High throughput multi-dimensional scaling (HiT-MDS)

High-Throughput Multidimensional Scaling (HiT-MDS)

Supplementary information to the paper submitted to German Conference on Bioinformatics (GCB) 2009:
CUDA-based multi-core implementation of MDS-based bioinformatics algorithms.
Thilo Fester, Falk Schreiber, and Marc Strickert.


Fig.1: Visualization of HiT-MDS (CUDA version). Left: Gene expression of 4824 genes covering 14 time points, embedded in 3D. Right: Yeast network of 4554 nodes, embedded in 3D from adjacency matrix.

Background

Multidimensional scaling (MDS) is a versatile method for

dimension reduction (= nonlinear alternative to projection to principal components),
reconstruction of data dissimilarity relationships (matrix) in the Euclidean output space,
conversion of one metric space (such as data compared by Manhattan distance) into Euclidean space,
dealing with missing data relationships (= zero force assumption).

High-Throughput MDS is an optimized version for rapid distance reconstruction, based on correlations of distances between input and output space.

Current batch version: fast and recommendable.

Versions available: MATLAB/GNU Octave, R and CUDA
Theoretical caveat: cannot reconstruct strict distance relationships of simplex edge points.

Previous serial version: slower, but without constraint to non-simplex data. [ Read more ].

Get batch HiT-MDS

CUDA version at [ sourceforge ] with SVN [ access ]

R: hitmds-r.zip (224 kb) Readme

MATLAB / GNU Octave: hitmds-ml.zip (227 kb) Readme

GPU enhanced MATLAB version (292 kb) using the Jacket plugin of AccelerEyes(R) for incredibly easy GPU computing.

Previous serial version of HiT-MDS

Home page related to previous slower serial version of HiT-MDS for functional genomics of developing barley seeds is found [ here ].

Credits

Y.-h Taguchi's MDS page [ powder page ]
GGobi visualization system [ GGobi / RGobi ]
John Aldo Lee's [ quality criteria ] and Jarkko Venna's [ work ]

	Marc Strickert, May 2009