💻 bioinformatics

Coding agents author interpretable single-cell embedding models from the literature

This paper demonstrates that coding agents can automatically generate interpretable, zero-shot single-cell embedding models by extracting and composing literature-cited gene programs into named axes, achieving biological quality comparable to data-driven methods while ensuring inherent batch robustness and interpretability without requiring training or prior gene-set databases.

Original authors: Brunn, N., Krissmer, S. M., Frosch, M., Frick, M., Prinz, M., Binder, H.

Published 2026-07-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Brunn, N., Krissmer, S. M., Frosch, M., Frick, M., Prinz, M., Binder, H.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Idea: The "Literature Librarian" Robot

Imagine you have a massive library of biology books (scientific papers) that describe how different cells in the body work. These books say things like, "A liver cell is defined by these 20 specific genes," or "A brain neuron is defined by those 15 genes."

Currently, when scientists analyze new cell data, they use powerful computers to look at the raw numbers and try to figure out the patterns themselves. It's like giving a student a pile of unorganized receipts and asking them to invent a new way to categorize spending without ever looking at a textbook. This often creates confusing results that are hard to explain and get messed up by technical errors (like different ways of measuring the data).

This paper introduces a new method: Instead of letting the computer guess the patterns, they use a coding agent (a smart AI robot) to act as a Librarian.

The robot is given a simple instruction: "We are studying the mouse brain. Please write a program that organizes cells based on what the scientific books say about them."

The robot then:

Reads the books (scientific literature) to find the specific gene lists for brain cells.
Writes a short, simple script (a "blueprint") that tells a computer how to score cells based on those specific gene lists.
Never sees the actual data it is about to analyze. It just writes the instructions based on what it learned from the books.

The Analogy: The "Recipe Book" vs. The "Taste Test"

The Old Way (Data-Driven): Imagine you are trying to sort a huge pile of fruit. You have never seen fruit before. You just look at the shapes and colors and try to group them. You might group a red apple with a red tomato because they look similar. It works, but you don't really know why they are grouped, and if the lighting changes (a "batch effect"), your groups might change completely.
The New Way (The Agent): Imagine you have a master recipe book that says, "Apples are red and sweet; Tomatoes are red and acidic." You hire a robot to write a sorting machine based only on those rules. The robot writes the code, but it never touches the fruit. When you finally feed the fruit into the machine, it sorts them perfectly because it was built on the "truth" of the recipe book, not on a guess.

What Makes This Special?

1. It's "Zero-Shot" (No Training Required)
Usually, AI models need to be "trained" on massive amounts of data to learn how to work. This robot doesn't need training. It just needs a description of the topic (e.g., "Human Pancreas"). It goes to the library, finds the rules, writes the code, and you are done. It's like hiring an expert who already knows everything and just needs a job description to start working.

2. It's "Auditable" (You Can Check the Work)
Because the robot writes a simple, named list of rules (e.g., "Axis 1: Liver Cells"), a human can read the code and say, "Yes, that matches what the science says."

Old way: The computer gives you a black box with 50 hidden numbers. You have to guess what they mean.
New way: The computer gives you a list: "This number is for Liver Cells, this one is for Heart Cells." If a number looks wrong, you can check the specific genes and the scientific paper the robot cited to see if it made a mistake.

3. It's "Batch-Proof" by Design
Scientific data often gets messed up by technical differences (like using different machines to measure cells).

Old way: You have to run a complex math correction step to fix these errors.
New way: Because the robot only looks for specific "marker genes" (the unique ID tags of a cell type) that are proven in the literature, it simply ignores the technical noise. It's like a security guard who only checks for a specific ID badge; if the lighting is bad or the camera is blurry, the guard still knows who belongs there because they are looking for the specific badge, not the general vibe.

4. You Can "Steer" the Structure
The authors showed that you can tell the robot to organize the results in a specific shape.

Example: They told the robot to organize cells from a developing mouse embryo into a family tree.
The robot arranged the axes so that the "depth" of the tree matched the age of the embryo. It didn't just find the pattern; it built the pattern exactly how the scientists asked, creating a map of development that is easy to read.

The Results: Does it Work?

The authors tested this "Robot Librarian" on real data from mice and humans (brain, pancreas, immune system).

Quality: The robot's organization was just as good as (and sometimes better than) the most advanced, data-hungry AI models currently used.
Reproducibility: If you ask the robot to do the job 10 times, it might pick slightly different genes each time, but the final result is almost identical. It's consistent.
Speed & Size: The "model" isn't a giant file. It's a tiny text script (a few kilobytes) that runs instantly on a standard computer. It doesn't need a supercomputer.

The Limitations (What the Paper Says)

The authors are honest about what this can't do yet:

It only knows what's in the books: If a new type of cell exists that scientists haven't written about yet, the robot won't find it. It can only organize what is already known.
It relies on the AI's knowledge: The robot uses a large language model to read the literature. If the model hallucinates (makes up a fake gene or paper), the blueprint could be wrong. However, because the output is a simple, readable script, a human can easily spot and fix these mistakes.

Summary

This paper presents a new way to analyze cell data: Ask an AI to write a rulebook based on scientific history, rather than letting the AI guess the rules from the data.

The result is a system that is transparent (you can read the rules), robust (it ignores technical noise), fast (no heavy training), and flexible (you can tell it how to organize the data). It turns the scattered knowledge of thousands of scientific papers into a single, usable, and auditable map for understanding cells.

Technical Summary: Coding Agents Author Interpretable Single-Cell Embedding Models

Problem
Single-cell RNA-sequencing (scRNA-seq) analysis relies on reducing high-dimensional expression profiles into low-dimensional embeddings to visualize, cluster, and order cells. Conventional methods learn these representations de novo from the expression matrix, producing dense latent spaces where dimensions are complex functions of many genes. These require post-hoc interpretation and dedicated batch-correction steps to handle technical variation. While scientific literature encodes a complementary, sparse, and compositional representation of cell states as validated marker-gene programs, this knowledge is scattered across papers and databases. Existing interpretable methods (e.g., UCell, expiMap) rely on fixed, human-curated gene-set databases that cannot be dynamically composed into structured axes or tailored to specific biological scenarios without training. Furthermore, foundation models often fail to outperform simpler baselines without fine-tuning and remain black boxes.

Methodology
The authors propose a framework where a coding agent authors interpretable single-cell embedding models directly from the scientific literature, given only a description of a biological scenario (species, tissue, and optional notes). The agent never sees the expression data, requires no training, and does not use a pre-existing gene-set database.

The workflow consists of two stages:

Authoring (Stage A): The agent receives a scenario configuration and a fixed Python template. It edits a single build_blueprint() function to curate named, literature-cited gene programs and compose them into embedding axes. The agent selects genes and programs relevant to the scenario, citing peer-reviewed resources (papers, atlases, or databases).
Refinement (Stage B, Optional): A discriminator agent critiques the blueprint's biological coverage, and a generator agent revises it to address findings, ensuring robust literature grounding.

The resulting blueprint is a short, declarative Python script (10–40 KB) rather than a trained model. It defines three types of axes:

Plain axes: Read out a single gene program.
Contrast axes: Oppose two programs to separate populations.
Gated axes: Restrict a program to a parent class, allowing the chaining of gates to encode hierarchical structures (e.g., developmental trees).

Embedding construction is a deterministic, CPU-only pass where each cell is scored on the curated marker genes of each axis. The output is a cell-by-axis table. The agent is constrained by a fixed API to ensure the output is standardized, auditable, and reproducible.

Key Contributions

Autonomous Model Authoring: Demonstrating that coding agents can synthesize interpretable embedding models from unstructured literature without training or data access.
Dynamic Curation: Unlike fixed-database methods, the agent dynamically curates gene programs specific to the biological scenario, enabling the discovery of fine-grained cell states (e.g., cortical subclasses) often missing from standard databases.
Inherent Interpretability and Auditability: Every dimension is a named axis composed of cited gene modules. The model is auditable down to individual genes, programs, and citations.
Steerable Topology: The agent can be instructed via natural language to arrange axes into specific topologies, such as a time-ordered developmental tree where depth reflects embryonic stage.
Batch Robustness by Construction: Because the embedding scores cells only on curated marker genes validated in the literature, it is inherently robust to batch effects without needing explicit correction or batch variables.

Results
The approach was validated across mouse and human tissues (mouse isocortex, human pancreas, human immune system, and mouse gastrulation).

Biological Quality: The zero-shot blueprints are competitive with conventional (PCA, scVI), foundation-model (scVI-census, UCE), and program-informed (UCell, expiMap) baselines. On the mouse isocortex, the blueprint outperformed fixed-database methods in resolving fine-grained subclasses (scIB Total 0.63–0.64 vs. 0.49–0.60) by curating ~23 named subclass axes not present in standard databases.
Batch Robustness: The blueprint achieves batch correction scores comparable to trained methods (scVI, expiMap) and significantly higher than prior-free methods. This invariance is structural: the model lacks channels for technical variation to load onto.
Reproducibility: Across 10 re-authoring runs, the embedding geometry showed high reproducibility (mean pairwise CKA ~0.85), despite the agent selecting partly different genes each time. The core biological vocabulary (genes, modules, axes) remained stable.
Developmental Tree: In the mouse gastrulation atlas, a single natural-language instruction steered the agent to organize axes into a tree where depth correlated with embryonic time (Spearman $\rho \approx 0.98$ per lineage). This structure was designed-in, contrasting with trajectory inference tools that recover graphs post-hoc from data geometry.
Efficiency: Authoring takes minutes (2–7 min) and yields kilobyte-scale scripts, contrasting with the megabyte-to-gigabyte footprints and longer training times of conventional methods.

Significance
The paper claims to provide a new paradigm for interpretable representation learning in single-cell genomics. It bridges the gap between the qualitative knowledge in the scientific literature and quantitative, auditable models. By treating the literature as a latent model of cell states, the approach offers a "literature lens" that is complementary to data-driven views. The authors emphasize that this method delivers a unique combination of properties: interpretability, batch robustness, dynamic adaptability to specific scenarios, and steerability into structured topologies, which no existing approach provides simultaneously. The work is presented as a step toward converting scattered, qualitative domain knowledge into quantitative, auditable models, provided the agent's output is disciplined by a structured API. The authors note limitations, including dependence on the quality of the literature and the language model's pretraining, and suggest future directions such as combining these priors with lightweight learned models to capture residual biological structure.