KnockNet - Modeling changes in transcriptomic state from a massive library of cellular signatures

Charles Blatti[blatti@illinois.edu], Casey Hanson[crhanso2@illinois.edu], Bryce Kille[kille2@illinois.edu], and Saurabh Sinha[sinhas@illinois.edu]

KnowEnG BD2K Center of Excellence
University of Illinois Urbana-Champaign

Table of Contents

1. Motivation
2. Installation
3. Tutorial
   1. Creating Gene Expression Datasets
   2. Example Training Runs
   3. Example Model Prediction
4. KnockNet Resources
   1. Parameters
   2. Input File Formats
   3. Output File Formats

Motivation

Learning signatures of gene knockdowns from paired transcriptional states.

Return to TOC

Installation

Local copy of KnockNet repository

If you wish to use the sample files necessary to complete this tutorial or the datasets from the paper, first clone this repository from github:

git clone https://github.com/knoweng-research/knocknet.git

Installing Tensorflow

Examples for creating the environment with tensorflow installed can be found here.

Return to TOC

Tutorial

This section of the README is meant to walk a user through a process of using Knocknet to find the related genes knockdown signature of their paired trancriptional datasets using the L1000 trained model or training a new model on an additional dataset.

Creating Gene Expression Datasets

Examples for pulling data from L1000 and formatting it into an appropriate dataset for training or testing can be found here.

Example Training Runs

Now that we have our data set prepared, we are ready to train the knocknet model. Examples for training the model can be found here.

Example Model Prediction

Now that we have a trained model, we are read to use it to predict gene knockout signatures in pair tumor progression datasets as shown here

Return to TOC

KnockNet Resources

Parameters

usage: train.py [-h] [--train_chkpt_dir TRAIN_CHKPT_DIR] [--log_dir LOG_DIR]
                [--data_dir DATA_DIR] [--data_batch_size DATA_BATCH_SIZE]
                [--data_serialized] [--data_mode DATA_MODE]
                [--data_batch_norm] [--training]
                [--train_max_steps TRAIN_MAX_STEPS]
                [--train_save_ckpt_secs TRAIN_SAVE_CKPT_SECS]
                [--train_save_summ_secs TRAIN_SAVE_SUMM_SECS]
                [--train_optimizer_str TRAIN_OPTIMIZER_STR]
                [--train_learning_rate TRAIN_LEARNING_RATE]
                [--reg_do_keep_prob REG_DO_KEEP_PROB]
                [--reg_kl_sparsity REG_KL_SPARSITY]
                [--reg_kl_scale REG_KL_SCALE] [--reg_l1_scale REG_L1_SCALE]
                [--reg_l2_scale REG_L2_SCALE] [--conv_depth CONV_DEPTH]
                [--conv_actv_str CONV_ACTV_STR] [--conv_batch_norm]
                [--conv_gene_pair] [--fcs_dimension_str FCS_DIMENSION_STR]
                [--fcs_actv_str FCS_ACTV_STR] [--fcs_batch_norm]
                [--fcs_res_block_size FCS_RES_BLOCK_SIZE]
                [--out_actv_str OUT_ACTV_STR]
                [--out_label_count OUT_LABEL_COUNT]
                [--eval_run_mode EVAL_RUN_MODE]
                [--eval_interval_secs EVAL_INTERVAL_SECS]

optional arguments:
  -h, --help            show this help message and exit

global parameters:
  --train_chkpt_dir TRAIN_CHKPT_DIR
                        Directory for saved model
  --log_dir LOG_DIR     Directory for summary and logs

data parameters:
  --data_dir DATA_DIR   Directory storing input data
  --data_batch_size DATA_BATCH_SIZE
                        Number of examples in a minibatch
  --data_serialized     Default .data files or serialized .tfrecords
  --data_mode DATA_MODE
                        Which part of the data to use: "all", "diff",
                        "exp_only"
  --data_batch_norm     Batch normalize outputs of in layer

training parameters:
  --training            Mark if training model
  --train_max_steps TRAIN_MAX_STEPS
                        Number of steps to run trainer
  --train_save_ckpt_secs TRAIN_SAVE_CKPT_SECS
                        Number of seconds to run trainer before saving ckpt
  --train_save_summ_secs TRAIN_SAVE_SUMM_SECS
                        Number of seconds to run trainer before saving
                        summaries
  --train_optimizer_str TRAIN_OPTIMIZER_STR
                        Optimizer function
  --train_learning_rate TRAIN_LEARNING_RATE
                        Initial learning rate

regularization parameters:
  --reg_do_keep_prob REG_DO_KEEP_PROB
                        Keep probability for training dropout
  --reg_kl_sparsity REG_KL_SPARSITY
                        Probability for activation of neuron
  --reg_kl_scale REG_KL_SCALE
                        Regulariztion scalar for KL activation loss
  --reg_l1_scale REG_L1_SCALE
                        Regulariztion scalar for L1 weights loss
  --reg_l2_scale REG_L2_SCALE
                        Regulariztion scalar for L2 weights loss

convolutional_layer parameters:
  --conv_depth CONV_DEPTH
                        Depth of convolutional_layer
  --conv_actv_str CONV_ACTV_STR
                        Activiation function
  --conv_batch_norm     Batch normalize outputs of conv layer
  --conv_gene_pair      Do gene pair convolution

fully_connected_layer parameters:
  --fcs_dimension_str FCS_DIMENSION_STR
                        Hidden layers of the model
  --fcs_actv_str FCS_ACTV_STR
                        Activiation function
  --fcs_batch_norm      Batch normalize outputs of fc layer
  --fcs_res_block_size FCS_RES_BLOCK_SIZE
                        Number of layers per residual block

output_layer parameters:
  --out_actv_str OUT_ACTV_STR
                        Activiation function
  --out_label_count OUT_LABEL_COUNT
                        Number of classes for training

evaluation and summary parameters:
  --eval_run_mode EVAL_RUN_MODE
                        Mode: "stats", "preds", "probs"
  --eval_interval_secs EVAL_INTERVAL_SECS
                        Time between sampling the evaluation metrics

Relating to input data files

Parameter	Type	Default	Description
data_dir	string	'./'	Directory storing input data.
data_batch_size	int	128	Number of examples in a minibatch.
data_serialized	bool	False	Default .data files or serialized .tfrecords.
data_mode	str	'all'	Which part of the data to use: "all", "diff", "exp_only".
data_batch_norm	bool	False	Batch normalize outputs of in layer.

Input File Formats

Expression .data format

An example input file in the correct data format can be found here

Metadata about data set

An example metadata file in the correct format can be found here

Serialize Format

Serialize examples are found here

Output File Formats

.preds Prediction File

Columns	Type	Description
true_labels	int	True class id of example
pred_labels	int	Predicted class id of example
correct_bool	bool	True if true class id matches predicted class id
true_prob_score	float	Predicted probability of the true class
pred_prob_score	float	Predicted probability of the predicted class
entropy	float	Entropy of the example's probability distribution over class states

.probs Probability File

Each row of this file contains the probability distribution over class states for an example

Return to TOC