3.22 NetTCR: Towards Accurate Prediction of T-cell Targets using Deep Learning
Leon Eyrich Jessen, DTU Technical University of Denmark
Thursday, 16th August from 11:10 - 11:30
Vanessa Isabell Jurtz(1), Leon Eyrich Jessen(1), Martin Closter Jespersen(1), Kamilla Kjærgaard Jensen(1), Bjoern Peters(2), Paolo Marcatili(1), Morten Nielsen(1). (1). Department of Bio and Health Informatics, Technical University of Denmark, Lyngby, Denmark. (2). La Jolla Institute for Allergy and Immunology, San Diego, California, USA.
The interaction between the Major Histocompatibility Complex type I (MHC-I), a peptide and the T-cell receptor (TCR) (MHCI::p::TCR) is a key determinant of immune response elicitation and therefore of paramount importance in infectious- and autoimmune diseases and cancer. Current state-of-the-art models developed by our group can with great precision model MHCI::p interactions. Using data from VDJdb and IEDB, we created an ensemble of convolutional neural networks, which to the best of our knowledge is the world’s first sequence based model capable of capturing the entire MHCI::p::TCR system. Due to limited data, we however currently can only model the interaction between the CDR3 region of the TCR’s beta chain with HLA-A*02:01 and 3 peptides. However, as the model framework is easily extendable, we will increase the breadth and thus improve the model, as soon as more data become available. Using the current model and an independent test set, we obtained AUC = 0.747.
TensorFlow is an open source software library for neural network models made by Google. Recently RStudio released Keras an API for accessing TensorFlow in R. Keras enables fast experimentation - Being able to go from idea to result with the least possible delay is key to doing good research.