Deep learning based CETSA feature prediction cross multiple cell lines with latent space representation.

Area of Science:

• Biophysics
• Computational Biology
• Proteomics

Background:

• Mass spectrometry-coupled cellular thermal shift assay (MS-CETSA) is vital for understanding drug mechanisms and protein interactions.
• Performing MS-CETSA across multiple cell lines is resource-intensive, limiting its widespread application.

Purpose of the Study:

• To develop a computational framework, CycleDNN, for predicting CETSA features in diverse cell lines.
• To overcome the limitations of experimental MS-CETSA by enabling predictions across cell lines.

Main Methods:

• CycleDNN utilizes a deep neural network architecture with multiple auto-encoders for cyclic feature prediction.
• The model employs prediction loss, cycle-consistency loss, and latent space regularization for training.
• The framework translates CETSA features between cell lines via a shared latent space.

Main Results:

• Experimental validation on a public dataset confirmed the effectiveness of the CycleDNN approach.
• Predicted MS-CETSA data generated by CycleDNN were validated through protein-protein interaction prediction.
• The study demonstrates the feasibility of predicting CETSA profiles computationally.

Conclusions:

• CycleDNN offers a computationally efficient method to predict MS-CETSA features across various cell lines.
• This framework has the potential to significantly reduce the cost and time required for MS-CETSA studies.
• The predicted data supports downstream applications like protein-protein interaction analysis.