Rxivist logo

Deciphering signaling specificity with interpretable deep neural networks

By Yunan Luo, Jianzhu Ma, Yang Liu, Qing Ye, Trey Ideker, Jian Peng

Posted 25 Mar 2018
bioRxiv DOI: 10.1101/288647

Protein kinase phosphorylation is a prevalent post-translational modification (PTM) regulating protein function and transmitting signals throughout the cell. Defective signal transductions, which are associated with protein phosphorylation, have been revealed to link to many human diseases, such as cancer. Defining the organization of the phosphorylation-based signaling network and, in particular, identifying kinase-specific substrates can help reveal the molecular mechanism of the signaling network. Here, we present DeepSignal, a deep learning framework for predicting the substrate specificity for kinase/SH2 sequences with or without mutations. Empowered by the memory and selection mechanism of recurrent neural network, DeepSignal can identify important specificity-defining residues to predict kinase specificity and changes upon mutations. Evaluated on several public benchmark datasets, DeepSignal significantly outperforms current methods on predicting substrate specificity on both kinase and SH2 domains. Further analysis in The Cancer Genome Atlas (TCGA) demonstrated that DeepSignal is able to aggregate mutations on both kinase/SH2 domains and substrates to quantify binding specificity changes, predict cancer genes related to signaling transduction, and identify novel perturbed pathways.

Download data

  • Downloaded 995 times
  • Download rankings, all-time:
    • Site-wide: 11,554 out of 89,211
    • In bioinformatics: 1,867 out of 8,418
  • Year to date:
    • Site-wide: 44,146 out of 89,211
  • Since beginning of last month:
    • Site-wide: 60,447 out of 89,211

Altmetric data


Downloads over time

Distribution of downloads per paper, site-wide


PanLingua

Sign up for the Rxivist weekly newsletter! (Click here for more details.)


News