Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data.

Area of Science:

• Computational biology
• Bioinformatics
• Systems biology

Background:

• Deep learning models offer powerful predictive capabilities but often lack biological interpretability.
• Generic deep neural networks provide limited insight into the molecular mechanisms driving predictions.
• Biological networks, with nodes representing molecules and edges representing interactions, offer a framework for mechanistic understanding.

Purpose of the Study:

• To develop an interpretable deep learning approach for biological discovery.
• To combine the predictive power of deep learning with the mechanistic insights from biological networks.
• To enhance the understanding of biological mechanisms underlying complex phenomena.

Main Methods:

• Introduction of Knowledge-Primed Neural Networks (KPNNs), a novel deep learning architecture designed for biological networks.
• Development of a learning method to stabilize node weights, enhance quantitative interpretability, and control for network connectivity.
• Validation using simulated data with known ground truth and application to real-world biological datasets.

Main Results:

• KPNNs successfully integrate deep learning with biological network structures for interpretable predictions.
• The developed learning method improves the interpretability of node weights in trained KPNNs.
• Demonstrated utility of KPNNs in five distinct biological applications using single-cell RNA-seq data from cancer and immune cells.

Conclusions:

• KPNNs represent a significant advancement in interpretable deep learning for biological research.
• The method is broadly applicable to any domain where prior knowledge can be encoded in networks.
• This approach facilitates deeper biological discovery by linking predictive models to underlying mechanisms.