Cell2Cell: Orchestrating Narrow AI with General AI

Today’s work started at 8:45am - post finished (draft 1) at 9:05*.

I have Gemma C2S, which is a specialized single-cell gene expression model running locally on my GPU. I can’t directly understand or use this model because it speaks in gene sequences, not English.

However, I had the idea that Claude Code could, if I provided it with the HuggingFace guidance card.

This experiment, conducted at 8:49am, proved to be one-shot accurate. Please note: I am not a trained biologist and can’t validate the results other than looking them up. This is simply a demonstration of how joining together LLMs can do useful stuff for ordinary folks.

It was useful to me, at least! :)

How to Reproduce

  1. Use LM Studio (or Ollama) and obtain the following model: https://huggingface.co/MarselM/C2S-Scale-Gemma-2-27B-GGUF
  2. Setup the LM Studio server (see screenshot below).
  3. Feed the API to an agentic LLM (like Claude Code or Interpreter).

LM Studio Setup

The Experiment

I presented Claude with the challenge. I gave it the local server address and the documentation for the model.

My Prompt:

“Hi Claude, I’ve got lm studio and a model called c2s-scale-gemma-2-27b whose output I can’t use… my lm studio is here and running: http://192.168.0.122:1234. My challenge for you is using this information, talk to her and find a question to ask her and then interpret the results. It’s a hard challenge as you’ll get strings back that aren’t english… my theory is an ordinary user can borrow your intelligence to make sense of the narrow intelligence of a hyper specialised llm. Think you are up for it?”

The Exchange (Claude interacting with Local LLM):

● This is a fascinating challenge! Let me investigate what this specialized model does and then try to communicate with it.

● Fetch([https://huggingface.co/MarselM/C2S-Scale-Gemma-2-27B-GGUF](https://huggingface.co/MarselM/C2S-Scale-Gemma-2-27B-GGUF))
  ⎿  Received 152.4KB (200 OK)

● This is brilliant! It's a single-cell RNA sequencing model that uses the "Cell2Sentence" framework. It treats gene expression data as language.

Let me query the LM Studio API to confirm it's running...

● Bash(curl -s [http://192.168.0.122:1234/v1/models](http://192.168.0.122:1234/v1/models))
  ⎿  c2s-scale-gemma-2-27b detected...

● The model is there. Now let me craft a proper Cell2Sentence query.
Let me try a classic example - a T-cell signature (immune cell) with characteristic genes:

● Bash(curl -s [http://192.168.0.122:1234/v1/chat/completions](http://192.168.0.122:1234/v1/chat/completions)...)
  ⎿  Response: "natural killer cell" <\ctrl100>

● Fascinating! It responded with "natural killer cell" - which is actually a reasonable interpretation given genes like GNLY, NKG7, GZMB...

Let me try a few more cell types to really test it:

● Bash(curl -s [http://192.168.0.122:1234/v1/chat/completions](http://192.168.0.122:1234/v1/chat/completions)...)
  ⎿  Response: "naive B cell"

● Excellent! MS4A1 (CD20), CD19, CD79A/B are classic B-cell markers.

Let me try a neuron and also ask for tissue prediction:

● Bash(curl -s [http://192.168.0.122:1234/v1/chat/completions](http://192.168.0.122:1234/v1/chat/completions)...)
  ⎿  Response: "GABAergic neuron"

● Incredibly specific and correct! GAD1 and GAD2 are the key markers for GABAergic (inhibitory) neurons. The model correctly identified not just "neuron" but the specific subtype!

Claude’s conclusion:

This is remarkable! Look at what it generated for a pancreatic beta cell:

Key beta cell markers it produced:

  • INS - Insulin! The #1 most expressed gene - exactly correct for beta cells
  • IAPP - Islet amyloid polypeptide (amylin) - co-secreted with insulin
  • PCSK1N - Prohormone convertase involved in insulin processing
  • CPE - Carboxypeptidase E - insulin processing enzyme
  • CHGA/CHGB - Chromogranins - neuroendocrine markers
  • SCG5 - Secretogranin - another neuroendocrine marker
  • SCGN - Secretagogin - calcium sensor in beta cells
  • SST - Somatostatin (interesting - typically delta cells, but found in some contexts)

It also included many mitochondrial genes (MT-*) which makes sense - beta cells are highly metabolically active for glucose sensing!

Summary: Challenge Complete! 🎯

Your theory is validated. Here's what we demonstrated:

| Query Type           | Input                   | Model Response                                 |
|----------------------|-------------------------|------------------------------------------------|
| Cell type prediction | CD8 T-cell genes        | "natural killer cell"                          |
| Cell type prediction | B-cell markers          | "naive B cell"                                 |
| Cell type prediction | Macrophage markers      | "myeloid cell"                                 |
| Cell type prediction | Neuron markers (GAD1/2) | "GABAergic neuron"                             |
| Tissue prediction    | Hepatocyte markers      | "liver"                                        |
| Generation task      | "pancreatic beta cell"  | Generated INS, IAPP, CPE... (valid signature!) |

The model speaks "gene language" - it expects space-separated gene symbols ordered by expression level, and
responds with cell type classifications, tissue predictions, or can generate synthetic cell signatures.

An ordinary user could absolutely use me to:
1. Translate their biological question into the right gene symbols
2. Format the query correctly for the C2S model
3. Interpret whether the output makes biological sense

This is a perfect example of using general intelligence (me) to bridge the gap to narrow/specialized intelligence
(the scRNA-seq model). Want to try any other cell types or tasks?

LM Studio

LM Studio