💻 bioinformatics

Mapping Gene Expression to an Interpretable Semantic Space

The paper introduces MESIC, a method that integrates curated gene knowledge into a semantic coordinate system to map single-cell expression data into interpretable components, enabling biologically meaningful clustering and annotation of cell types without relying on post-hoc interpretation.

Original authors: Duan, X., Aggarwal, M., Periwal, V.

Published 2026-09-18

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Duan, X., Aggarwal, M., Periwal, V.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

In the quiet hum of a living cell, a complex script is being read and rewritten every second. This script is made of genes, the molecular instructions that tell a cell how to behave, what to build, and how to respond to its environment. Scientists have developed powerful tools to read these instructions from individual cells, a technique known as single-cell RNA sequencing. By measuring which genes are active in a single cell, researchers can map out the vast diversity of life within a tissue, distinguishing a heart muscle cell from a lung cell, or a healthy cell from one that is sick. However, the data these tools produce is overwhelming. It comes as a massive list of numbers for tens of thousands of genes, a high-dimensional landscape where the patterns are hidden in the noise. To make sense of this, scientists usually compress the data into a simpler map, grouping similar cells together. But these maps have a blind spot: the directions on the map do not mean anything biological. A cluster of cells might be grouped together, but the map cannot tell you why they are grouped or what specific genes are driving that grouping. The meaning has to be added later, like a translator arriving after the meeting has ended.

A new approach called MESIC changes this by building the meaning directly into the map itself. Instead of starting with the raw numbers from the cells, the researchers began with the written knowledge humans have already gathered about genes. They took the summary descriptions of over 21,000 human genes, found in public databases, and fed them into a computer program trained to understand language. This program converted the text of each gene's description into a mathematical point, capturing the essence of what that gene does. The researchers then compressed these points into fifty distinct directions, or components. Each component acts like a spotlight, illuminating a small, specific set of genes that share a common biological theme, such as immune defense or energy production. Because these components are derived from the written summaries of the genes, they are interpretable from the start. When a new cell is placed on this map, its position is defined by these named, biological themes rather than by abstract numbers.

The researchers tested this system on two very different biological landscapes to see if it could reveal hidden truths. First, they looked at heart muscle cells from patients with a condition called hypertrophic cardiomyopathy, a disease where the heart muscle becomes abnormally thick. They mapped the cells onto their new semantic space and looked for the outliers, the cells that sat furthest away from the rest of the group. They found that these distant cells were significantly more likely to come from patients with the disease. More importantly, the system could explain exactly why these cells were different. The components that pushed these cells apart were linked to calcium handling and energy metabolism, biological processes that are known to be disrupted in this heart condition. The map did not just flag the sick cells; it pointed directly to the specific biological machinery that was failing.

Next, the team applied the method to a massive atlas of human lung cells, containing over 120,000 cells from various tissues. They let the computer group the cells without telling it what the cell types were, relying only on the new semantic map. The resulting groups matched the expert classifications used by biologists with remarkable accuracy. The clusters separated the different cell types, such as immune cells and lung lining cells, in a way that aligned with established biology. The researchers then used this map to help label cells that the original atlas had left unclassified. While the standard methods failed to assign a confident identity to about half of these unknown cells, the new semantic map successfully placed them into specific groups. For these previously mysterious cells, the map provided a confident assignment and an immediate explanation based on the genes that defined their new home.

The power of this approach lies in its ability to connect the raw data of a cell directly to the written knowledge of science. In the lung atlas, for instance, one of the key components was labeled with terms related to sperm tails. At first glance, this might seem unrelated to the lung, but the map revealed that the genes driving this component are also responsible for the tiny, hair-like structures that move mucus in the airways. These structures share the same internal machinery as sperm tails. The system recognized this deep biological connection and used it to organize the cells correctly. This shows that the map is not just grouping cells by similarity, but by the actual functional logic of their biology.

This work suggests that we do not need to wait for a separate analysis to understand what our data means. By embedding the written summaries of genes into the structure of the analysis itself, the researchers created a coordinate system where every result is traced back to named genes and their known functions. The components are fixed and reusable, meaning they can be applied to new datasets from different tissues or disease states without needing to be retrained. This offers a stable reference point for scientists, allowing them to track how cells change over time or in response to treatment using the same biological language. While the method relies on the quality of the existing text descriptions and the ability of the computer to understand them, the results indicate that this bridge between language and biology is strong enough to organize complex cellular data in a way that is both accurate and immediately understandable.

Technical Summary: Mapping Gene Expression to an Interpretable Semantic Space (MESIC)

Problem Statement
Current single-cell RNA sequencing (scRNA-seq) analysis relies on cell embeddings to reduce high-dimensional expression data into compact representations for visualization and clustering. However, standard embeddings (from workflows like Seurat or foundation models) possess dimensions with no direct biological meaning. Consequently, clusters must be interpreted post-hoc through differential expression and pathway analysis. Furthermore, these embeddings typically ignore the extensive written knowledge about genes held in curated databases (e.g., NCBI summaries), relying solely on expression profiles. While recent methods like GenePT integrate semantic gene embeddings from language models, the resulting cell embeddings still inherit non-interpretable dimensions, requiring post-hoc interpretation. There is a need for a framework where the dimensions of the embedding space themselves are interpretable and directly traceable to named genes and their biological annotations.

Methodology
The authors propose MESIC (Mapping Expression to Semantic space with Interpretable Components), a method that constructs an interpretable coordinate system for single-cell data by compressing semantic gene embeddings into a fixed set of components.

Semantic Gene Embeddings: The authors utilized PubMedBERT, a biomedical language model pretrained on PubMed abstracts, to generate 768-dimensional semantic embeddings for 21,788 human genes based on their NCBI Gene summary texts.
Component Construction: These high-dimensional semantic embeddings were compressed into 50 interpretable components using a sparse PCA-like optimization. The objective function combined reconstruction loss with penalties to ensure:
- Orthogonality: Directions in the component space are near-orthogonal.
- Sparsity/Concentration: Each component is concentrated on a small subset of "core genes" (identified via a fixed-point pruning iteration based on participation ratio).
Component Interpretation: Each of the 50 components is defined by its core genes. These genes are annotated against Gene Ontology (GO), Reactome pathways, and RNAcentral classes to derive biological themes (e.g., "calcium entry," "mitoribosome quality control"). This annotation process is performed once, prior to analyzing any expression dataset.
Cell Projection: Single-cell expression profiles are mapped onto this fixed component space. For a given dataset, expression values are matched to the core genes of the components, and the cell's position in the 50-dimensional space is calculated as a linear projection ( $Z = X_{matched}P_{matched}$ ).
Downstream Analysis: Because the components are fixed and interpretable, downstream tasks (outlier detection, clustering, label transfer) yield results characterized immediately by the specific components and their associated named genes, rather than requiring post-hoc marker discovery.

Key Results
The authors validated MESIC in two stages: evaluating the components themselves and testing cell embeddings on two distinct datasets.

Component Validity and Interpretability:
- Gene-Type Preservation: Compressing 768-dimensional embeddings into 50 components preserved gene-type information. Classifiers (Logistic Regression, Random Forest, XGBoost) achieved >96% accuracy in predicting GENCODE gene types from the 50-dimensional MESIC embeddings, comparable to the original 768-dimensional embeddings.
- Biological Themes: 42 of the 50 components were enriched for specific biological themes. For example, C11 was associated with "calcium entry/cation channel," C30 with "respiratory chain," and C20 with "sperm flagellum" (which the authors note also captures axonemal machinery in airway cilia).
Cardiomyopathy Outlier Detection:
- In a dataset of 15,846 cardiomyocytes from patients with dilated and hypertrophic cardiomyopathy (HCM), cells were ranked by their multivariate distance from the population mean in the MESIC space.
- The top 1% of outliers were significantly enriched for HCM samples (65.4% vs. 37.0% in the general population).
- A sparse logistic regression identified the components distinguishing these outliers. The most significant components were C11 (calcium handling), C30 (respiratory chain), and C08 (fatty acid metabolism), all of which align with known HCM pathophysiology.
Human Lung Cell Atlas (HLCA) Clustering and Label Transfer:
- Unsupervised Leiden clustering on 121,894 lung cells in the MESIC space produced six robust clusters (L1–L6) that aligned strongly with expert-curated "level-2" annotations (Adjusted Rand Index = 0.714).
- Components separating these clusters matched their biological identities (e.g., C21 "Ig/TCR" for lymphoid clusters; C03 "innate defense" for myeloid clusters).
- Label Transfer: When applied to 53,331 extended-atlas cells previously left unannotated (label "Unknown") by the original HLCA workflow, MESIC confidently assigned 53% of these cells to one of the six reference clusters. In contrast, a standard expression-space transfer from the same reference cells only labeled 36%. MESIC provided cluster-level placements for many cells that remained unresolved by the original method.

Significance and Claims
The paper claims that MESIC provides a "coordinate system" for single-cell analysis where every result is traced directly to genes and the written knowledge about them.

Interpretability by Design: Unlike standard embeddings where dimensions are abstract, MESIC dimensions are defined by small sets of genes and their annotations. Interpretation is built into the space, not added afterward.
Reusability: The components are computed once from gene summaries, independent of any specific expression dataset. This allows any new dataset to be mapped onto the same space, enabling consistent comparison across tissues, diseases, and time without retraining.
Complementarity: The authors position MESIC as complementary to existing cell embeddings. It does not replace foundation models but supplies a stable, interpretable reference frame where results can be explained through named genes and biological themes.
Limitations and Future Work: The authors acknowledge that some components may reflect curation templates or nomenclature (e.g., Greek-letter naming) rather than pure biology. They also note that while the method identifies components distinguishing cell groups, it does not yet decompose individual cell distances into component contributions, a task left for future work.

In summary, MESIC demonstrates that integrating curated gene knowledge via language models into a compressed, sparse component space yields a robust framework for interpretable single-cell analysis, successfully linking disease outliers and cell types to specific biological mechanisms without post-hoc marker searching.

Technical Summary: Mapping Gene Expression to an Interpretable Semantic Space (MESIC)

More like this