Cell Identity Codes: Understanding Cell Identity from Gene Expression Profiles using Deep Neural Networks.

Area of Science:

• Computational biology
• Genomics
• Biomedical informatics

Background:

• Accurate cell type identification is crucial in biomedical research.
• Traditional methods using marker genes are limited due to lack of exclusive markers.
• Existing machine learning approaches struggle with the complexity of gene expression profiles (GEPs).

Purpose of the Study:

• To develop a novel machine learning approach for robust cell type identification.
• To create a concise numerical representation of cell identity from GEPs.
• To improve the accuracy of cell type classification using deep learning.

Main Methods:

• Analysis of 1040 GEPs from 16 human tissues and cell types using deep neural networks.
• Development and comparison of deep autoencoder architectures.
• Introduction of a 'cell identity code' (CIC) as a 30-value numerical vector encoding GEPs.
• Development of a classifier autoencoder for cell type identification.

Main Results:

• A deep autoencoder architecture was identified that encodes GEPs into a 30-value CIC.
• The CIC accurately reconstructs the original GEP, comparable to experimental replicates.
• The CIC captures biological information, linking to different cellular pathways and processes.
• The model demonstrated accurate classification of unseen cell types and resilience to noise.
• A classifier autoencoder achieved high accuracy in identifying cell types from GEPs or CICs.

Conclusions:

• Deep autoencoders can generate a meaningful cell identity code (CIC) from GEPs.
• The CIC serves as a powerful, compact representation for understanding cell identity.
• The classifier autoencoder provides an accurate and robust method for cell type identification.