TxPert: using multiple knowledge graphs for prediction of transcriptomic perturbation effects.

Area of Science:

• Computational biology
• Genomics
• Systems biology

Background:

• Accurate prediction of cellular responses to genetic perturbations is vital for understanding diseases and developing therapies.
• Exploring all possible genetic perturbations is costly, necessitating methods that generalize to new conditions.

Purpose of the Study:

• To develop a deep learning method, TxPert, for predicting transcriptomic perturbation effects.
• To leverage multiple knowledge graphs of gene relationships for enhanced prediction accuracy.

Main Methods:

• TxPert utilizes a latent-transfer-based deep learning approach.
• It integrates information from multiple knowledge graphs, including biological databases and high-throughput screens.
• The method focuses on predicting gene (product)-gene (product) relationships.

Main Results:

• Combining different knowledge graphs improved prediction performance.
• TxPert achieved performance comparable to experimental reproducibility for single unseen perturbations.
• For double unseen perturbations and cross-cell line predictions, TxPert outperformed existing methods by 8-25%.

Conclusions:

• TxPert offers a powerful deep learning framework for predicting transcriptomic responses to genetic perturbations.
• The integration of diverse knowledge graphs enhances the generalizability and accuracy of perturbation effect predictions.
• This method holds promise for accelerating therapeutic development and biological discovery.