LSTM_protein

jLSTM_protein

This is the recommended new JAVA-package of "Long Short-Term Memory" for Protein classification (jLSTM_protein). The implementation of the LSTM neural network is the same as in the C package and performs and behaves identically but faster. jLSTM_protein is multithreaded and therefore uses effectively multicore and -processor machines. Using more than one thread results in faster computation as in the C package. Each thread individually computes the gradients with a local weight matrix and updates a global weight matrix performing asynchronous stochastic gradient. jLSTM_protein uses Biojava 1.6 for reading FASTA sequences. The Biojava package is included and needs no separate installation.

Download, tryout the software implementation with examples (Biojava included) jLSTM.tar.bz2
Please read the Documentation to get information about this software.
For Remote Homology detection experiments adding positive training sequences with PSI-BLAST searches gives best results. To do this we offer a complete environment based on Perl and NCBI PSI-BLAST. Here is the README.
For the SCOP 1.53 benchmark here are the datasets. The positive training sequences are already extended. datasetsSCOP1.53.tar.gz (gzip 39,7 MB) or datasetsSCOP1.53.tar.bz2 (bzip2 30,8 MB)

SCOP 1.67 benchmark datasets (Positive training sequences are already extended (UniRef50)): SCOP167-superfamily.tar.bz2 SCOP167-fold.tar.bz2

Results for SCOP 1.67 benchmark (Average ROC and ROC50 scores):

	Superfamily level		Fold level
	ROC	ROC50	ROC	ROC50

GPkernel	0.902	0.591	0.844	0.514
GPextended	0.869	0.542	0.753	0.371
GPboost	0.797	0.375	0.688	0.298
SVM-Pairwise	0.849	0.555	0.724	0.359
Mismatch	0.878	0.543	0.814	0.467
eMOTIF	0.857	0.551	0.698	0.308
LA-kernel	0.919	0.686	0.834	0.504
PSI-BLAST	0.575	0.175	0.501	0.010
jLSTM	0.942	0.773	0.821	0.571

Håndstad et al. BMC Bioinformatics 2007 8:23 doi:10.1186/1471-2105-8-23.

(j)LSTM as Logistic Regression with the Spectrum Kernel (new)

LSTM logistic regression / spectrum kernel is a stripped down LSTM which can be interpreted as logistic regression with the spectrum kernel for sequence classification. For any step in the DNA sequence and a given k a k-mer string vector is build and fed into the network. The LSTM architecture is just two memory cells and no input- or output gates. The memory cells are not connected with each other. The squashing function h is the identity function. LSTM in this version weighs important k-mers for the classification and therefore can be used as an additional pattern recognizer based on k-mers.

The software is included in the jLSTM_protein package. (see above)
There is a README with further information.
A DNA data set is included for testing.

LSTM_protein

We offer a first C-package of "Long Short-Term Memory" for Protein classification (LSTM_protein).

Please read the Documentation to get information about this software.
A short description how to install this software.
Download, compile and tryout the software implementation with examples in tar.gz format LSTM.tar.gz
For Remote Homology detection experiments adding positive training sequences with PSI-BLAST searches gives best results. To do this we offer a complete environment based on Perl and NCBI PSI-BLAST. Here is the README.
For the SCOP benchmark here are the datasets. The positive training sequences are already extended. datasetsSCOP1.53.tar.gz (gzip 39,7 MB) or datasetsSCOP1.53.tar.bz2 (bzip2 30,8 MB)

Please cite:

Sepp Hochreiter, Martin Heusel, and Klaus Obermayer. "Fast Model-based Protein Homology Detection without Alignment." Bioinformatics 2007; doi: 10.1093/bioinformatics/btm247. Abstract

Updated 08/12/2010

LICENSE, WARRANTY, AND LIABILITY
This programm is freely available under the GNU General Public License (GPL).
You should have received a copy of the GNU General Public License along with this program; if not, write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.