concise 0.4.3

CONCISE (COnvolutional Neural for CIS-regulatory Elements) is a model for predicting PTR features like mRNA half-life from cis-regulatory elements using deep learning.

CONCISE

CONCISE (COnvolutional neural Network for CIS-regulatory Elements) is a model for predicting any quatitative outcome (say mRNA half-life) from cis-regulatory sequence using deep learning.

• Developed by the Gagneur Lab (computational biology): https://www.gagneurlab.in.tum.de

• Free software: MIT license

• Documentation: https://concise-bio.readthedocs.io

Features

• Very simple API

• Serializing the model to JSON - allows to analyze the results in any langugage of choice

• Helper function for hyper-parameter random search

• CONCISE uses TensorFlow at its core and is hence able of using GPU computing

Installation

After installing the following prerequisites:

1. Python (3.4 or 3.5) with pip (see Python installation guide and pip documentation)

2. TensorFlow python package (see TensorFlow installation guide or Installing Tensorflow on AWS GPU-instance)

install CONCISE using pip:

pip install concise

Getting Started

import pandas as pd
import concise

# read-in and prepare the data
dt = pd.read_csv("./data/pombe_half-life_UTR3.csv")

X_feat, X_seq, y, id_vec = concise.prepare_data(dt,
                                                features=["UTR3_length", "UTR5_length"],
                                                response="hlt",
                                                sequence="seq",
                                                id_column="ID",
                                                seq_align="end",
                                                trim_seq_len=500,
                                              )

######
# Train CONCISE
######

# initialize CONCISE
co = concise.Concise(motif_length = 9, n_motifs = 2,
                     init_motifs = ("TATTTAT", "TTAATGA"))

# train:
# - on a GPU if tensorflow is compiled with GPU support
# - on a CPU with 5 cores otherwise
co.train(X_feat[500:], X_seq[500:], y[500:], n_cores = 5)

# predict
co.predict(X_feat[:500], X_seq[:500])

# get fitted weights
co.get_weights()

# save/load from a file
co.save("./Concise.json")
co2 = Concise.load("./Concise.json")

######
# Train CONCISE in 5-fold cross-validation
######

# intialize
co3 = concise.Concise(motif_length = 9, n_motifs = 2,
                      init_motifs = ("TATTTAT", "TTAATGA"))

cocv = concise.ConciseCV(concise_object = co3)

# train
cocv.train(X_feat, X_seq, y, id_vec,
           n_folds=5, n_cores=3, train_global_model=True)

# out-of-fold prediction
cocv.get_CV_prediction()

# save/load from a file
cocv.save("./Concise.json")
cocv2 = ConciseCV.load("./Concise.json")

Where to go from here:

• See the example file scripts/example-workflow.py

• Read the API Documenation https://concise-bio.readthedocs.io/en/latest/documentation.html

History

0.1.0 (2016-09-15)

• First release on PyPI.

0.1.1 (2016-09-17)

• Minor documentation changes

• Renamed some internal variables

0.2.0 (2016-09-21)

• Introduced new feature: regress_out_feat

• Major renaming of variables for concistency

0.3.0 (2016-11-30)

• Added L-BFGS optimizer in addition to Adam. Use optimizer=”lbfgs” in Concise()

0.3.1 (2016-11-30)

• New function: best_kmers for motif efficient initialization

0.4.0 (2017-02-07)

• refactor: Removed regress_out feature

• feature: multi-task learning

0.4.1 (2017-02-09)

• bugfix: multi-task learning

0.4.2 (2017-02-09)

• same as 0.4.1 (pypi upload failed for 0.4.1)

0.4.3 (2017-02-09)

• feat: added early_stop_patience argument