Deep Learning Based CETSA Feature Prediction cross Multiple Cell Lines with Latent Space Representation

Cellular Thermal Shift Assay (CETSA), a biophysical principle-based technique measuring the protein thermal stability inside cell, has contributed significantly to the understanding of drug mechanisms of action (MoAs) and the dissection of protein interaction states as well as cell biology processes, etc. One of the barriers for CETSA applications is that CETSA experiments must be conducted on the specific cell lines of interest, which is typically time-consuming and costly in terms of reagents and machine time. In this work, we made efforts to explore the prediction of CETSA features cross different cell lines by proposing a computational framework based on a novel deep neural network called CycleDNN. Given a set of cell lines {C_1, C_2, ..., C_n}, the proposed CycleDNN consists of encoders {E_1, E_2, ..., E_n} and decoders {D_1, D_2, ..., D_n}. E_i encodes the CETSA features from cell line C_i into Z in the latent space Z, and D_i decodes Z into the CETSA features in cell line C_i (i = 1, 2, ..., n). Any encoder and decoder can be combined into an auto-encoder to perform CETSA feature prediction from one cell line to the other. In this way, the auto-encoders form a cyclic feature prediction across multiple cell lines. The prediction loss, cycle-consistency loss, and latent space regularization loss are adopted to guide the model's training. The experimental results on a public CETSA dataset demonstrate the effectiveness of our proposed approach. Moreover, the validity of the predicted CETSA data from our model has also been verified by using a decision tree for Protein-Protein Interaction prediction.