ParaBTM: A Parallel Processing Framework for Biomedical Text Mining on Supercomputers.

A prevailing way of extracting valuable information from biomedical literature is to apply text mining methods on unstructured texts. However, the massive amount of literature that needs to be analyzed poses a big data challenge to the processing efficiency of text mining. In this paper, we address this challenge by introducing parallel processing on a supercomputer. We developed paraBTM, a runnable framework that enables parallel text mining on the Tianhe-2 supercomputer. It employs a low-cost yet effective load balancing strategy to maximize the efficiency of parallel processing. We evaluated the performance of paraBTM on several datasets, utilizing three types of named entity recognition tasks as demonstration. Results show that, in most cases, the processing efficiency can be greatly improved with parallel processing, and the proposed load balancing strategy is simple and effective. In addition, our framework can be readily applied to other tasks of biomedical text mining besides NER.

1. Introduction

With the rapid development of biotechnology, the amount of biomedical literature is growing exponentially. For instance, PubMed (https://www.ncbi.nlm.nih.gov/pubmed/), the most recognized biomedical literature database, indexes over 28 million entries for biomedical literature. Most of that information is presented in the form of unstructured texts. It is almost impossible for any domain expert to digest such a massive amount of information within a short period of time. Therefore, automated tools are essential for a systematic understanding of literature. To deal with the literature big data challenge, text mining methods are commonly applied to extract relevant knowledge from vast amounts of literature, and this has become a prominent trend in recent years [1].

Typical tasks of biomedical text mining include named entity recognition and relation extraction. One of the most fundamental tasks of biomedical text mining is named entity recognition (NER). Its task is to recognize target entities that represents key concepts from unstructured biomedical texts, such as proteins, genes, mutations, diseases, etc. There are some existing start-of-art biomedical tools that use text mining methods to identify some specific types of entities, such as mutations [2,3], genes [4,5], and diseases [6,7]. Most of these tools can achieve satisfactory recognition performance (F score over 80%) on standard datasets.

Relation extraction (RE) is a process that typically follows NER and aims to discover semantic connections between entities. Nowadays, there are a number of RE tools using different methods to identify biomedical entity interactions [8,9], such as drug–gene relationships [10,11,12], gene–disease relationships [9,13,14] and protein–protein interaction [15]. Some of them can achieve high F scores (over 80%) on several annotated datasets.

NER and RE are the preliminary steps in mining information from literature. With the uncovered facts, it is possible to construct a complex knowledge graph, which can assist new knowledge discovery and hypotheses generation. In order to achieve this goal, it is necessary to process as many articles as possible. However, text mining procedures are time consuming. BioContext, for instance, an integrated text mining system for large-scale extraction and contextualization of biomolecular events, took nearly 3 months to complete a full run of the system, which analyzed 20 million MEDLINE abstracts and several hundred thousand PMC open access full-texts using 100 concurrent processes [16]. In addition, some text mining tools, like GNormPlus [17] require a substantial amount of memory (≥5 GB), due to the necessity of loading a large gene dictionary and complementary data structures. Consequently, commodity servers cannot fulfil the computation and storage demands of large-scale text mining. Cloud-based solutions in Map-Reduce mode can partially fulfil computational resource demands. However, practically speaking, many text mining components were written in different languages, and they are dependent on a complex collection of third-party libraries, which prevents them from being readily transplanted into a high-level framework, like Hadoop and Spark. In addition, we dived into the details of load balancing, which cannot be readily supported by Map-Reduce.

An alternative solution to address this computational challenge is to harness the power of high performance computers. High performance computers (HPC) like Tianhe-2 [18] represent high-end computing infrastructures that have traditionally been used to solve compute-intensive scientific and engineering problems. The system configuration of Tianhe-2 is listed in Table 1.

Table 1

The system configuration of Tianhe-2.

Items	Content
Manufacturer	NUDT
Cores	3,120,000
Memory	1,024,000 GB
CPU	Intel Xeon E5-2692v2 12 C 2.2 GHz
Interconnect	TH Express-2
Linpack Performance(Rmax)	33,862.7 TFlop/s
Theoretical Peak(Rpeak)	54,902.4 TFlop/s
HPCG [TFlop/s]	580.109
Operating System	Kylin Linux
MPI	MPICH2 with a customized GLEX channel

Although the software stack on Tianhe-2 is designed and optimized for compute-intensive tasks, its high-performance architecture does provide the capability and capacity of big data processing. Nonetheless, to employ Tianhe-2 for big data processing is not a trivial task, which requires expert knowledge of the system architecture and parallel programming. The programming model is MPI-based (message passing interface) [18], which adds an extra dimension of complexity to normal programming languages like C/C++, Python, Java, etc. Most existing text mining tools are implemented without parallel processing. Therefore, it is necessary to develop an enabling framework that can support parallel text mining without the need to rewrite the original code. In this paper, we develop a parallel processing framework for text mining on the Tianhe-2 supercomputer. The framework integrates text mining tools as plugins. It unifies the input–output stream, implements the parallel processing across multiple compute nodes using the MPI model, and it applies a carefully crafted load balancing strategy to improve the parallelization efficiency. Without a loss of generality, we demonstrate the effectiveness of our framework using multiple NER tools as the demonstration plugins, which can recognize genes, mutations and diseases appearing in biomedical literature. More sophisticated tools of biomedical text mining can be readily integrated into the framework. In the remaining of this paper, we will introduce how paraBTM works and evaluate its performance on Tianhe-2.

2. Results and Discussion

To verify the effectiveness of the parallel framework, we constructed a corpus named 60K, which consists of 60,000 randomly selected articles from PubMed. For NER plugins, we chose three state-of-the-art tools (GNormPlus [17], tmVar [2], DNorm [7]), developed by NCBI (National Centre for Biotechnology Information). We measured the performance in terms of the total processing time and the average processing time (across all processes), and the total time includes the time of initialization and the actual processing time of different plugins.

The 60 K corpus is presented in the NXML format, which is a standard format provided by NCBI. Titles, abstracts, and full-texts from NXML files are extracted and re-written in the PubTator format. All input and output files processed by paraBTM should follow the PubTator format and the PubTator format starts with:

|t|