DeepSeqProt

DeepSeqProt is an unsupervised deep learning framework for protein sequence data

---

DEPENDENCIES

Nonstandard Python Libraries:

• PyTorch
• NumPy
• pandas
• Biopython

These can be installed by running conda install --file requirements.txt

---

USAGE

> python DeepSeqProt.py -i [input] -o [output] -t [train] [options]

-i [fasta file]         Input fasta file, each sequence will be passed through the trained model [REQUIRED]
-o [string]             Name of the output directory and prefix of several output files [REQUIRED]
-t [fasta | model file] Either a fasta file of sequences used to train the model or a previously trained model file [REQUIRED]
-w [integer list]       List of integers representing the length and width of the encoder architecture [2000, 1500, 1000, 500, 1]
-n [integer]            Number of discerete embeddings used to populate latent space [1000]
-a [integer]            Number of latent space dimensions, this paramater introduces a simple convolutional layer to transform the default 20 channels into 'a' channels prior to quantization
-c [float]              Beta, hyperparameter used to weight commitment loss [0.001]
-d [float]              Weight decay [0.9]
-b [integer]            Batch size [32]
-l [float]              Learning rate [1e-4]
-m [float]              Maximum number of training updates [1e6]
-h                      Print help

---

OUTPUT

Contents of output directory:

• fastas/ Directory containing fasta files of each embedding with their associated input sequences

• output_coordinates.txt Assigned embedding and coordinates in latent space for each sequence in tab delimted format

• output.pt Pickled PyTorch object of the trained model which can be used for future runs

• output_log.txt Log file

---

VALIDATION & BENCHMARKING

For the sake of reproducibility we also include several scripts used to validate and benchmark DeepSeqProt during development

run_benchmarks.sh

> ./run_benchmarks.sh -i [fasta file]

Will run each clustering algorithm used during development to benchmark DeepSeqProt and format it into a *_coordinates.txt file which can be used by validate.py. Note that each algorithm (CD-HIT, MMseqs2, MCL) must be installed and included in your PATH

validate.py

> python validate.py -i [coordinates file] -r [reference table]

Calculates several summary statistics from a *_coordinates.txt file generated either by DeepSeqProt.py or run_benchmarks.sh. Requires a tab-delimited reference table in the format:

Entry	Organism	Protein families	Gene ontology IDs

for each sequence in *_coordinates.txt. We recommend using the Uniprot database for this purpose, an example table can be found here. In addition to the libraries used by DeepSeqProt validate.py also requires the nonstandard Python libraries scikit-learn and goatools