c2s-mechinterp

Mechanistic interpretability of Cell2Sentence (C2S-Scale) language models.

Code and analysis pipelines for the paper "Same ceiling, different machinery: mechanistic interpretability of a text language model trained to read cells as sentences" (Ihor Kendiukhov, Institute of Medical Genetics and Applied Genomics, University of Tübingen).

C2S-Scale represents a cell as a sentence of gene names — the genes are ranked by expression and written out as plain text, so a general-purpose language model (Gemma-2) can be trained on cells the way it is trained on text. This repository opens that model up and asks what it actually computes.

Interactive feature atlas: https://biodyn-ai.github.io/c2s-atlas/ Atlas source code: https://github.com/Biodyn-AI/c2s-atlas

---

What the study found

	Result
✅ Interpretable features	68% of sparse-autoencoder features per layer carry a significant ontology annotation; 119,808 features released in the public atlas
✅ Features persist across depth	50.4% of layer-0 features are still recoverable at layer 25 (cosine 0.88) — against 2–3% for an encoder-style model under the identical pipeline
✅ Coherent causal circuits	27,633 significant cross-layer edges; 85.2% join features that share an ontology term
✅ Geometry beyond co-expression	functionally related genes are closer than a co-expression-matched permutation null predicts (z = 3.9–6.6)
✅ Causally used	steering along internal directions moves the model's own predictions; norm-matched random directions do not
❌ No regulatory logic	96.7% of gene knockdowns are detected, but only 2.1% of transcription factors give a target-specific response
❌ Scale hurts legibility	at matched dictionary capacity the 27B model has 31× more dead features and 19 points lower annotation than the 2B
❌ No gain over the input	on cell-level tasks the model sits within 0.007 of a linear readout of its own input encoding

Full provenance for every number: docs/05-results.md.

The protocol in one picture

a cell  ─►  top-512 gene names, ranked      (expression values discarded)
            │
            ▼
        C2S-Scale (Gemma-2, 2B / 27B)  ── residual stream tapped
            │
   ┌────────┼─────────┬──────────────┬─────────────┐
   ▼        ▼         ▼              ▼             ▼
Stage 1   Stage 2   Regulatory     Stage 3      Stage 4
feature   circuit   logic test     geometry     public
atlas     tracing   (perturbation) + steering   atlas

Every model claim is judged against three controls — see docs/02-controls.md. They are the methodological core of the paper, and each one changed a conclusion.

Install

git clone https://github.com/Biodyn-AI/c2s-mechinterp.git
cd c2s-mechinterp
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Version pins that matter (both cost us a day):

• transformers==4.46.3 — 5.x breaks against torch 2.4 with a DTensor import error.
• tiktoken + protobuf are required for the Gemma-2 tokenizer.

A GPU is needed for extraction, circuit tracing and steering. The 2B model runs on a 24 GB card (RTX 4090 / A5000); the 27B model needs 80 GB for forward passes, though its sparse autoencoders can be retrained on a 24 GB card from stored activations.

Quickstart — a minimal Stage-1 atlas

# 1. extract gene-token activations for a few layers
python -m c2s_mechinterp.extract_activations \
    --model vandijklab/C2S-Scale-Gemma-2-2B --h5ad data/k562.h5ad \
    --layers 0 8 16 --max-cells 1200 --max-genes 512 \
    --unit gene-token --store-dtype float32 --out runs/demo

# 2. train a TopK sparse autoencoder on one layer
python -m c2s_mechinterp.train_sae --run runs/demo --layer 16 --expansion 4 --k 32 --epochs 6

# 3. catalogue and annotate the features
python -m c2s_mechinterp.feature_catalog  --run runs/demo --layer 16
python -m c2s_mechinterp.annotate_features --run runs/demo --layer 16 --ontology-dir data/ontologies

Full commands for every result in the paper: docs/04-reproduce.md.

Repository map

Path	Contents
c2s_mechinterp/	Core pipeline: cell-sentence construction, activation extraction, TopK SAEs, feature catalogues, ontology annotation, co-activation modules, cross-layer tracking, causal circuit tracing, perturbation response, scaling comparison
geometry/	Stage 3: gene-representation geometry with the co-expression-matched null, cell-state trajectory geometry, intrinsic-dimension probes
manifold/	Cell-cycle manifold, activation steering, lap-walk navigation, continuity and plateau tests, information-matched transfer experiments (supporting analyses)
paper/	Manuscript source and the scripts that regenerate every figure from the result artefacts
results/	Result write-ups and the small JSON artefacts the paper's numbers are read from
scripts/	Smoke tests and unit checks (e.g. sub-word→gene attribution)
docs/	Protocol, controls, data, reproduction and results documentation

manifold/_ported_from_encoder_routes/ holds reference implementations ported from earlier encoder-model studies, kept verbatim for comparability; they are not part of the C2S pipeline.

Data

No primary data is redistributed here. All datasets are public — see docs/03-data.md for accessions and preparation steps.

Related repositories

Repository	Purpose
Biodyn-AI/c2s-mechinterp	This repository — analysis pipelines
Biodyn-AI/c2s-atlas	The interactive SAE feature atlas (React application + released feature data)

Citation

See CITATION.cff. Please also cite the C2S-Scale model and the datasets used.

License

MIT — see LICENSE.