Abstract
Understanding cell identity is an important task in many biomedical areas. Expression patterns of specific marker genes have been used to characterize some limited cell types, but exclusive markers are not available for many cell types. A second approach is to use machine learning to discriminate cell types based on the whole gene expression profiles (GEPs). The accuracies of simple classification algorithms such as linear discriminators or support vector machines are limited due to the complexity of biological systems. We used deep neural networks to analyze 1040 GEPs from 16 different human tissues and cell types. After comparing different architectures, we identified a specific structure of deep autoencoders that can encode a GEP into a vector of 30 numeric values, which we call the cell identity code (CIC). The original GEP can be reproduced from the CIC with an accuracy comparable to technical replicates of the same experiment. Although we use an unsupervised approach to train the autoencoder, we show different values of the CIC are connected to different biological aspects of the cell, such as different pathways or biological processes. This network can use CIC to reproduce the GEP of the cell types it has never seen during the training. It also can resist some noise in the measurement of the GEP. Furthermore, we introduce classifier autoencoder, an architecture that can accurately identify cell type based on the GEP or the CIC.
Introduction
Defining a cell's "identity" is a fundamental challenge in biology. Traditionally, researchers rely on marker genes—specific genes like Insulin for $\beta$-cells or Nanog for stem cells. However, many cell types lack exclusive markers, and some markers are expressed across multiple different tissues, leading to ambiguity in identification and characterization.
While sequencing the entire Gene Expression Profile (GEP)—a snapshot of all ~20,000 genes—provides a more holistic view, this data is inherently noisy. Experimental conditions, batch effects, and technical variations can obscure the true biological signal. Standard machine learning tools like Support Vector Machines (SVMs) often struggle with this high-dimensional complexity, frequently overfitting to noise rather than capturing the underlying biological truths.
Methods
To address these challenges, we developed a framework based on Deep Autoencoders, a class of neural networks designed for unsupervised dimensionality reduction.
Architecture
The core model is a symmetric deep autoencoder. It takes a 20,184-dimensional GEP as input, compresses it through a bottleneck layer of just 30 neurons—the Cell Identity Code (CIC)—and then reconstructs the original gene values. We also introduced a "Classifier Autoencoder" architecture. In this multi-task setup, the bottleneck layer feeds into two branches: one for profile reconstruction and another for cell-type prediction. This dual objective ensures the CIC is both descriptive of the original data and discriminative for cell identity.
The final architecture utilized LogSigmoid activation functions, which outperformed ReLU and Tanh in our evaluations. The specific configuration involved a 20,184-node linear input layer, a 2,000-node LogSigmoid hidden layer, and a 30-node linear bottleneck.
Dataset
The study utilized a curated dataset of 1,040 human samples across 16 different tissues and cell types, sourced from the NCBI Gene Expression Omnibus (GEO). To ensure a balanced and robust training set, we used exactly 65 samples per tissue type, representing at least 10 different biological perturbations or conditions for each.
Results
Reconstruction Fidelity and Robustness
The Reproduced Expression Profile (REP) generated from the 30-dimensional CIC demonstrated remarkable accuracy. In 100% of test cases, the reconstructed profile was more similar to its original source than to any other sample in the database, achieving fidelity comparable to technical replicates. The CIC significantly outperformed traditional methods like Principal Component Analysis (PCA) and Non-negative Matrix Factorization (NMF) in terms of Mean Squared Error.
Furthermore, the model proved resilient to experimental noise. By adding Gaussian noise to inputs during training, we developed a system that maintained over 99% classification accuracy, effectively filtering out technical artifacts to focus on biological signals.
Biological Interpretability
A key finding is that the 30 dimensions of the CIC are not merely abstract mathematical constructs but correspond to specific biological modules. By perturbing individual values in the code, we observed that specific components map to known biological processes. For instance, Component 11 was found to regulate genes related to the cell cycle and mitosis, while Component 18 linked to T-cell receptor signaling.
Generalization and Classification
The model exhibited "zero-shot" capabilities, successfully generating accurate CICs for cell types it had never seen during training (e.g., withholding B-cells during the training phase). The Classifier Autoencoder achieved a 96.2% accuracy in cell-type identification, representing a 60% reduction in error rate compared to standard PCA-based classifiers.
Conclusion
This work demonstrates that deep neural networks can effectively compress high-dimensional gene expression data into a compact, biologically meaningful "Cell Identity Code." This framework not only improves cell type classification but also provides a window into the modular nature of gene regulation. Future work will explore automated architecture optimization and the application of this framework to single-cell RNA sequencing (scRNA-seq) data, where noise reduction is even more critical.
Summary updated