/DeepTCR

Deep Learning Methods for Parsing T-Cell Receptor Sequencing (TCRSeq) Data

Primary LanguageJupyter NotebookMIT LicenseMIT

DeepTCR

Deep Learning Methods for Parsing T-Cell Receptor Sequencing (TCRSeq) Data

DeepTCR is a python package that has a collection of unsupervised and supervised deep learning methods to parse TCRSeq data. To see examples of how the algorithms can be used on an example datasets, see the subdirectory 'tutorials' for a collection of tutorial use cases across multiple datasets. For complete documentation for all available methods, click here.

While DeepTCR will run with Tensorflow-CPU versions, for optimal training times, we suggest training these algorithms on GPU's (requiring CUDA, cuDNN, and tensorflow-GPU).

DeepTCR now has the added functionality of being able to analyze paired alpha/beta chain inputs as well as also being able to take in v/d/j gene usage and the contextual HLA information the TCR-Sequences were seen in (i.e. HLA alleles for a repertoire from a given human sample). For detailed instructions on how to upload this type of data, refer to the documentation for loading data into DeepTCR.

For questions or help, email: johnwilliamsidhom@gmail.com

Publication

For full description of algorithm and methods behind DeepTCR, refer to the following manuscript:

Sidhom, J. W., Larman, H. B., Pardoll, D. M., & Baras, A. S. (2021). DeepTCR is a deep learning framework for revealing sequence concepts within T-cell repertoires. Nat Commun 12, 1605

Dependencies

See requirements.txt for all DeepTCR dependencies. Installing DeepTCR from Github repository or PyPi will install all required dependencies. It is recommended to create a virtualenv and installing DeepTCR within this environment to ensure proper versioning of dependencies.

In the most recent release (DeepTCR 2.0, fifth release), the package now uses python 3.7 & Tensorflow 2.0. Since this has required an overhaul in a lot of the code, there could be some bugs so we would greatly appreciate if you post any issues to the issues page and I will do my best to fix them as quickly as possible. One can find the latest DeepTCR 1.x version under the v1 branch if you still want to use that version. Or one can specifically pip install the specific version desired.

Instructions on how to create a virtual environment can be found here: https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/

Installation

In order to install DeepTCR:

pip3 install DeepTCR

Or to install latest updated versions from Github repo:

Either download package, unzip, and run setup script:

python3 setup.py install

Or use:

pip3 install git+https://github.com/sidhomj/DeepTCR.git

Release History

1.1

Initial release including two methods for unsupervised learning (VAE & GAN). Also included ability to handle paired alpha/beta data.

1.2

Second release included major refactoring in code to streamline and share methods across classes. Included ability for algorithm to accept v/d/j gene usage. Added more analytical fetures and visualization methods. Removed GAN from unsupervised learning techniques.

1.2.7

On-graph clustering method introduced for repertoire classifier to improve classification performance.

1.2.13

Ability for HLA information to be incorporated in the analysis of TCR-Seq.

1.2.24

Added ability to do regression for sequence-based model.

1.3

Third release including improved repertoire classification architecture. Details in method will follow in manuscript.

1.4

Fourth release includes major refactoring of code and adding more features including:

  • Multi-Model Inference. When training the supervised sequence or repertoire classifier, in Monte-Carlo or K-Fold Cross Validation, a separate model will be stored for each cross-validation. When using the inference engine, users can choose to do an ensemble inference of some or many of the trained models.
  • HLA Supertype Integration. Previous versions allowed users to provide HLA alleles for additional dimension of featurization for the TCR. In this version, when providing HLA (either via the Get_Data or Load_Data methods), one now has the option of assigning the HLA-A and B genes to known supertypes for a more biologically functional representation of HLA.
  • VAE now has an optional method by which to find a minimal number of latent features to model the underlying distribution by incorporating a sparsity regularization on the latent layer. When using this feature, the VAE will provide a more compact latent space even if the initial latent_dim is unnecessarily high to model the distribution of data.
  • Supervised models now have an additional option to use Multi-Sample Dropout to improve training and generalization.
  • Incorporation of LogoMaker so now when Representative Sequences are generated along with enriched motifs, seq logos are made and saved directly in the results folder under Motifs.
  • Improved Motif Identification algorithm behind supervised method Representative_Sequences that uses a multinomial linear model to identify which motifs are associated to predicted probabilites from neural network.
  • Supervised Repertoire Model now able to do regression. By providing per-instance label with regression value with Load_Data method, this will automatically use the average of all instance level labels as the sample level value to regress the model against.

2.0

Fifth release:

  • Upgrading to use python 3.7 & Tensorflow 2.0
  • For large repertoires, we have incorporated the ability to randomly subsample the repertoire over the course of training. Two methods of sub-sampling exist. 1) Completely randomly sampled from across the entire repertoire vs 2) randomly sampled as a probability function of the frequency of the TCR (at the amino acid level), meaning that a TCR with a 25% frequency will be sample at that probability.

2.1.0

  • Upgrading to Tensorflow 2.7
  • Improved handling of inference with previously unseen V/D/J gene usage.
  • Improved computational efficiency for loading data from large files (~2x improvement in speed, 50% or more decrease in peak memory consumption)