/master_thesis

Primary LanguagePython

Deep Generative Models for the Computational Design of T-cell Receptor CDR3 sequences

This repository includes several Variational Autoencoder variants to fit the probability distribution and generate for T cell receptor beta chain CDR3 sequences.

Setting up environment

Our work is inspired by Vampire. Before running our VAE models, you should set up your environment as described in Vampire repository.

Preprocessing Data

  • Amino acid sequences are onehot-encoded into 30*21 matrix. Processing step can be finished by executing preprocess_adaptive.py in Vampire.
  • Nucleotide sequence are onehot-encoded into 90*5 matrix. Processing step can be finished by executing preprocess_nt.py.

Processed Data

The data we used to train the models are cohort2 from Immunosequencing identifies signatures of cytomegalovirus exposure history and HLA-mediated effects on the T-cell repertoire. Processed data can be downloaded from Zenodo.

Running

Before running .py scripts, you should execute

conda activate vampire

VAE-seq (vae_seq.py)

  • Input and output are CDR3 amino acid sequence.
  • Encoder and decoder consists dense layers.

Supervised VAE (svae.py)

  • Input data is CDR3 amino acid sequence and the corresponding V- and J- genes.
  • Output data is CDR3 amino acid sequence.
  • Encoder and decoder consists dense layers.

Nucleotide VAE (nt_vae.py)

  • Input data is CDR3 nucleotide sequence.
  • Output data is CDR3 amino acid sequence.
  • Encoder and decoder consists dense layers.

Recurrent neural network VAE (rnn_vae.py)

  • Input and output are CDR3 amino acid sequence.
  • Encode consists of bidirectional GRU layer, decoder consists of undirectional GRU layer.

Result analysis

The scripts that plot distance heatmap of generations' probability distribution and scatter of frequency estimation are modeller_plot.py and plot_heatmap.py.