan) dominate numerically and are distributed throughout the tissue.

Generated with Claude Science AI Workbench (Anthropic, 2026) · squidpy 1.8.2 · Jackson et al., Nature 578, 2020

Figure 2 — Neighborhood enrichment: which cell types co-locate?

The permutation-based z-scores reveal the tissue's immune architecture at a glance. Three patterns stand out:

Immune clustering (z = +29 to +36): T cells, macrophages, and stromal cells form a tightly co-located immune compartment — the canonical TIL niche.

Tumor-immune avoidance (z = −21 to −28): Apoptotic tumor cells strongly avoid macrophages. Immune exclusion at the single-cell spatial level.

Tumor self-clustering (z = +33 to +48): Each tumor subtype clusters with itself, consistent with clonal expansion in distinct spatial niches.

Generated with Claude Science AI Workbench (Anthropic, 2026) · squidpy 1.8.2 · Jackson et al., Nature 578, 2020

Interpretation note: A z-score above +10 is conventionally significant with n_perms=1000. Values of +29 to +48 indicate extremely non-random co-localization.

Figure 3 — Co-occurrence: the spatial scale of immune clustering

Co-occurrence shows at what distance cell types interact. The steep decay curves confirm that immune clustering is a contact-range phenomenon — not a tissue-wide gradient. Macrophages are enriched 7× near other macrophages at minimal distance, then fall sharply.

Generated with Claude Science AI Workbench (Anthropic, 2026) · squidpy 1.8.2 · Jackson et al., Nature 578, 2020

Figure 4 — Graph centrality: who connects the tissue?

Apoptotic tumor cells have degree centrality 0.83 and closeness centrality 0.84 — nearly 5× higher than any other cell type. They are physically positioned at the crossroads of the tissue, spatially interleaved with every other population.

Generated with Claude Science AI Workbench (Anthropic, 2026) · squidpy 1.8.2 · Jackson et al., Nature 578, 2020

Figure 5 — Ripley's L: clustering vs. complete spatial randomness

Apoptotic tumor cells spike massively above the 95% CSR envelope across all distances — extreme, scale-independent clustering. Combined with the centrality result: self-clustering + network centrality is characteristic of a dominant tumor clone that has physically reorganized the tissue architecture.

Generated with Claude Science AI Workbench (Anthropic, 2026) · squidpy 1.8.2 · Jackson et al., Nature 578, 2020

---

What Claude Science actually did (under the hood)

The agent's execution log was visible in real time. It:

1. Created a conda environment from scratch (squidpy==1.8.2, scanpy==1.11.5)

2. Downloaded and validated the IMC dataset (4,668 cells × 34 markers)

3. Ran all five Squidpy analyses with appropriate parameters

4. Noticed its own figure title was inaccurate and self-corrected it before saving

5. Authored a 23-cell reproducible notebook with embedded outputs

6. Bundled the dataset, requirements.txt, and README automatically

The self-correction on the Ripley's L title is worth noting. It cross-checked the figure against the data and caught a misleading generalization. That's not what most analysis scripts do.

---

Bonus: the same pipeline on Visium data

The IMC analysis above uses point-cloud coordinates. I reran the same prompt on a 10x Genomics Visium section (grid-based spots, 55 µm resolution). The agent automatically switched to coord_type="grid" and n_neighs=6 for the hexagonal lattice without being instructed to — it inferred this from the data format.

Figure 6 — Visium: spatial clusters

On the Visium hexagonal grid, cluster boundaries directly reflect anatomical structure. The sharper compartmentalization here vs. the IMC data is expected: Visium spots are physically constrained to a grid, so spatial patterns emerge more cleanly at tissue scale.

Generated with Claude Science AI Workbench (Anthropic, 2026) · squidpy 1.8.2 · 10x Genomics Visium

Figure 7 — Visium: gene expression overlay

Gene expression mapped onto spatial coordinates reveals domain-specific marker gradients. Unlike IMC (protein-level), Visium captures transcriptomic heterogeneity at spot resolution. Claude Science generated both the continuous expression overlay and the categorical cluster map in a single run.

Generated with Claude Science AI Workbench (Anthropic, 2026) · squidpy 1.8.2 · 10x Genomics Visium

Figure 8 — Visium: neighborhood enrichment

The neighborhood enrichment heatmap on Visium shows a clean tissue-layer organization: adjacent anatomical compartments co-enrich (positive z-scores), distant layers avoid each other (negative z-scores). You can trace these directly back to the cluster map in Figure 6.

Generated with Claude Science AI Workbench (Anthropic, 2026) · squidpy 1.8.2 · 10x Genomics Visium

The full Visium pipeline — grid graph parameters, spatial autocorrelation, and multi-sample batch correction — is in Pack 1 notebooks 03, 06, and 06b.

Honest assessment: is it useful?

✓ Works well for

• Standard Squidpy pipelines

• Exploratory analysis on new datasets

• Generating a reproducible baseline notebook

• Researchers new to spatial omics

• Rapid figure drafts for lab meetings

✗ Limitations

• Requires manual approval for each tool call

• Can't customize beyond what you describe

• No Visium HD / large dataset chunking

• Figure aesthetics are functional, not polished

• Session restarts lose progress

Bottom line: For getting from raw data to interpretable spatial figures in one session without writing code, it genuinely works. For publication-level customization, you still need to go hands-on. Pack 1 covers exactly that layer.

---

SpatiaBio Pack 1

Squidpy Foundations — 16 Notebooks ($19)

Everything Claude Science did above — plus memory optimization for Visium HD, batch correction across samples, ligand-receptor analysis, and Nature-style publication figure templates.

Get it for $19 →

Sister blog

The biology behind the cells you're mapping

NeoantigenLab covers neoantigen biology, HLA typing, pVACseq, and cancer immunotherapy — the immunology context for what spatial analysis reveals.

Visit NeoantigenLab →