Ledidi skill
Ledidi designs minimal edits to DNA, RNA and protein sequences for a target model output; its skill guides agents.
- 119 ★ on the repository
- Last commit
- Apache-2.0
Minimal access
Why this level
- The skill leads to training runs on your machine; GPU and memory use depend on the model.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/ledidiDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill leads to training runs on your machine; GPU and memory use depend on the model.
This is third-party code. Review the repository files before installing.
What it does
Ledidi inverts normal training: the oracle model stays frozen and the data is optimized. The library learns a continuous weight matrix, samples edits to a template sequence through Gumbel-softmax and pushes them until the model outputs the desired result, while an input loss keeps the number of edits small. Any differentiable PyTorch model that takes a one-hot sequence becomes an editor. The package ships a Claude Code skill. It acts as a router: a short `SKILL.md` sends the agent to reference files by topic. They cover the objective and the trade-off parameter `l`, multi-task and multi-model oracles, constraints and priors, in-painting, custom losses, affinity catalogs, pruning, design validation and out-of-memory recovery. Some API mistakes pass silently, so the skill requires reading the reference before writing code.
Who it is for. Bioinformaticians and researchers designing regulatory sequences or proteins with trained models.
Good fit when
- A motif must be inserted into DNA or a binding site knocked out with minimal edits.
- A cell-type specific element is designed for a target predicted activity.
- An agent writes Ledidi code and must know the silent API pitfalls.
- Several oracle models are balanced in one design.
Not a fit when
- Discrete substitutions from a motif library are needed: use `tangermeme.design`.
- There is no trained differentiable model to score sequences.
- The task has nothing to do with biological sequences.
Example request
Design minimal edits to this sequence so the model predicts higher activity in K562 cells, then validate the outcome.Limitations
Ledidi works with a model you supply, and design quality is bounded by that model. The library defaults to `device='cuda'`, so on a machine without a GPU you must set `cpu` explicitly. The `tangermeme` dependency is required and brings its own skill, which is worth installing too. The bioRxiv preprint describes the method, while laboratory validation of designs is up to the user. Use `--force` to refresh the skill after upgrading the package. The last commit dates from September 2026.
How to disable. Delete the `~/.claude/skills/ledidi` directory and the package with `pip uninstall ledidi`. Design results remain in your files.
FAQ
How do I install the agent skill?
After `pip install ledidi` run `ledidi-install-skills`: it copies the skill to `~/.claude/skills/ledidi`. `--print-path` shows the source directory and `--force` refreshes it.
What is tangermeme and do I need its skill?
It is a required dependency handling encoding, FASTA I/O, motifs and plots. Its skill, installed with `tangermeme-install-skills`, complements the Ledidi skill.
What if there is no GPU?
The library defaults to `cuda`, so pass `device='cpu'`. The README gives a separate install command for CPU-only PyTorch.
When is another tool better?
For discrete design, such as implanting motifs from a catalog or greedy substitution, `tangermeme.design` fits better.
Where is the method described?
In the bioRxiv preprint and the documentation on readthedocs, both linked from the README.
Discussion
No questions or reviews yet.
Related
SkillsСкилл Claude Code учит агента библиотеке tangermeme для анализа геномных нейросетей: DeepLIFT/SHAP, мутагенез, дизайн.
Agent skills for Claude Code, Codex and Cursor that add artificial dendrites to a PyTorch model and analyze training results.
17 skills for computational biology: RNA-seq, single-cell, protein structure, docking, literature and figures
A template for data scientists: AGENTS.md, CLAUDE.md and skills for Python, SQL, Polars and model diagnostics.
Around 200 skills for Claude Code on bioinformatics, chemistry and biostatistics, from RNA-seq analysis to molecular docking
NVIDIA's skill for Claude Code and Codex: the agent turns a dataset description into a config, validates it, previews it and generates synthetic data
README in short
The README presents the idea of the method, installation, a usage example and documentation links. A separate section covers working with a coding agent.