Autoencoder based local T cell repertoire density can be used to classify samples and T cell receptors

Abstract

Recent advances in T cell repertoire (TCR) sequencing allow for the characterization of repertoire properties, as well as the frequency and sharing of specific TCR. However, there is no efficient measure for the local density of a given TCR. TCRs are often described either through their Complementary Determining region 3 (CDR3) sequences, or theirV/J usage, or their clone size. We here show that the local repertoire density can be estimated using a combined representation of these components through distance conserving autoencoders and Kernel Density Estimates (KDE). We present ELATE–an Encoder-based LocAl Tcr dEnsity and show that the resulting density of a sample can be used as a novel measure to study repertoire properties. The cross-density between two samples can be used as a similarity matrix to fully characterize samples from the same host. Finally, the same projection in combination with machine learning algorithms can be used to predict TCR-peptide binding through the local density of known TCRs binding a specific target.

Counting is Almost All You Need

“…The network is trained with a dropout of 0.2 and a ReLU. An MSE loss function is implemented to compare each input sequence to the resulting decoded sequence (13). The current version differs from the ELATE encoder ((13)), since it includes a variational term.…”

“…An MSE loss function is implemented to compare each input sequence to the resulting decoded sequence (13). The current version differs from the ELATE encoder ((13)), since it includes a variational term. Instead of encoding an input as a single point, we encode it as a distribution over the latent space.…”

“…A TCR autoencoder is a model that preserves the information about input 𝑡 𝑖 V gene and CDR3 sequence , while reducing the dimension to a low dimension representation 𝑧 𝑖 . The training of the TCR autoencoder includes several steps of data processing (13). The first step is representing each of the amino acid per position as well as the V genes by an embedding vector.…”

“…We here show that counting arguments actually produce better results than the current SOTA ML or Bayesian methods. We then further improve on that by including the similarity between different TCRs using the combination of a Variational AutoEncoder (VAE) (13), and a novel attention model to include not only the relative importance of positive samples, but also their quantity, named attTCR (attention TCR). Finally, we propose an intermediate solution between the counting and attTCR - gTCR that uses a graph of the TCR repertoire co-occurences to predict the class of a sample.…”

“…To address the similarity between TCRs, one can use either a sequence similarity (how similar are 237 the TCR CDR3 and V sequence), or a functional similarity (how often they co-appear in the same 238 sample). For the sequence similarity, we projected each sequence using an improvement of the 239 ELATE (Encoder based LocAl Tcr dEnsity) TCR autoencoder (13). ELATE was enlarged to become a 240 cyclic variational autoencoder, and the TCR representation method was improved (see Methods).…”