Supplementary File 1 Additional GPU-powered Tools for Bioinformatics and Computational Biology

A field of Computational Biology where GPU acceleration can yield a relevant speed-up is related to the analysis of spectral data derived from, e.g., mass-spectrometry experiments. FastPaSS [4] is a tool that accelerates the identification of a spectrum in a spectral library by means of the SpectraST similarity scoring algorithm [14]. The core of the matching algorithm is represented by a dot product of two vectors corresponding to the normalized intensities of the spectra, a calculation that is well-suitable for GPU acceleration. According to the results shown in [4], using a Nvidia GeForce 8600 GTS, FastPaSS allows a 8× speed-up with respect to a sequential execution optimized using a pre-caching of files. Tempest [19] extends the task of spectral matching by performing the whole similarity scoring— including the generation of theoretical fragmented spectra—directly on the GPU. Tempest exploits the possibility of asynchronous execution of CUDA kernels to implement a heterogeneous execution scheme, in which the CPU performs database digestion while the GPU performs the scoring. Tempest implements two scoring functions: the first based on cross-correlation, and the second on dot product. According to [19], Tempest’s accelerated scoring function based on the dot product allows a two orders of magnitude speed-up with respect to the correlation-based method. Still, the approach based on correlation is more accurate and Tempest’s speed-up is about 10× with respect to the CPU, using a Nvidia GeForce GTX 480. The acceleration of similarity scoring by means of GPUs was also investigated in [16]. Here, it was shown that the spectral dot product can be strongly accelerated by intensively exploiting the shared memory and parallel reduction techniques, with a two-order magnitude speed-up using a Nvidia GeForce GTX 280. A different example of GPU-powered spectral analysis concerns feature detection, used to identify and quantify the proteins contained in a sample; in particular, features must be separated from signal noise and baseline artifacts. Hussong et al. [11] employed an adaptive wavelet transform for this task, a process that can be parallelized on the GPU. Their implementation exploits the shared memory to store the spectrum, allowing a 200× speed-up on real world data sets. Since the methodology was implemented on an early CUDA architecture (namely, Nvidia Tesla C870), the amount of available memory was limited to 16 KB (see Table 2 in Supplementary File 2). Thus, for larger signals, this tool automatically switches to texture memory, affecting the speed-up.