Scientific software gets locked in the language it was written in. A bridge like
rpy2 lets an agent call an R tool; it does not let anyone
extend or build on it natively — and the classics go unmaintained as their
language and dependencies move on.
Bio-Babel rebuilds them natively in the target ecosystem, and ships each package with a machine-readable contract so an agent that has never seen it still calls it correctly. Nothing in that is specific to R, or to graphics. R → Python is where we are proving it: seventeen classics so far, each its own repository and its own caretaker.
“And the whole earth was of one language, and of one speech.” Genesis 11:1
Why Babel
The punishment in the story was never difference — a field's diversity of thought is its wealth. It was that ideas could no longer travel. Bioinformatics lives a small version of that: a beautiful tool blooms on one side of the language line, and the half of the field on the other side can reach across for it at best — never build on it.
Callable from another language, but not natively extended or built upon. A bridge gets you
the call and nothing more: you cannot write a new geom against a tool you reach through
rpy2, and you cannot compose it with anything native. The tool is reachable
and still stranded.
It is in the object system, the data containers, the numerical conventions, the ecosystem dependencies. Match every signature and a reimplementation will still run while silently diverging — or run correctly while sitting outside the idioms of the ecosystem it now belongs to.
The rebuild has to behave like the original. And an agent that has never seen it has to
use it correctly. Neither implies the other — a faithful port whose API the model has to
guess at fails just as hard. So every package ships a _biobabel/ contract
alongside the implementation.
The Stack
The visualization stack is the first worked example: R's grid → {gtable, scales} → ggplot2
chain and everything downstream of it, carried into pure Python on a Cairo backend.
No matplotlib. Publication-quality output by construction — and one
self-consistent API for an agent to reason about, instead of four.
Tier 1 · Foundation
R's grid — viewports, grobs, units, the whole device model — reimplemented
on Cairo, plus the scales transform/break/palette machinery and the
gtable layout algebra. Everything above stands on these three.
grid_py261 symbolsscales226 symbolsgtable_py21 symbolsTier 2 · Grammar
geoms, stats, scales, facets, coords, themes and guides — composed with +,
exactly as in R. The largest single contract in the project, and the one an agent is most
likely to hallucinate against.
ggplot2_py626 symbols · 18 concepts · 30 idiomsTier 2–3 · Extensions
The extension packages that make R plots publication-ready: repelled labels, alluvial flows, independent scales, facet surgery, plot composition and the two heatmap standards of genomics.
patchworkcompose with | and /complexheatmapannotated heatmap listsggrepel_pynon-overlapping labelsggalluvialflow diagramsggh4xfacet & axis hooksggnewscalemultiple colour scalespheatmapclustered heatmapsTier 2–3 · Analysis
Single-cell trajectories, intercellular communication, CNV calling. These build their entire visualization layer on the tiers below — so a Monocle 2 trajectory plot renders like the R tutorial without adapter code.
monocle2pyDDRTree · BEAM · pseudotimemonocle3AnnData-native trajectoriesnichenetrligand → target modellingtradeseqNB-GAM lineage DEcopykatCNV aneuploidy detectionddrtreereversed graph embeddingThe spire
Every tier below publishes a _biobabel/ contract through a Python entry
point. biobabel discovers them and serves 12 read-only tools over MCP to
Claude Code, Cursor, Codex and Continue. It runs no business logic and never executes
code.
biobabelcontract registry + MCP serverA controller drives coding agents through a 13-step workflow — preparation, implementation, validation. Implementation operates on the package, not the function: the whole export surface is resolved into a dependency graph and cut into slices, each built and tested only after the ones it depends on.
An agent writes each package's _biobabel/ contract from that package's
own source, never from a neighbour's. Seven deterministic checks gate it — among them that
every public export is covered, and that each anti-pattern's detector fires on its own bad
example and stays silent on the good one.
biobabel enumerates one Python entry-point group — no central registry —
and serves what it finds as 12 read-only MCP tools. An agent sees exactly the packages the
user has installed; a package without a contract is invisible.
The catalog
Each is its own repository, its own distribution, its own maintainer. Versions track the
upstream R package — rgrid-python 4.5.3.post6 tracks R grid 4.5.3;
a Python-side fix without an R bump becomes a PEP 440 .postN.
ports R grid · imports as grid_py
R's grid graphics engine — viewports, grobs, units, the device model — reimplemented on a Cairo backend. The floor everything else stands on.
ports R scales · imports as scales
Breaks, labels, transforms and palettes. The scale machinery the grammar depends on, down to the tick-placement heuristics.
ports R gtable · imports as gtable_py
Layout tables of grobs — the skeleton a plot is assembled into: panels, axes, strips and guides.
ports R ggplot2 · imports as ggplot2_py
The grammar of graphics, whole. geoms, stats, scales, facets, coords, themes and guides — composed with +, exactly as in R.
ports R patchwork · imports as patchwork
Compose plots into arbitrary layouts with |, / and insets. Alignment that actually respects the gtable.
ports R ComplexHeatmap · imports as complexheatmap
Annotated heatmaps, splits and heatmap lists — the de-facto standard for genomics figures.
ports R pheatmap · imports as pheatmap
Pretty heatmaps with row/column clustering and annotation bars, built on the original grid model.
ports R ggrepel · imports as ggrepel_py
Text and label geoms that repel each other and the plot edges. The reason gene labels are readable.
ports R ggalluvial · imports as ggalluvial
Alluvial diagrams and flow plots for categorical longitudinal data — clone tracking, fate mapping, cohort flows.
ports R ggh4x · imports as ggh4x
Hooks into facets, axes and strips that the grammar does not expose on its own.
ports R ggnewscale · imports as ggnewscale
Multiple independent colour and fill scales in a single plot, without fighting the guide system.
ports R monocle 2.9.0 · imports as monocle2py
Single-cell trajectory inference — DDRTree, BEAM, pseudotime — with its entire visualization layer on the Bio-Babel stack. No matplotlib, no seaborn.
ports R monocle3 · imports as monocle3
AnnData-native trajectories, clustering and differential expression.
ports R nichenetr · imports as nichenetr
Ligand → target modelling of intercellular communication, with the prior networks intact.
ports R tradeSeq · imports as tradeseq
NB-GAM trajectory-based differential expression along lineages.
ports R copykat · imports as copykat
CNV-based aneuploidy detection from scRNA-seq — separating tumour from stroma.
ports R DDRTree · imports as ddrtree
Reversed graph embedding for principal-graph learning. The engine underneath Monocle 2.
Every library in the catalog is under active testing — APIs stabilizing, edges still being filed. Use it for verifiable work, and tell us when it is wrong.
The framework
Fidelity and usability are met by two agent-driven systems, coupled by the contract that each package carries: a build side that reconstructs the software, and a read side that serves its operating knowledge to the agent. Two of the three are not public yet.
Drives coding agents through the 13-step workflow that reconstructs a package natively. A controller keeps state in a report tree on disk rather than in the model's context, so progress survives an agent losing it — and implementation is ordered by a dependency-graph partition, package-wide, instead of proceeding function by function.
Coming soon
Writes each package's _biobabel/ contract from that package's own source, in
an isolated session that can never read a neighbour's. Seven deterministic checks gate the
result, so coverage of the public surface cannot silently shrink and no detector ships
without firing on its own bad example.
Assembles those contracts from every installed package through one Python entry-point group and serves them as 12 read-only tools over MCP. There is no central registry, and the server neither plans nor executes: an agent sees exactly what the user installed, and a package without a contract is invisible.
Open source ↗├── __init__.py # get_manifest() → PackageManifest
├── package.yaml # identity · tier · class · foundation
├── skill.md # the narrative, for the LLM
├── symbols/*.yaml # 261 signatures + requires/writes/mutates
├── concepts/*.yaml # 6 invariants, "for R users / for Python users"
├── idioms/*.yaml # 15 verbatim runnable templates
├── anti_patterns/* # 4 detectable mistakes + the correct form
├── templates/*.py # real code that actually runs
└── detectors.py # the AST checks, owned by the package
# the agent guesses — R muscle memory
grid.viewport(width=unit(0.5, "npc")) ✗ silently wrong
# the agent asks first
biobabel.describe_symbol(pkg="grid_py",
symbol="Unit")
→ Unit(value, units) — class, not a function.
→ anti-pattern: unit_kw — lowercase unit() is R.
grid_py.Viewport(width=Unit(0.5, "npc")) ✓
All prefixed biobabel., all returning the same envelope, all served over
line-delimited JSON-RPC on stdio. The server holds no business logic and
never executes your code.
Discovery 8
list_packagesdescribe_packagelist_workflowsdescribe_workflowlist_symbolsdescribe_symbollist_templatesdescribe_templateConcept 3
describe_conceptlist_idiomsdescribe_idiomValidation 1
check_code — static AST scan_biobabel/, no registration. There is no
reflection fallback and no degraded mode.site-packages. The
producer declares itself.exec, no subprocess, no eval. Running a
snippet is the calling agent's job.An ablation, not a demo. Same model, same 38 tasks, same workspace tools, same judge. The single independent variable is whether the agent can reach the contract server. Both arms get an identical tool set — that identity is the fairness guarantee.
The agent writes solution.py blind; the harness runs it exactly once and the
judge grades the real artifact. Nothing can be ground out by retrying — this measures what
the model actually knows.
Both arms iterate freely, so accuracy saturates and the signal moves to cost: how many rounds, how many blind executions, how many output tokens it takes to get there.
| Metric | Baseline | Bio-Babel | Δ |
|---|
114 cells per arm · pass threshold 0.85 rubric score · tasks span
grid_py 9, ggplot2_py 9, scales 7,
gtable_py 5, cross-library 4, analysis 4.
bio-babel-MCPBench ↗
The honest line: contract lookups are not free. In execution mode the Bio-Babel arm spends +272% input tokens to buy −9.7 rounds and −50% output tokens. We report the cost because a benchmark that only reports the win is an advertisement.
Principles
Behaviour trusts the reference; the API feels native to the target language. Not a line-for-line transliteration — a sibling.
No rpy2, no second interpreter, no R install. If you can
pip install it, it works.
Every public function is checked against its reference — numerically where possible, visually where not. Divergence is a bug, not a feature.
A port is alive when someone is learning it and someone is maintaining it. Every package has a caretaker, not just an author.
🔒 The AI pipeline behind these ports is still in internal development and is not yet public. It will be. In the meantime, the libraries it produces are what we are asking you to judge us on.
Community
Open an issue with a classic package you wish existed on the other side — and the workflow that is painful without it.
If a port behaves differently from its reference, that is the most valuable bug report we can receive.
Tell us which classic you would like to help bring over; we will help you scope it.
Rough drafts welcome. Our pipeline assists the draft; review and stewardship need more hands.
The scarcest resource here is long-term maintainers. If a library matters to your work, consider becoming its caretaker.