Table of Contents

1. Topsy-Turvy
   1. Summary
   2. Files and Folders
      1. Generating PLM Embeddings
      2. Training the Model
      3. Evaluating Performance
   3. Saved Models

Topsy-Turvy

Summary

Topsy-Turvy is a novel method that synthesizes protein sequence information and inherent network structures transferrable across species to construct and/or enrich the PPI network for the target species.

For more information about the model architectures (and downloading the pretained models and datasets), go to [[]].

Files and Folders

All the relevant test and evaluation codes are found inside the `topsy_turvy` folder. The major files for training/evaluation are:

1. embedding.py
2. train.py
3. evaluate.py

Generating PLM Embeddings

`embedding.py` is used to produce the PLM embeddings from the input sequence file in fasta format. It can be run using

python embedding.py 
      [--seqs SEQ-FASTA-FILE] 
      [--o OUTPUT-DEST-FILE]
      [--d GPU-DEVICE-ID]

Training the Model

`train.py` is used to train the model, given the sequence and network information for a source network. It can be run using

python train.py [-h] 
      [--pos-pairs POS_PAIRS]               # Positive edgelist for training 
      [--neg-pairs NEG_PAIRS]               # Negative edgelist for training
      [--pos-test POS_TEST]                 # Positive edgelist for testing 
      [--neg-test NEG_TEST]                 # Negative edgelist for testing
      [--embedding EMBEDDING]               # PLM embedding obtained using `embedding.py`
      [--augment]                           # If (p, q) in training edgelist, add (q,p) for training too
      [--protein-size PROTEIN_SIZE]         # Maximum protein size to use in training data: default = 800
      [--projection-dim PROJECTION_DIM]     # Dimension of the projection layer: default 100
      [--dropout-p DROPOUT_P]               # Parameter p for the embedding dropout layer
      [--hidden-dim HIDDEN_DIM]             # Number of hidden units for comparison layer in contact prediction
      [--kernel-width KERNEL_WIDTH]         # The width of the conv. filter for contact prediction
      [--use-w]                             # Use the weight matrix in the interaction prediction or not
      [--do-pool]                           # Use the max pool layer
      [--pool-width POOL_WIDTH]             # The size of the max pool in the interaction model
      [--sigmoid]                           # Use sigmoid activation at the end of the interaction model: Default false
      [--negative-ratio NEGATIVE_RATIO]     # Number of negative training samples for each positive training sample
      [--epoch-scale EPOCH_SCALE]           # Report the heldout performance every multiple of this many epochs 
      [--num-epochs NUM_EPOCHS]             # Total number of epochs
      [--batch-size BATCH_SIZE]             # Minibatch size 
      [--weight-decay WEIGHT_DECAY]         # L2 regularization
      [--lr LR]                             # Learning rate
      [--lambda LAMBDA_]                    # The weight on the similarity objective
      # Use these parameter for Topsy-turvy training 
      [--use_glider]                        # Use this to train with Topsy-Turvy.
      [--glider_param GLIDER_PARAM]         # g_t param: default 0.2 
      [--glider_thresh GLIDER_THRESH]       # g_p param: Default 92.5
      # Output and device information
      [-o OUTPUT] 
      [--save-prefix SAVE_PREFIX] 
      [-d DEVICE]
      [--checkpoint CHECKPOINT] 

In order to use the `train.py` in Topsy-Turvy mode, add `–use_glider` option in the train.py.

Evaluating Performance

`evaluate.py` is used to perform evaluation, given the positive and negative test examples. It can be run using

python evaluate.py
       [--model MODEL]                        # Model Location 
       [--pos-pairs POS_PAIRS]
       [--neg-pairs NEG_PAIRS]                # Positive and Negative Pairs
       [--embeddings EMBED]                   # PLM embedding file
       [--fasta FASTA]                        # If there is no PLM file, use FASTA to generate the PLM internally
       [--outfile OUTFILE]                    # Output prefix
       [--device DEVICE]                      # GPU device
       [--plot-prediction-distributions]
       [--plot-curves]

Saved Models

The best pre-trained models for D-SCRIPT and Topsy-Turvy are provided in the folder `Pretrained-Models`