CAS@home

CAS@home
Dark mode version of CAS logo
Protein structure prediction, the kind of computation TreeThreader performed on CAS@home
Project
Status	Completed
Category	Biology and Medicine
Compute	CPU
Requires	None
Development
Developer	Institute of High Energy Physics, Chinese Academy of Sciences
Author	Wu Wenjing
Sponsor	Chinese Academy of Sciences
Initial release	September 30, 2010  (16 years ago)
Completed	April 27, 2020  (6 years ago)
Software
Written in	C
Operating system	Windows, Linux
BOINC statistics
Stats as of	August 8, 2019  (7 years ago)
Performance	409 GigaFLOPs
Active users	194
Total users	16,759
Active hosts	430
Total hosts	37,289
Analytics
GPU performance	0
CPU performance	1.3 TFLOPS
Metadata
Website	http://casathome.ihep.ac.cn/

CAS@home (Chinese Academy of Sciences at Home) was a volunteer computing project built on BOINC and hosted by the Computing Centre of the Institute of High Energy Physics (IHEP), part of the Chinese Academy of Sciences (CAS) in Beijing. Launched in 2010, it was the first volunteer computing project based in mainland China, and was created to introduce Chinese researchers to volunteer computing while donating spare processing power from volunteers around the world to scientists at CAS and other Chinese research institutes. Over its lifetime the project hosted several scientific applications, but it became best known for TreeThreader, a protein structure prediction program developed by researchers at the Institute of Computing Technology (ICT), CAS, which used the donated computing power to identify the most likely three-dimensional shape of a protein from its amino acid sequence.

History

CAS@home was officially launched in January 2010 with the backing of the Chinese Academy of Sciences, and a dedicated BOINC server was established at IHEP during the first quarter of that year. A hands-on volunteer computing workshop with international experts and more than twenty participants was held at IHEP on 9 March 2010 to introduce the concept to Chinese researchers. The project itself, announced publicly on the BOINC website on 7 September 2010, described itself as encouraging Chinese scientists to adopt volunteer computing and volunteer thinking for their research, through workshops, pilot applications, and Chinese-language outreach.

The first application made available on the project was SCThread (also referred to as Short-Cut Threading), a protein structure prediction tool developed by scientists at the Institute of Computing Technology, CAS, with testing of volunteer jobs beginning in mid-2010. Tsinghua University's Centre for Micro and Nano Mechanics also prepared a fluid and solid motion simulation application for the platform, and IHEP physicists worked on an application named BOSS for simulating particle collisions at the Beijing Electron-Positron Collider (BEPC), using VirtualBox-based virtualization.

The project's day-to-day maintenance and development was coordinated by Wu Wenjing of the IHEP Computing Centre, working with Wu Jie and Kan Wenxiao, under the leadership of Chen Gang, with additional support from David Anderson of the Space Sciences Laboratory at UC Berkeley, Francois Grey of Tsinghua University and the Citizen Cyberscience Centre at CERN, and Lei Yang of Tsinghua University.

By 2013, the main scientific application running on CAS@home was TreeThreader, a protein structure prediction program written in C by researchers at ICT, which compared an input protein sequence against a library of roughly fifty thousand known protein structural templates to identify the closest match.

A ribbon diagram of a protein's three-dimensional structure, the kind of output TreeThreader generated from an input amino acid sequence

At that time, the project had accumulated around 23,000 active hosts contributing roughly 1.3 TFLOPS of real-time computing power, with around 10,000 active users (about a third of them from China), and had validated 7 million CPU hours of work since November 2012, with a peak of one million CPU hours validated in a single month. Wu Wenjing of IHEP presented this status update at the BOINC Workshop in Grenoble in 2013.

In late 2013 and into 2014, the project ran successive large computing campaigns for TreeThreader, including a batch of fifty thousand protein sequences completed in early 2014, with a further round of around eighty thousand sequences planned to follow. A network problem between ICT and IHEP interrupted job submission for TreeThreader in mid-2013, requiring a temporary pause in job generation while the issue was diagnosed.

FALCON@home and CASP participation

TreeThreader's underlying remote-homologue recognition approach was published by the project's developers as TOPO, a method for improving protein fold recognition by identifying common structural frameworks shared between a query protein and previously solved structures, even when sequence similarity to those structures was low. By extracting the conserved regions of a structural template and aligning the query sequence against those conserved regions rather than against the full-length template, the approach avoided the unreliable alignments that more variable template regions tend to produce in conventional threading methods.

This work was deployed as a public-facing service named FALCON@home (also written FALCON_TOPO), which let biologists worldwide submit their own protein sequences for structure prediction over the web, using CAS@home as its volunteer computing backend. At its peak the service drew on more than twenty thousand volunteer CPUs and was able to process over a thousand protein sequences per day, a throughput its developers presented as a way to keep up with the rapidly growing number of protein sequences in need of structural annotation.

In the 11th Critical Assessment of protein Structure Prediction (CASP11) in 2014, the FALCON@home-based server FALCON_TOPO was ranked 12th in the template-based modeling category and 17th in the free modeling category among competing structure prediction groups, while an enhanced variant called FALCON_EnvFold was ranked 9th in the template-based modeling category. Prediction quality in CASP is commonly assessed using the TM-score, a measure of structural similarity between a predicted model and the experimentally solved structure that ranges from 0 to 1:

$\text{TM-score}=\max \left[\frac{1}{L_{target}}\sum _{i=1}^{L_{aligned}}\frac{1}{1+{\left(\frac{d_i}{d_0}\right)}^2}\right]$

where $L_{target}$ is the number of residues in the target protein, $L_{aligned}$ is the number of aligned residue pairs, $d_i$ is the distance between the $i$th pair of aligned residues, and $d_0$ is a length-dependent scaling factor. For one CASP11 target the developers highlighted as an example of FALCON@home's strength in remote homologue identification, the server's prediction reached a TM-score of 0.84 against the experimentally determined structure. The developers also used FALCON@home to predict structures for 6,033 mouse proteins in six days, including hundreds of mitochondrial proteins, and reported a strong negative correlation between a protein's surface residue ratio and its half-life, suggesting the platform's usefulness for large-scale proteome-wide studies as well as individual structure prediction.

Decline and retirement

After the large TreeThreader campaign completed in October 2015, the project's computing activity slowed considerably, with only sporadic batches of around two hundred sequences processed each month, largely driven by submissions through the FALCON public service rather than dedicated CAS@home campaigns. A further TreeThreader campaign was planned between February and May 2016 to help prepare results ahead of that year's CASP competition. No further public updates were issued after this period, and the project's website and BOINC server eventually went offline. BOINCstats, a cross-project statistics tracker, subsequently marked CAS@home as a retired project once its scheduler stopped responding.

Scientific publications

Publications using BOINC-computed data

1. Wang, Chao, Haicang Zhang, Wei-Mou Zheng, Dong Xu, Jianwei Zhu, Bing Wang, Kang Ning, Shiwei Sun, Shuai Cheng Li and Dongbo Bu. FALCON@home: a high-throughput protein structure prediction server based on remote homologue recognition. Bioinformatics 32, no. 3 (2016): 462 to 464.

Other project-related publications

1. Zhu, Jianwei, Haicang Zhang, Chao Wang, Bin Ling, Wei-Mou Zheng and Dongbo Bu. TOPO: Improving remote homologue recognition via identifying common protein structure framework. arXiv preprint arXiv:1507.03197 (2015).