Living on the Edge: Supercomputing Powers Protein Analysis

New Berkeley Lab algorithm allows biologists to harness the capabilities of massively parallel supercomputers to make sense of a genomic ‘data deluge’

Did you know that the tools used for analyzing relationships between social network users or ranking web pages can also be extremely valuable for making sense of big science data? On a social network like Facebook, each user (person or organization) is represented as a node and the connections (relationships and interactions) between them are called edges. By analyzing these connections, researchers can learn a lot about each user—interests, hobbies, shopping habits, friends, etc.

In biology, similar graph-clustering algorithms can be used to understand the proteins that perform most of life’s functions. It is estimated that the human body alone contains about 100,000 different protein types, and almost all biological tasks—from digestion to immunity—occur when these microorganisms interact with each other. A better understanding of these networks could help researchers determine the effectiveness of a drug or identify potential treatments for a variety of diseases.

Today, advanced high-throughput technologies allow researchers to capture hundreds of millions of proteins, genes and other cellular components at once and in a range of environmental conditions. Clustering algorithms are then applied to these datasets to identify patterns and relationships that may point to structural and functional similarities. Though these techniques have been widely used for more than a decade, they cannot keep up with the torrent of biological data being generated by next-generation sequencers and microarrays. In fact, very few existing algorithms can cluster a biological network containing millions of nodes (proteins) and edges (connections).

That’s why a team of researchers from the Department of Energy’s (DOE’s) Lawrence Berkeley National Laboratory (Berkeley Lab) and Joint Genome Institute (JGI) took one of the most popular clustering approaches in modern biology—the Markov Clustering (MCL) algorithm—and modified it to run quickly, efficiently and at scale on distributed-memory supercomputers. In a test case, their high-performance algorithm—called HipMCL—achieved a previously impossible feat: clustering a large biological network containing about 70 million nodes and 68 billion edges in a couple of hours, using approximately 140,000 processor cores on the National Energy Research Scientific Computing Center’s (NERSC) Cori supercomputer. A paper describing this work was recently published in the journal Nucleic Acids Research.

“The real benefit of HipMCL is its ability to cluster massive biological networks that were impossible to cluster with the existing MCL software, thus allowing us to identify and characterize the novel functional space present in the microbial communities,” says Nikos Kyrpides, who heads JGI’s Microbiome Data Science efforts and the Prokaryote Super Program and is co-author on the paper. “Moreover we can do that without sacrificing any of the sensitivity or accuracy of the original method, which is always the biggest challenge in these sort of scaling efforts.”

“As our data grows, it is becoming even more imperative that we move our tools into high performance computing environments, ” he adds.  “If you were to ask me how big is the protein space? The truth is, we don’t really know because until now we didn’t have the computational tools to effectively cluster all of our genomic data and probe the functional dark matter.” 

To get a grip on this torrent of data, researchers rely on cluster analysis, or clustering. This is essentially the task of grouping objects so that items in the same group (cluster) are more similar than those in other clusters. For more than a decade, computational biologists have favored MCL for clustering proteins by similarities and interactions.

“One of the reasons that MCL has been popular among computational biologists is that it is relatively parameter free; users don’t have to set a ton of parameters to get accurate results and it is remarkably stable to small alterations in the data. This is important because you might have to redefine a similarity between data points or you might have to correct for a slight measurement error in your data. In these cases, you don’t want your modifications to change the analysis from 10 clusters to 1,000 clusters,” says Aydin Buluç, a CRD scientist and one of the paper’s co-authors.

But, he adds, the computational biology community is encountering a computing bottleneck because the tool mostly runs on a single computer node, is computationally expensive to execute and has a big memory footprint—all of which limit the amount of data this algorithm can cluster.

One of the most computationally and memory intensive steps in this analysis is a process called random walk. This technique quantifies the strength of a connection between nodes, which is useful for classifying and predicting links in a network. Because there are many different ways of traveling between nodes in a network, this step repeats numerous times. Algorithms like MCL will continue running this random walk process until there is no longer a significant difference between the iterations. It sounds simple enough, but for protein networks with millions of nodes and billions of edges, this can become an extremely computationally and memory intensive problem. With HipMCL, Berkeley Lab computer scientists used cutting-edge mathematical tools to overcome these limitations. 

The next step is to continue to rework HipMCL and other computational biology tools for future exascale systems, which will be able to compute quintillion calculations per second. This will be essential as genomics data continues to grow at a mindboggling rate—doubling about every five to six months.  This will be done as part of DOE Exascale Computing Project’s Exagraph co-design center.

This article has been republished from materials provided by Berkeley Lab Computing Sciences. Note: material may have been edited for length and content. For further information, please contact the cited source.

Reference: Azad, A., Pavlopoulos, G. A., Ouzounis, C. A., Kyrpides, N. C., & Buluç, A. (n.d.). HipMCL: a high-performance parallel implementation of the Markov clustering algorithm for large-scale networks. Nucleic Acids Research. https://doi.org/10.1093/nar/gkx1313